Upload 34 files

LTX2.3-1.0.4-new/API issues-API问题办法.txt

@@ -0,0 +1,50 @@
+1. 复制LTX桌面版的快捷方式到LTX_Shortcut
+
+2. 运行run.bat
+----
+1. Copy the LTX desktop shortcut to LTX_Shortcut
+
+2. Run run.bat
+----
+
+
+
+【问题描述 / Problem】
+系统强制使用 FAL API 生成图片，即使本地有 GPU 可用。
+System forces FAL API generation even when local GPU is available.
+
+【原因 / Cause】
+LTX 强制要求 GPU 有 31GB VRAM 才会使用本地显卡，低于此值会强制走 API 模式。
+LTX requires 31GB VRAM to use local GPU. Below this, it forces API mode.
+
+================================================================================
+【修复方法 / Fix Method】
+================================================================================
+
+运行: API issues.bat.bat (以管理员身份)
+Run: API issues.bat.bat (as Administrator)
+
+================================================================================
+================================================================================
+
+【或者手动 / Or Manual】
+
+1. 修改 VRAM 阈值 / Modify VRAM Threshold
+   文件路径 / File: C:\Program Files\LTX Desktop\resources\backend\runtime_config\runtime_policy.py
+   第16行 / Line 16:
+   原 / Original: return vram_gb < 31
+   改为 / Change:  return vram_gb < 6
+
+2. 清空 API Key / Clear API Key
+   文件路径 / File: C:\Users\<用户名>\AppData\Local\LTXDesktop\settings.json
+   原 / Original: "fal_api_key": "xxxxx"
+   改为 / Change:  "fal_api_key": ""
+
+【说明 / Note】
+- VRAM 阈值改为 6GB，意味着 6GB 及以上显存都会使用本地显卡
+- VRAM threshold set to 6GB means 6GB+ VRAM will use local GPU
+- 清空 fal_api_key 避免系统误判为已配置 API
+- Clear fal_api_key to avoid system thinking API is configured
+- 修改后重启程序即可生效
+- Restart LTX Desktop after changes
+================================================================================

LTX2.3-1.0.4-new/UI/i18n.js

@@ -0,0 +1,646 @@
+/**
+ * LTX UI i18n — 与根目录「中英文.html」思路类似，但独立脚本、避免坏 DOM/错误路径。
+ * 仅维护文案映射；动态节点由 index.js 在语言切换后刷新。
+ */
+(function (global) {
+    const STORAGE_KEY = 'ltx_ui_lang';
+
+    const STR = {
+        zh: {
+            tabVideo: '视频生成',
+            tabBatch: '智能多帧',
+            tabMotion: '视频迁移',
+            tabImage: '图像生成',
+            promptLabel: '视觉描述词 (Prompt)',
+            promptPlaceholder: '在此输入视觉描述词 (Prompt)...',
+            seedLabel: '随机种子 (Seed)',
+            seedRandom: '随机',
+            seedFixed: '固定',
+            clearVram: '释放显存',
+            clearingVram: '清理中...',
+            settingsTitle: '系统高级设置',
+            langToggleAriaZh: '切换为 English',
+            langToggleAriaEn: 'Switch to 中文',
+            sysScanning: '正在扫描 GPU...',
+            sysBusy: '运算中...',
+            sysOnline: '在线 / 就绪',
+            sysStarting: '启动中...',
+            sysOffline: '未检测到后端 (Port 3000)',
+            advancedSettings: '高级设置',
+            deviceSelect: '工作设备选择',
+            gpuDetecting: '正在检测 GPU...',
+            outputPath: '输出与上传存储路径',
+            outputPathPh: '例如: D:\\LTX_outputs',
+            savePath: '保存路径',
+            outputPathHint:
+                '系统默认会在 C 盘保留输出文件。请输入新路径后点击保存按钮。',
+            lowVram: '低显存优化',
+            lowVramDesc:
+                '尽量关闭 fast 超分、在加载管线后尝试 CPU 分层卸载（仅当引擎提供 Diffusers 式 API 才可能生效）。每次生成结束会卸载管线。说明：整模型常驻 GPU 时占用仍可能接近满配（例如约 24GB），要明显降占用需更短时长/更低分辨率或 FP8 等小权重。',
+            vramLimitLabel: '可用最高显存上限 (GB, 0为全开优先显存)',
+            vramLimitPh: '例如: 12 (0表示无限制)',
+            saveLabel: '保存',
+            modelCheckpointLabel: '视频模型（蒸馏版）',
+            modelCheckpointDefault: '默认官方蒸馏模型',
+            modelCheckpointHint: '推荐使用 distilled-fp8；仅显示 LTX 2.3 22B 蒸馏模型，避开 dev 模型。',
+            modelCheckpointSaved: '已选择模型',
+            modelCheckpointNone: '未找到可切换的蒸馏模型',
+            modelCheckpointLoadFail: '模型列表加载失败',
+            modelLoraSettings: '模型与LoRA设置',
+            modelFolder: '模型文件夹',
+            modelFolderPh: '当前 LTX 模型目录',
+            loraFolder: 'LoRA文件夹',
+            loraFolderPh: '模型目录\\loras',
+            loraFolderPath: 'LoRA 文件夹路径（可选）',
+            loraFolderPathPlaceholder: '留空使用 模型目录\\loras',
+            saveScan: '保存并扫描',
+            loraPlacementHint: '将 LoRA 文件放到当前模型目录下的 <code>loras</code> 文件夹。',
+            loraPlacementHintWithDir:
+                '将 LoRA 文件放到当前模型目录: <code>{dir}</code>\\loras',
+            basicEngine: '基础画面 / Basic EngineSpecs',
+            qualityLevel: '清晰度级别',
+            aspectRatio: '画幅比例',
+            ratio169: '16:9 电影宽幅',
+            ratio916: '9:16 移动竖屏',
+            ratio11: '1:1 方形',
+            ratio43: '4:3 经典横幅',
+            ratio34: '3:4 经典竖幅',
+            ratio219: '21:9 超宽银幕',
+            ratio921: '9:21 超长竖屏',
+            ratioRef: '跟随参考图',
+            ratioCustom: '自定义尺寸',
+            ratioRefMissing: '请先上传参考图',
+            resPreviewPrefix: '最终发送规格',
+            fpsLabel: '帧率 (FPS)',
+            durationLabel: '时长 (秒)',
+            cameraMotion: '镜头运动方式',
+            motionStatic: 'Static (静止机位)',
+            motionDollyIn: 'Dolly In (推近)',
+            motionDollyOut: 'Dolly Out (拉远)',
+            motionDollyLeft: 'Dolly Left (向左)',
+            motionDollyRight: 'Dolly Right (向右)',
+            motionJibUp: 'Jib Up (升臂)',
+            motionJibDown: 'Jib Down (降臂)',
+            motionFocus: 'Focus Shift (焦点)',
+            audioGen: '生成 AI 环境音 (Audio Gen)',
+            selectModel: '选择模型',
+            selectLora: '选择 LoRA',
+            defaultModel: '使用默认模型',
+            noLora: '不使用 LoRA',
+            loraStrength: 'LoRA 强度',
+            genSource: '生成媒介 / Generation Source',
+            startFrame: '起始帧 (首帧)',
+            endFrame: '结束帧 (尾帧)',
+            uploadStart: '上传首帧',
+            uploadEnd: '上传尾帧 (可选)',
+            refAudio: '参考音频 (A2V)',
+            uploadAudio: '点击上传音频',
+            sourceHint:
+                '💡 若仅上传首帧 = 图生视频/音视频；若同时上传首尾帧 = 首尾插帧。',
+            motionTransferTitle: '视频迁移 / Video Transfer',
+            motionRefVideoLabel: '参考视频',
+            motionVideoUploadText: '点击或拖拽视频',
+            motionVideoUploadHint: '用于动作或运镜迁移',
+            motionTargetImageLabel: '目标主体图',
+            motionImageUploadText: '点击或拖拽图片',
+            motionImageUploadHint: '作为主体/首帧引导',
+            motionTransferModeLabel: '迁移类型',
+            motionModeAction: '动作迁移',
+            motionModeCamera: '运镜迁移',
+            motionModeRepaint: '视频重绘',
+            motionControlType: '控制类型',
+            motionControlCanny: 'Canny 轮廓',
+            motionControlDepth: 'Depth 深度',
+            motionControlPose: 'Pose 姿态',
+            motionControlStrength: '控制强度',
+            motionTransferHint:
+                '动作迁移使用 Pose 姿态控制；运镜迁移使用原始参考视频作为 IC-LoRA guide。',
+            motionRefVideoName: '参考视频',
+            motionTargetImageName: '目标主体图',
+            motionUploadOk: '✅ {label}上传成功: {name}',
+            motionUploadFail: '❌ {label}上传失败: {message}',
+            motionClearRefVideo: '🧹 已清除参考视频',
+            motionClearTargetImage: '🧹 已清除目标主体图',
+            motionErrNeedVideo: '请先上传参考视频',
+            motionErrNeedImage: '请先上传目标主体图',
+            motionDefaultPromptNotice: '视频迁移未填写提示词，已使用默认视频迁移提示词',
+            motionStartLog: '正在发起视频迁移: {type}, 控制强度 {strength}',
+            motionStartMeta: 'FPS {fps}, 时长 {duration}s',
+            uploadFileStart: '正在上传{label}: {name}...',
+            fileReadFail: '读取本地文件失败',
+            downloadLabel: '下载',
+            queueTitle: '任务队列',
+            queueIdle: '空闲',
+            queueQueued: '排队中',
+            queueRunning: '执行中',
+            queueComplete: '已完成',
+            queueError: '失败',
+            queueCancelled: '已取消',
+            queueWaiting: '等待 {n}',
+            queueRunningSummary: '执行中 1 / 排队 {n}',
+            queueNoTasks: '暂无任务',
+            queueViewResult: '查看结果',
+            queuePosition: '队列第 {n} 位',
+            queueTaskTypeVideo: '视频',
+            queueTaskTypeMotion: '迁移',
+            queueTaskTypeBatch: '批量',
+            queueTaskTypeImage: '图像',
+            queueSubmitLog: '📥 已加入队列: {id}（前面还有 {n} 个任务）',
+            queueDoneLog: '✅ 队列任务完成: {label}',
+            queueFailLog: '❌ 队列任务失败: {label} - {error}',
+            queueCancelLog: '🛑 队列任务已取消: {label}',
+            replayRun: '重跑',
+            replayLoad: '载入参数',
+            replayLabel: 'Replay',
+            replayMissing: '⚠️ 这个历史任务没有可重放参数',
+            replayQueuedLog: '↻ Replay 已加入队列: {id}',
+            replayLoadedLog: '↗ 已载入 Replay 参数，可微调后重新渲染',
+            replayFailed: 'Replay 失败',
+            previewLoadSeed: '载入种子',
+            previewLoadParams: '载入参数',
+            previewNoReplaySeed: '⚠️ 当前预览没有可载入的种子',
+            previewNoReplayParams: '⚠️ 当前预览没有可载入的参数',
+            previewSeedLoadedLog: '已载入种子 {seed}，并切换为固定种子',
+            previewNoDownload: '❌ 当前没有可下载的预览内容',
+            imgPreset: '预设分辨率 (Presets)',
+            imgOptSquare: '1:1 Square (1024x1024)',
+            imgOptLand: '16:9 Landscape (1280x720)',
+            imgOptPort: '9:16 Portrait (720x1280)',
+            imgOptCustom: 'Custom 自定义...',
+            width: '宽度',
+            height: '高度',
+            samplingSteps: '采样步数 (Steps)',
+            smartMultiFrameGroup: '智能多帧',
+            workflowModeLabel: '工作流模式（点击切换）',
+            wfSingle: '单次多关键帧',
+            wfSegments: '分段拼接',
+            uploadImages: '上传图片',
+            uploadMulti1: '点击或拖入多张图片',
+            uploadMulti2: '支持一次选多张，可多次添加',
+            batchStripTitle: '已选图片 · 顺序 = 播放先后',
+            batchStripHint: '在缩略图上按住拖动排序；松手落入虚线框位置',
+            batchFfmpegHint:
+                '💡 <strong>分段模式</strong>：2 张 = 1 段；3 张 = 2 段再拼接。<strong>单次模式</strong>：几张图就几个 latent 锚点，一条视频出片。<br>多段需 <code style="font-size:9px;">ffmpeg</code>：装好后加 PATH，或设环境变量 <code style="font-size:9px;">LTX_FFMPEG_PATH</code>，或在 <code style="font-size:9px;">%LOCALAPPDATA%\\LTXDesktop\\ffmpeg_path.txt</code> 第一行写 ffmpeg.exe 完整路径。',
+            bgmLabel: '成片配乐（可选，统一音轨）',
+            bgmUploadHint: '上传一条完整 BGM（生成完成后会替换整段成片的音轨）',
+            mainRender: '开始渲染',
+            waitingTask: '等待分配渲染任务...',
+            libHistory: '历史资产 / ASSETS',
+            libLog: '系统日志 / LOGS',
+            refresh: '刷新',
+            logReady: '> LTX-2 Studio Ready. Expecting commands...',
+            resizeHandleTitle: '拖动调整面板高度',
+            batchNeedTwo: '💡 请上传至少2张图片',
+            batchSegTitle: '视频片段设置（分段拼接）',
+            batchSegClip: '片段',
+            batchSegDuration: '时长',
+            batchSegSec: '秒',
+            batchSegPrompt: '片段提示词',
+            batchSegPromptPh: '此片段的提示词，如：跳舞、吃饭...',
+            batchKfPanelTitle: '单次多关键帧 · 时间轴',
+            batchTotalDur: '总时长',
+            batchTotalSec: '秒',
+            batchPanelHint:
+                '为每一张图设置独立持续时间：第 1 张从 0 s 开始，下一张在上一张持续结束后接上，最后一张也会延续自己的时长。因后端按<strong>整数秒</strong>建序列，实际请求里的整段时长为合计秒数<strong>向上取整</strong>（至少 2），略长于小数合计时属正常。镜头与 FPS 仍用左侧「视频生成」。',
+            batchKfTitle: '关键帧',
+            batchStrength: '引导强度',
+            batchFrameDuration: '持续时长',
+            batchFrameDurationTitle: '这一张图在时间轴上占用的时长（秒）；总时长 = 每张图持续时长之和',
+            batchGapTitle: '间隔',
+            batchSec: '秒',
+            batchAnchorStart: '片头',
+            batchAnchorEnd: '片尾',
+            batchAnchorLast: '最后一帧开始',
+            batchThumbDrag: '按住拖动排序',
+            batchThumbRemove: '删除',
+            batchAddMore: '＋ 继续添加',
+            batchGapInputTitle: '这一张图在时间轴上占用的时长（秒）；总时长 = 每张图持续时长之和',
+            batchStrengthTitle: '与 Comfy guide strength 类似，中间帧可调低（如 0.2）减轻闪烁',
+            batchTotalPillTitle: '等于下方各「持续时长」之和，无需单独填写',
+            defaultPath: '默认路径',
+            phase_loading_model: '加载权重',
+            phase_encoding_text: 'T5 编码',
+            phase_validating_request: '校验请求',
+            phase_uploading_audio: '上传音频',
+            phase_uploading_image: '上传图像',
+            phase_inference: 'AI 推理',
+            phase_downloading_output: '下载结果',
+            phase_complete: '完成',
+            gpuBusyPrefix: 'GPU 运算中',
+            progressStepUnit: '步',
+            loaderGpuAlloc: 'GPU 正在分配资源...',
+            warnGenerating: '⚠️ 当前正在生成中，请等待完成',
+            warnBatchPrompt: '⚠️ 智能多帧请至少填写：顶部主提示词、本页全局补充词或某一「片段提示词」',
+            warnNeedPrompt: '⚠️ 请输入提示词后再开始渲染',
+            warnVideoLong: '⚠️ 时长设定为 {n}s 极长，可能导致显存溢出或耗时较久。',
+            errBatchMinImages: '请上传至少2张图片',
+            errSingleKfPrompt: '单次多关键帧请至少填写顶部主提示词或本页全局补充词',
+            loraNoneLabel: '无',
+            modelDefaultLabel: '默认',
+            tabTts: 'TTS 语音',
+            ttsStatusBarDetecting: '🔍 正在检测 TTS 模型...',
+            ttsTextTitle: '合成文本 / Text',
+            ttsTextHint: '支持在文本开头加英文括号描述声音，例如：<code style="font-size:10px;">(年轻女声，温柔甜美)</code>',
+            ttsTextPlaceholder: '输入要合成的文本内容...',
+            ttsModeTitle: '合成模式 / Mode',
+            ttsModeTextOnly: '🗣️ 文字转语音（含声音设计）',
+            ttsModeClone: '🎙️ 声音克隆',
+            ttsModeUltimate: '⭐ 终极克隆（最高还原度）',
+            ttsRefLabel: '📎 参考音频（Reference）',
+            ttsRefUploadHint: '点击上传参考音频 (.wav / .mp3)',
+            ttsUltimateLabel: '📝 参考音频对应的文本转录（可选）',
+            ttsUltimatePlaceholder: '与参考音频完全一致的文本内容...',
+            ttsParamsTitle: '高级参数 / Parameters',
+            ttsCfgLabel: 'CFG 强度',
+            ttsStepsLabel: '推理步数',
+            ttsResultTitle: '生成结果 / Output',
+            ttsDownload: '⬇️ 下载音频',
+            ttsGenBtn: '🎙️ 开始生成语音',
+            ttsGenBusy: '⏳ 生成中...',
+            ttsErrNoText: '❌ TTS: ��输入合成文本',
+            ttsErrNoRef: '❌ TTS: 声音克隆模式需要上传参考音频',
+            ttsStatusReady: '✅ VoxCPM2 就绪 — 模型目录: ',
+            ttsStatusNoPkq: '❌ voxcpm 包未安装，请在 LTX Python 环境中运行: pip install voxcpm',
+            ttsStatusNoDir: '❌ 请将 VoxCPM2 文件夹放到: ',
+            ttsStatusNotAvail: '⚠️ TTS 不可用，请检查配置',
+            ttsStatusConnErr: '❌ 无法连接后端 TTS 接口: ',
+        },
+        en: {
+            tabVideo: 'Video',
+            tabBatch: 'Frames',
+            tabMotion: 'Transfer',
+            tabImage: 'Image',
+            promptLabel: 'Prompt',
+            promptPlaceholder: 'Describe the scene...',
+            seedLabel: 'Seed',
+            seedRandom: 'Random',
+            seedFixed: 'Fixed',
+            clearVram: 'Clear VRAM',
+            clearingVram: 'Clearing...',
+            settingsTitle: 'Advanced settings',
+            langToggleAriaZh: 'Switch to English',
+            langToggleAriaEn: 'Switch to Chinese',
+            sysScanning: 'Scanning GPU...',
+            sysBusy: 'Busy...',
+            sysOnline: 'Online / Ready',
+            sysStarting: 'Starting...',
+            sysOffline: 'Backend offline (port 3000)',
+            advancedSettings: 'Advanced',
+            deviceSelect: 'GPU device',
+            gpuDetecting: 'Detecting GPU...',
+            outputPath: 'Output & upload folder',
+            outputPathPh: 'e.g. D:\\LTX_outputs',
+            savePath: 'Save path',
+            outputPathHint:
+                'Outputs default to C: drive. Enter a folder and click Save.',
+            lowVram: 'Low-VRAM mode',
+            lowVramDesc:
+                'Tries to reduce VRAM (engine-dependent). Shorter duration / lower resolution helps more.',
+            vramLimitLabel: 'Max VRAM Limit (GB, 0 for unlimited)',
+            vramLimitPh: 'e.g. 12 (0 for unlimited)',
+            saveLabel: 'Save',
+            modelCheckpointLabel: 'Video model (distilled)',
+            modelCheckpointDefault: 'Default official distilled model',
+            modelCheckpointHint: 'Recommended: distilled-fp8. Only LTX 2.3 22B distilled models are shown; dev models are hidden.',
+            modelCheckpointSaved: 'Model selected',
+            modelCheckpointNone: 'No switchable distilled models found',
+            modelCheckpointLoadFail: 'Failed to load model list',
+            modelLoraSettings: 'Model & LoRA folders',
+            modelFolder: 'Models folder',
+            modelFolderPh: 'Current LTX models directory',
+            loraFolder: 'LoRAs folder',
+            loraFolderPh: 'models\\loras',
+            loraFolderPath: 'LoRA folder path (optional)',
+            loraFolderPathPlaceholder: 'Empty = models\\loras',
+            saveScan: 'Save & scan',
+            loraHint: 'Put .safetensors / .ckpt LoRAs here, then refresh lists.',
+            basicEngine: 'Basic / Engine',
+            qualityLevel: 'Quality',
+            aspectRatio: 'Aspect ratio',
+            ratio169: '16:9 widescreen',
+            ratio916: '9:16 portrait',
+            ratio11: '1:1 square',
+            ratio43: '4:3 classic',
+            ratio34: '3:4 portrait',
+            ratio219: '21:9 ultrawide',
+            ratio921: '9:21 tall',
+            ratioRef: 'Match reference',
+            ratioCustom: 'Custom size',
+            ratioRefMissing: 'upload a reference first',
+            resPreviewPrefix: 'Output',
+            fpsLabel: 'FPS',
+            durationLabel: 'Duration (s)',
+            cameraMotion: 'Camera motion',
+            motionStatic: 'Static',
+            motionDollyIn: 'Dolly in',
+            motionDollyOut: 'Dolly out',
+            motionDollyLeft: 'Dolly left',
+            motionDollyRight: 'Dolly right',
+            motionJibUp: 'Jib up',
+            motionJibDown: 'Jib down',
+            motionFocus: 'Focus shift',
+            audioGen: 'AI ambient audio',
+            selectModel: 'Model',
+            selectLora: 'LoRA',
+            defaultModel: 'Default model',
+            noLora: 'No LoRA',
+            loraStrength: 'LoRA strength',
+            genSource: 'Source media',
+            startFrame: 'Start frame',
+            endFrame: 'End frame (optional)',
+            uploadStart: 'Upload start',
+            uploadEnd: 'Upload end (opt.)',
+            refAudio: 'Reference audio (A2V)',
+            uploadAudio: 'Upload audio',
+            sourceHint:
+                '💡 Start only = I2V / A2V; start + end = interpolation.',
+            motionTransferTitle: 'Video Transfer',
+            motionRefVideoLabel: 'Reference video',
+            motionVideoUploadText: 'Click or drop a video',
+            motionVideoUploadHint: 'Used for action or camera transfer',
+            motionTargetImageLabel: 'Target subject image',
+            motionImageUploadText: 'Click or drop an image',
+            motionImageUploadHint: 'Guides subject identity / first frame',
+            motionTransferModeLabel: 'Transfer type',
+            motionModeAction: 'Action transfer',
+            motionModeCamera: 'Camera transfer',
+            motionModeRepaint: 'Video repaint',
+            motionControlType: 'Control type',
+            motionControlCanny: 'Canny contour',
+            motionControlDepth: 'Depth',
+            motionControlPose: 'Pose',
+            motionControlStrength: 'Control strength',
+            motionTransferHint:
+                'Action transfer uses Pose control; camera transfer uses the raw reference video as an IC-LoRA guide.',
+            motionRefVideoName: 'reference video',
+            motionTargetImageName: 'target subject image',
+            motionUploadOk: '✅ {label} uploaded: {name}',
+            motionUploadFail: '❌ {label} upload failed: {message}',
+            motionClearRefVideo: '🧹 Reference video cleared',
+            motionClearTargetImage: '🧹 Target subject image cleared',
+            motionErrNeedVideo: 'Upload a reference video first',
+            motionErrNeedImage: 'Upload a target subject image first',
+            motionDefaultPromptNotice: 'No prompt entered for video transfer; using the default video transfer prompt.',
+            motionStartLog: 'Starting video transfer: {type}, strength {strength}',
+            motionStartMeta: 'FPS {fps}, duration {duration}s',
+            uploadFileStart: 'Uploading {label}: {name}...',
+            fileReadFail: 'Failed to read local file',
+            downloadLabel: 'Download',
+            queueTitle: 'Task Queue',
+            queueIdle: 'Idle',
+            queueQueued: 'Queued',
+            queueRunning: 'Running',
+            queueComplete: 'Done',
+            queueError: 'Error',
+            queueCancelled: 'Cancelled',
+            queueWaiting: '{n} waiting',
+            queueRunningSummary: 'Running 1 / Queued {n}',
+            queueNoTasks: 'No tasks',
+            queueViewResult: 'View result',
+            queuePosition: 'Queue #{n}',
+            queueTaskTypeVideo: 'Video',
+            queueTaskTypeMotion: 'Transfer',
+            queueTaskTypeBatch: 'Batch',
+            queueTaskTypeImage: 'Image',
+            queueSubmitLog: '📥 Added to queue: {id} ({n} ahead)',
+            queueDoneLog: '✅ Queue task complete: {label}',
+            queueFailLog: '❌ Queue task failed: {label} - {error}',
+            queueCancelLog: '🛑 Queue task cancelled: {label}',
+            replayRun: 'Replay',
+            replayLoad: 'Load',
+            replayLabel: 'Replay',
+            replayMissing: '⚠️ This item has no replay payload',
+            replayQueuedLog: '↻ Replay added to queue: {id}',
+            replayLoadedLog: '↗ Replay settings loaded; adjust and render again',
+            replayFailed: 'Replay failed',
+            previewLoadSeed: 'Load Seed',
+            previewLoadParams: 'Load Params',
+            previewNoReplaySeed: '⚠️ Current preview has no seed to load',
+            previewNoReplayParams: '⚠️ Current preview has no params to load',
+            previewSeedLoadedLog: 'Loaded seed {seed} and switched to fixed seed',
+            previewNoDownload: '❌ No downloadable preview right now',
+            imgPreset: 'Resolution presets',
+            imgOptSquare: '1:1 (1024×1024)',
+            imgOptLand: '16:9 (1280×720)',
+            imgOptPort: '9:16 (720×1280)',
+            imgOptCustom: 'Custom...',
+            width: 'Width',
+            height: 'Height',
+            samplingSteps: 'Steps',
+            smartMultiFrameGroup: 'Smart multi-frame',
+            workflowModeLabel: 'Workflow',
+            wfSingle: 'Single pass',
+            wfSegments: 'Segments',
+            uploadImages: 'Upload images',
+            uploadMulti1: 'Click or drop multiple images',
+            uploadMulti2: 'Multi-select OK; add more anytime.',
+            batchStripTitle: 'Order = playback',
+            batchStripHint: 'Drag thumbnails to reorder.',
+            batchFfmpegHint:
+                '💡 <strong>Segments</strong>: 2 images → 1 clip; 3 → 2 clips stitched. <strong>Single</strong>: N images → N latent anchors, one video.<br>Stitching needs <code style="font-size:9px;">ffmpeg</code> on PATH, or <code style="font-size:9px;">LTX_FFMPEG_PATH</code>, or <code style="font-size:9px;">%LOCALAPPDATA%\\LTXDesktop\\ffmpeg_path.txt</code> with full path to ffmpeg.exe.',
+            bgmLabel: 'Full-length BGM (optional)',
+            bgmUploadHint: 'Replaces final mix audio after generation.',
+            mainRender: 'Render',
+            waitingTask: 'Waiting for task...',
+            libHistory: 'Assets',
+            libLog: 'Logs',
+            refresh: 'Refresh',
+            logReady: '> LTX-2 Studio ready.',
+            resizeHandleTitle: 'Drag to resize panel',
+            batchNeedTwo: '💡 Upload at least 2 images',
+            batchSegTitle: 'Segment settings',
+            batchSegClip: 'Clip',
+            batchSegDuration: 'Duration',
+            batchSegSec: 's',
+            batchSegPrompt: 'Prompt',
+            batchSegPromptPh: 'e.g. dancing, walking...',
+            batchKfPanelTitle: 'Single pass · timeline',
+            batchTotalDur: 'Total',
+            batchTotalSec: 's',
+            batchPanelHint:
+                'Set an independent duration for each image: the first starts at 0s, each next image starts after the previous duration, and the last image also keeps its own duration. Backend uses whole seconds (ceil, min 2). Motion & FPS use the Video panel.',
+            batchKfTitle: 'Keyframe',
+            batchStrength: 'Strength',
+            batchFrameDuration: 'Duration',
+            batchFrameDurationTitle: 'How long this image occupies on the timeline; total = sum of all image durations',
+            batchGapTitle: 'Gap',
+            batchSec: 's',
+            batchAnchorStart: 'start',
+            batchAnchorEnd: 'end',
+            batchAnchorLast: 'last starts',
+            batchThumbDrag: 'Drag to reorder',
+            batchThumbRemove: 'Remove',
+            batchAddMore: '+ Add more',
+            batchGapInputTitle: 'How long this image occupies on the timeline; total = sum of all image durations',
+            batchStrengthTitle: 'Guide strength (lower on middle keys may reduce flicker)',
+            batchTotalPillTitle: 'Equals the sum of durations below',
+            defaultPath: 'default',
+            phase_loading_model: 'Loading weights',
+            phase_encoding_text: 'T5 encode',
+            phase_validating_request: 'Validating',
+            phase_uploading_audio: 'Uploading audio',
+            phase_uploading_image: 'Uploading image',
+            phase_inference: 'Inference',
+            phase_downloading_output: 'Downloading',
+            phase_complete: 'Done',
+            gpuBusyPrefix: 'GPU',
+            progressStepUnit: 'steps',
+            loaderGpuAlloc: 'Allocating GPU...',
+            warnGenerating: '⚠️ Already generating, please wait.',
+            warnBatchPrompt: '⚠️ Enter main prompt, page extra prompt, or a segment prompt.',
+            warnNeedPrompt: '⚠️ Enter a prompt first.',
+            warnVideoLong: '⚠️ Duration {n}s is very long; may OOM or take a long time.',
+            errBatchMinImages: 'Upload at least 2 images.',
+            errSingleKfNeedPrompt: 'Enter main or page extra prompt for single-pass keyframes.',
+            loraNoneLabel: 'none',
+            modelDefaultLabel: 'default',
+            loraPlacementHintWithDir:
+                'Place LoRAs into the current models directory: <code>{dir}</code>\\loras',
+            loraPlacementHint: 'Place LoRAs in the <code>loras</code> folder under the current models directory.',
+            tabTts: 'TTS',
+            ttsStatusBarDetecting: '🔍 Detecting TTS model...',
+            ttsTextTitle: 'Synthesis Text',
+            ttsTextHint: 'Supports descriptors in brackets at start, e.g. <code style="font-size:10px;">(Young female, soft and sweet)</code>',
+            ttsTextPlaceholder: 'Enter text to synthesize...',
+            ttsModeTitle: 'Synthesis Mode',
+            ttsModeTextOnly: '🗣️ Text to Speech (with sound design)',
+            ttsModeClone: '🎙️ Voice Cloning',
+            ttsModeUltimate: '⭐ Ultimate Clone (Max similarity)',
+            ttsRefLabel: '📎 Reference Audio',
+            ttsRefUploadHint: 'Click to upload reference (.wav / .mp3)',
+            ttsUltimateLabel: '📝 Audio Transcript (Optional)',
+            ttsUltimatePlaceholder: 'Text content matches reference audio exactly...',
+            ttsParamsTitle: 'Parameters',
+            ttsCfgLabel: 'CFG Strength',
+            ttsStepsLabel: 'Inference Steps',
+            ttsResultTitle: 'Output',
+            ttsDownload: '⬇️ Download Audio',
+            ttsGenBtn: '🎙️ Start Synthesis',
+            ttsGenBusy: '⏳ Generating...',
+            ttsErrNoText: '❌ TTS: Please enter synthesis text',
+            ttsErrNoRef: '❌ TTS: Voice cloning requires reference audio',
+            ttsStatusReady: '✅ VoxCPM2 Ready — Model dir: ',
+            ttsStatusNoPkq: '❌ voxcpm package not installed. Run: pip install voxcpm',
+            ttsStatusNoDir: '❌ Put the VoxCPM2 folder at: ',
+            ttsStatusNotAvail: '⚠️ TTS unavailable, check config',
+            ttsStatusConnErr: '❌ Cannot connect to TTS API: ',
+        },
+    };
+
+    function getLang() {
+        return localStorage.getItem(STORAGE_KEY) === 'en' ? 'en' : 'zh';
+    }
+
+    function setLang(lang) {
+        const L = lang === 'en' ? 'en' : 'zh';
+        localStorage.setItem(STORAGE_KEY, L);
+        document.documentElement.lang = L === 'en' ? 'en' : 'zh-CN';
+        try {
+            applyI18n();
+        } catch (err) {
+            console.error('[i18n] applyI18n failed:', err);
+        }
+        updateLangButton();
+        if (typeof global.onUiLanguageChanged === 'function') {
+            try {
+                global.onUiLanguageChanged();
+            } catch (e) {
+                console.warn('onUiLanguageChanged', e);
+            }
+        }
+    }
+
+    function t(key) {
+        const L = getLang();
+        const table = STR[L] || STR.zh;
+        if (Object.prototype.hasOwnProperty.call(table, key)) return table[key];
+        if (Object.prototype.hasOwnProperty.call(STR.zh, key)) return STR.zh[key];
+        return key;
+    }
+
+    function applyI18n(root) {
+        root = root || document;
+        root.querySelectorAll('[data-i18n]').forEach(function (el) {
+            var key = el.getAttribute('data-i18n');
+            if (!key) return;
+            if (el.tagName === 'OPTION') {
+                el.textContent = t(key);
+            } else {
+                el.textContent = t(key);
+            }
+        });
+        root.querySelectorAll('[data-i18n-placeholder]').forEach(function (el) {
+            var key = el.getAttribute('data-i18n-placeholder');
+            if (key) el.placeholder = t(key);
+        });
+        root.querySelectorAll('[data-i18n-title]').forEach(function (el) {
+            var key = el.getAttribute('data-i18n-title');
+            if (key) el.title = t(key);
+        });
+        root.querySelectorAll('[data-i18n-html]').forEach(function (el) {
+            var key = el.getAttribute('data-i18n-html');
+            if (key) el.innerHTML = t(key);
+        });
+        root.querySelectorAll('[data-i18n-value]').forEach(function (el) {
+            var key = el.getAttribute('data-i18n-value');
+            if (key && (el.tagName === 'INPUT' || el.tagName === 'BUTTON')) {
+                el.value = t(key);
+            }
+        });
+    }
+
+    function updateLangButton() {
+        var btn = document.getElementById('lang-toggle-btn');
+        if (!btn) return;
+        btn.textContent = getLang() === 'zh' ? 'EN' : '中';
+        btn.setAttribute(
+            'aria-label',
+            getLang() === 'zh' ? t('langToggleAriaZh') : t('langToggleAriaEn')
+        );
+        btn.classList.toggle('active', getLang() === 'en');
+    }
+
+    function toggleUiLanguage() {
+        try {
+            setLang(getLang() === 'zh' ? 'en' : 'zh');
+        } catch (err) {
+            console.error('[i18n] toggleUiLanguage failed:', err);
+        }
+    }
+
+    /** 避免 CSP 拦截内联 onclick；确保按钮一定能触发 */
+    function bindLangToggleButton() {
+        var btn = document.getElementById('lang-toggle-btn');
+        if (!btn || btn.dataset.i18nBound === '1') return;
+        btn.dataset.i18nBound = '1';
+        btn.removeAttribute('onclick');
+        btn.addEventListener('click', function (ev) {
+            ev.preventDefault();
+            toggleUiLanguage();
+        });
+    }
+
+    function boot() {
+        document.documentElement.lang = getLang() === 'en' ? 'en' : 'zh-CN';
+        try {
+            applyI18n();
+        } catch (err) {
+            console.error('[i18n] applyI18n failed:', err);
+        }
+        updateLangButton();
+        bindLangToggleButton();
+    }
+
+    global.getUiLang = getLang;
+    global.setUiLang = setLang;
+    global.t = t;
+    global.applyI18n = applyI18n;
+    global.toggleUiLanguage = toggleUiLanguage;
+    global.updateLangToggleButton = updateLangButton;
+
+    if (document.readyState === 'loading') {
+        document.addEventListener('DOMContentLoaded', boot);
+    } else {
+        boot();
+    }
+})(typeof window !== 'undefined' ? window : global);

LTX2.3-1.0.4-new/UI/index.css

@@ -0,0 +1,982 @@
+:root {
+            --accent:      #2563EB;     /* Refined blue – not too bright, not purple */
+            --accent-hover:#3B82F6;
+            --accent-dim:  rgba(37,99,235,0.14);
+            --accent-ring: rgba(37,99,235,0.35);
+            --bg:          #111113;
+            --panel:       #18181B;
+            --panel-2:     #1F1F23;
+            --item:        rgba(255,255,255,0.035);
+            --border:      rgba(255,255,255,0.08);
+            --border-2:    rgba(255,255,255,0.05);
+            --text-dim:    #71717A;
+            --text-sub:    #A1A1AA;
+            --text:        #FAFAFA;
+        }
+
+        * { box-sizing: border-box; -webkit-font-smoothing: antialiased; min-width: 0; }
+        body {
+            background: var(--bg); margin: 0; color: var(--text);
+            font-family: -apple-system, "SF Pro Display", "Segoe UI", sans-serif;
+            display: flex; height: 100vh; overflow: hidden;
+            font-size: 13px; line-height: 1.5;
+        }
+
+        .sidebar {
+            width: 460px; min-width: 460px;
+            background: var(--panel);
+            border-right: 1px solid var(--border);
+            display: flex; flex-direction: column; z-index: 20;
+            overflow-y: auto; overflow-x: hidden;
+        }
+
+        /* Scrollbar */
+        ::-webkit-scrollbar { width: 5px; height: 5px; }
+        ::-webkit-scrollbar-track { background: transparent; }
+        ::-webkit-scrollbar-thumb { background: rgba(255,255,255,0.08); border-radius: 10px; }
+        ::-webkit-scrollbar-thumb:hover { background: rgba(255,255,255,0.18); }
+
+        .sidebar-header { padding: 24px 24px 4px; }
+
+        .lang-toggle {
+            background: #333;
+            border: 1px solid #555;
+            color: var(--text-dim);
+            padding: 4px 10px;
+            border-radius: 6px;
+            font-size: 11px;
+            cursor: pointer;
+            transition: background 0.15s, color 0.15s, border-color 0.15s;
+            font-weight: 700;
+            min-width: 44px;
+            flex-shrink: 0;
+        }
+        .lang-toggle:hover {
+            background: var(--item);
+            color: var(--text);
+            border-color: var(--accent);
+        }
+        .lang-toggle.active {
+            background: #333;
+            color: var(--text);
+            border-color: #555;
+        }
+        .sidebar-section { padding: 8px 24px 18px; border-bottom: 1px solid var(--border); }
+
+        .setting-group {
+            background: rgba(255,255,255,0.025);
+            border: 1px solid var(--border-2);
+            border-radius: 10px;
+            padding: 14px;
+            margin-bottom: 12px;
+        }
+        .group-title {
+            font-size: 10px; color: var(--text-dim); font-weight: 700;
+            text-transform: uppercase; letter-spacing: 0.7px;
+            margin-bottom: 12px; padding-bottom: 5px;
+            border-bottom: 1px solid var(--border-2);
+        }
+
+        /* Mode Tabs */
+        .tabs {
+            display: flex; gap: 4px; margin-bottom: 14px;
+            background: rgba(255,255,255,0.04);
+            padding: 4px; border-radius: 10px;
+            border: 1px solid var(--border-2);
+        }
+        .tab {
+            flex: 1; padding: 9px 0; text-align: center; border-radius: 7px;
+            cursor: pointer; font-size: 12px; color: var(--text-dim);
+            transition: all 0.2s; font-weight: 600;
+            display: flex; align-items: center; justify-content: center;
+        }
+        .tab.active { background: var(--accent); color: #fff; box-shadow: 0 1px 6px rgba(10,132,255,0.45); }
+        .tab:hover:not(.active) { background: rgba(255,255,255,0.06); color: var(--text); }
+        html[lang="en"] .tabs { gap: 3px; padding: 3px; }
+        html[lang="en"] .tab { font-size: 11px; padding: 8px 0; font-weight: 700; }
+        html[lang="en"] .tab svg { width: 13px; height: 13px; margin-right: 4px !important; flex-shrink: 0; }
+
+        .label-group { display: flex; justify-content: space-between; align-items: center; margin-bottom: 6px; }
+        label { display: block; font-size: 11px; color: var(--text-dim); font-weight: 600; text-transform: uppercase; letter-spacing: 0.5px; margin-bottom: 6px; }
+        .val-badge { font-size: 11px; color: var(--accent); font-family: "SF Mono", ui-monospace, monospace; font-weight: 600; }
+
+        input[type="text"], input[type="number"], select, textarea {
+            width: 100%; background: var(--panel-2);
+            border: 1px solid var(--border);
+            border-radius: 7px; color: var(--text);
+            padding: 8px 11px; font-size: 12.5px; outline: none; margin-bottom: 9px;
+            /* Only transition border/shadow – NOT background-image to prevent arrow flicker */
+            transition: border-color 0.15s, box-shadow 0.15s;
+        }
+        input:focus, select:focus, textarea:focus {
+            border-color: var(--accent);
+            box-shadow: 0 0 0 2px var(--accent-ring);
+        }
+        select {
+            -webkit-appearance: none; -moz-appearance: none; appearance: none;
+            /* Stable grey arrow – no background shorthand so it won't animate */
+            background-color: var(--panel-2);
+            background-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='12' height='12' viewBox='0 0 24 24' fill='none' stroke='%2371717A' stroke-width='2.5' stroke-linecap='round' stroke-linejoin='round'%3E%3Cpolyline points='6 9 12 15 18 9'/%3E%3C/svg%3E");
+            background-repeat: no-repeat;
+            background-position: right 10px center;
+            background-size: 12px;
+            padding-right: 28px;
+            cursor: pointer;
+            /* Explicitly do NOT transition background properties */
+            transition: border-color 0.15s, box-shadow 0.15s;
+        }
+        select:focus { background-color: var(--panel-2); }
+        select option { background: #27272A; color: var(--text); }
+        textarea { resize: vertical; min-height: 78px; font-family: inherit; }
+
+        .slider-container { display: flex; align-items: center; gap: 12px; margin-bottom: 14px; }
+        input[type="range"] { flex: 1; accent-color: var(--accent); height: 4px; cursor: pointer; border-radius: 2px; }
+
+        .upload-zone {
+            border: 1px dashed var(--border); border-radius: 10px;
+            padding: 18px 10px; text-align: center; cursor: pointer;
+            background: rgba(255,255,255,0.03); margin-bottom: 10px; position: relative;
+            transition: all 0.2s;
+        }
+        .upload-zone:hover, .upload-zone.dragover { background: var(--accent-dim); border-color: var(--accent); }
+        .upload-zone.has-images {
+            padding: 12px; background: rgba(255,255,255,0.025);
+        }
+        .upload-zone.has-images .upload-placeholder-mini {
+            display: flex; align-items: center; gap: 8px; justify-content: center;
+            color: var(--text-dim); font-size: 11px;
+        }
+        .upload-zone.has-images .upload-placeholder-mini span {
+            background: var(--item); padding: 6px 12px; border-radius: 6px;
+        }
+        #batch-images-placeholder { display: block; }
+        .upload-zone.has-images #batch-images-placeholder { display: none; }
+
+        /* 批量模式：上传区下方的横向缩略图条 */
+        .batch-thumb-strip-wrap {
+            margin-top: 10px;
+            margin-bottom: 4px;
+        }
+        .batch-thumb-strip-head {
+            display: flex;
+            flex-direction: column;
+            gap: 2px;
+            margin-bottom: 8px;
+        }
+        .batch-thumb-strip-title {
+            font-size: 11px;
+            font-weight: 700;
+            color: var(--text-sub);
+        }
+        .batch-thumb-strip-hint {
+            font-size: 10px;
+            color: var(--text-dim);
+        }
+        .batch-images-container {
+            display: flex;
+            flex-direction: row;
+            flex-wrap: nowrap;
+            gap: 10px;
+            overflow-x: auto;
+            overflow-y: visible;
+            padding: 6px 4px 14px;
+            margin: 0 -4px;
+            scrollbar-width: thin;
+            scrollbar-color: var(--border) transparent;
+            align-items: center;
+        }
+        .batch-images-container::-webkit-scrollbar { height: 6px; }
+        .batch-images-container::-webkit-scrollbar-thumb {
+            background: var(--border);
+            border-radius: 3px;
+        }
+        .batch-image-wrapper {
+            flex: 0 0 72px;
+            width: 72px;
+            height: 72px;
+            position: relative;
+            border-radius: 10px;
+            overflow: hidden;
+            background: var(--item);
+            border: 1px solid var(--border);
+            cursor: grab;
+            touch-action: none;
+            user-select: none;
+            -webkit-user-select: none;
+            transition:
+                flex-basis 0.38s cubic-bezier(0.22, 1, 0.36, 1),
+                width 0.38s cubic-bezier(0.22, 1, 0.36, 1),
+                min-width 0.38s cubic-bezier(0.22, 1, 0.36, 1),
+                margin 0.38s cubic-bezier(0.22, 1, 0.36, 1),
+                opacity 0.25s ease,
+                border-color 0.2s ease,
+                box-shadow 0.2s ease,
+                transform 0.28s cubic-bezier(0.22, 1, 0.36, 1);
+        }
+        .batch-image-wrapper:active { cursor: grabbing; }
+        .batch-image-wrapper.batch-thumb--source {
+            flex: 0 0 0;
+            width: 0;
+            min-width: 0;
+            height: 72px;
+            margin: 0;
+            padding: 0;
+            border: none;
+            overflow: hidden;
+            opacity: 0;
+            background: transparent;
+            box-shadow: none;
+            pointer-events: none;
+            /* 收起必须瞬时：若与占位框同时用 0.38s 过渡，右侧缩略图会与「突然出现」的槽位不同步而闪一下 */
+            transition: none !important;
+        }
+        /* 按下瞬间：冻结其它卡片与槽位动画，避免「槽位插入 + 邻居过渡」两帧打架 */
+        .batch-images-container.is-batch-settling .batch-image-wrapper:not(.batch-thumb--source) {
+            transition: none !important;
+        }
+        .batch-images-container.is-batch-settling .batch-thumb-drop-slot {
+            animation: none;
+            opacity: 1;
+        }
+        /* 拖动时跟手的浮动缩略图（避免原槽位透明后光标下像「黑块」） */
+        .batch-thumb-floating-ghost {
+            position: fixed;
+            left: 0;
+            top: 0;
+            z-index: 99999;
+            width: 76px;
+            height: 76px;
+            border-radius: 12px;
+            overflow: hidden;
+            pointer-events: none;
+            will-change: transform;
+            box-shadow:
+                0 20px 50px rgba(0, 0, 0, 0.45),
+                0 10px 28px rgba(0, 0, 0, 0.28),
+                0 0 0 1px rgba(255, 255, 255, 0.18);
+            transform: translate3d(0, 0, 0) scale(1.06) rotate(-1deg);
+        }
+        .batch-thumb-floating-ghost img {
+            width: 100%;
+            height: 100%;
+            object-fit: cover;
+            display: block;
+            pointer-events: none;
+        }
+        .batch-thumb-drop-slot {
+            flex: 0 0 72px;
+            width: 72px;
+            height: 72px;
+            box-sizing: border-box;
+            border-radius: 12px;
+            border: 2px dashed rgba(255, 255, 255, 0.22);
+            background: linear-gradient(145deg, rgba(255, 255, 255, 0.09), rgba(255, 255, 255, 0.03));
+            pointer-events: none;
+            transition: border-color 0.35s ease, box-shadow 0.35s ease, opacity 0.35s ease;
+            animation: batch-slot-breathe 2.4s ease-in-out infinite;
+            box-shadow: inset 0 0 0 1px rgba(255, 255, 255, 0.06);
+        }
+        @keyframes batch-slot-breathe {
+            0%, 100% { opacity: 0.88; }
+            50% { opacity: 1; }
+        }
+        .batch-image-wrapper .batch-thumb-img-wrap {
+            width: 100%;
+            height: 100%;
+            border-radius: 9px;
+            overflow: hidden;
+            /* 必须让事件落到外层 .batch-image-wrapper，否则 HTML5 drag 无法从 draggable 父级启动 */
+            pointer-events: none;
+        }
+        .batch-image-wrapper .batch-thumb-img {
+            width: 100%;
+            height: 100%;
+            object-fit: cover;
+            display: block;
+            pointer-events: none;
+            user-select: none;
+            -webkit-user-drag: none;
+        }
+        .batch-thumb-remove {
+            position: absolute;
+            top: 3px;
+            right: 3px;
+            z-index: 5;
+            box-sizing: border-box;
+            min-width: 22px;
+            height: 22px;
+            padding: 0 5px;
+            margin: 0;
+            border: 1px solid rgba(255, 255, 255, 0.12);
+            border-radius: 6px;
+            background: rgba(0, 0, 0, 0.5);
+            font-family: inherit;
+            font-size: 14px;
+            font-weight: 400;
+            line-height: 1;
+            color: rgba(255, 255, 255, 0.9);
+            opacity: 0.72;
+            cursor: pointer;
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            transition: background 0.12s, opacity 0.12s, border-color 0.12s;
+            pointer-events: auto;
+        }
+        .batch-image-wrapper:hover .batch-thumb-remove {
+            opacity: 1;
+            background: rgba(0, 0, 0, 0.68);
+            border-color: rgba(255, 255, 255, 0.2);
+        }
+        .batch-thumb-remove:hover {
+            background: rgba(80, 20, 20, 0.75) !important;
+            border-color: rgba(255, 180, 180, 0.35);
+            color: #fff;
+        }
+        .batch-thumb-remove:focus-visible {
+            opacity: 1;
+            outline: 2px solid var(--accent-dim, rgba(120, 160, 255, 0.6));
+            outline-offset: 1px;
+        }
+        .upload-icon { font-size: 18px; margin-bottom: 6px; opacity: 0.45; }
+        .upload-text { font-size: 11px; color: var(--text); }
+        .upload-hint { font-size: 10px; color: var(--text-dim); margin-top: 3px; }
+        .preview-thumb { width: 100%; height: auto; max-height: 100px; object-fit: contain; border-radius: 8px; display: none; margin-top: 10px; }
+        .clear-img-overlay {
+            position: absolute; top: 8px; right: 8px; background: rgba(255,59,48,0.85); color: white;
+            width: 20px; height: 20px; border-radius: 10px; display: none; align-items: center; justify-content: center;
+            font-size: 11px; cursor: pointer; z-index: 5;
+        }
+
+        .btn-outline {
+            background: var(--panel-2);
+            border: 1px solid var(--border);
+            color: var(--text-sub); padding: 5px 12px; border-radius: 7px;
+            font-size: 11.5px; font-weight: 600; cursor: pointer;
+            transition: background 0.15s, border-color 0.15s, color 0.15s;
+            display: inline-flex; align-items: center; justify-content: center; gap: 5px;
+            white-space: nowrap;
+        }
+        .btn-outline:hover:not(:disabled) { background: rgba(255,255,255,0.08); color: var(--text); border-color: rgba(255,255,255,0.18); }
+        .btn-outline:active { opacity: 0.7; }
+        .btn-outline:disabled { opacity: 0.3; cursor: not-allowed; }
+
+        .btn-icon {
+            padding: 5px; background: transparent; border: none; color: var(--text-dim);
+            border-radius: 6px; cursor: pointer; display: flex; align-items: center; justify-content: center;
+            transition: color 0.15s, background 0.15s;
+        }
+        .btn-icon:hover { color: var(--text-sub); background: rgba(255,255,255,0.07); }
+
+        .btn-primary {
+            width: 100%; padding: 13px;
+            background: var(--accent); border: none;
+            border-radius: 9px; color: #fff; font-weight: 700; font-size: 13.5px;
+            letter-spacing: 0.2px; cursor: pointer; margin-top: 14px;
+            transition: background 0.15s;
+        }
+        .btn-primary:hover:not(:disabled) { background: var(--accent-hover); }
+        .btn-primary:active { opacity: 0.82; }
+        .btn-primary:disabled { background: rgba(255,255,255,0.08); color: var(--text-dim); cursor: not-allowed; }
+
+        .btn-danger {
+            width: 100%; padding: 12px; background: #DC2626; border: none;
+            border-radius: 9px; color: #fff; font-weight: 700; font-size: 13.5px;
+            cursor: pointer; margin-top: 8px; display: none; transition: background 0.15s;
+        }
+        .btn-danger:hover { background: #EF4444; }
+
+        /* Workspace */
+        .workspace { flex: 1; display: flex; flex-direction: column; background: #0A0A0A; position: relative; overflow: hidden; }
+        .viewer { flex: 2; display: flex; align-items: center; justify-content: center; padding: 16px; background: #0A0A0A; position: relative; min-height: 40vh; }
+        .monitor {
+            width: 100%; height: 100%; max-width: 1650px; border-radius: 10px; border: 1px solid var(--border);
+            overflow: hidden; position: relative; background: #070707;
+            display: flex; align-items: center; justify-content: center;
+            background-image: radial-gradient(rgba(255,255,255,0.02) 1px, transparent 1px);
+            background-size: 18px 18px;
+        }
+        .monitor img, .monitor video {
+            width: auto; height: auto; max-width: 100%; max-height: 100%;
+            object-fit: contain; display: none; z-index: 2; border-radius: 3px;
+        }
+
+        .progress-container { position: absolute; bottom: 0; left: 0; width: 100%; height: 2px; background: var(--border-2); z-index: 10; }
+        #progress-fill { width: 0%; height: 100%; background: var(--accent); transition: width 0.5s; }
+        #loading-txt { font-size: 12px; color: var(--text-sub); font-weight: 600; z-index: 5; position: absolute; display: none; }
+
+
+
+        .spinner {
+            width: 12px; height: 12px;
+            border: 2px solid rgba(255,255,255,0.2);
+            border-top-color: currentColor;
+            border-radius: 50%;
+            animation: spin 1s linear infinite;
+        }
+        @keyframes spin { to { transform: rotate(360deg); } }
+        
+        .loading-card {
+            display: flex; align-items: center; justify-content: center;
+            flex-direction: column; gap: 6px; color: var(--text-dim); font-size: 10px;
+            background: rgba(37,99,235,0.07) !important;
+            border-color: rgba(37,99,235,0.3) !important;
+        }
+        .loading-card .spinner { width: 28px; height: 28px; border-width: 3px; color: var(--accent); }
+        .loading-card:hover { background: rgba(37,99,235,0.14) !important; border-color: var(--accent) !important; }
+
+        .library { flex: 1.5; border-top: 1px solid var(--border); padding: 14px 20px; display: flex; flex-direction: column; background: #0F0F11; overflow-y: hidden; }
+        #log-container { flex: 1; overflow-y: auto; padding-right: 4px; }
+        #log { font-family: ui-monospace, "SF Mono", monospace; font-size: 10.5px; color: var(--text-dim); line-height: 1.7; }
+
+        /* History wrapper: scrollable area for thumbnails only */
+        #history-wrapper {
+            flex: 1;
+            overflow-y: auto;
+            min-height: 110px;   /* always show at least one row */
+            padding-right: 4px;
+        }
+        #history-container {
+            display: grid;
+            grid-template-columns: repeat(auto-fill, minmax(150px, 1fr));
+            justify-content: start;
+            gap: 10px; align-content: flex-start;
+            padding-bottom: 4px;
+        }
+        /* Pagination row: hidden, using infinite scroll instead */
+        #pagination-bar {
+            display: none;
+        }
+
+        .history-card {
+            width: 100%; max-width: 200px; aspect-ratio: 16 / 9;
+            background: #1A1A1E; border-radius: 7px;
+            overflow: hidden; border: 1px solid var(--border);
+            cursor: pointer; position: relative; transition: border-color 0.15s, transform 0.15s;
+        }
+        .history-card:hover { border-color: var(--accent); transform: translateY(-1px); }
+        .history-card img, .history-card video {
+            width: 100%; height: 100%; object-fit: cover;
+            background: #1A1A1E;
+        }
+        .history-audio-thumb {
+            width: 100%; height: 100%; display: flex; flex-direction: column;
+            align-items: center; justify-content: center; gap: 6px;
+            color: var(--text); background: linear-gradient(135deg, rgba(35,35,40,0.95), rgba(18,18,22,0.98));
+            font-size: 11px; font-weight: 700; text-align: center; padding: 10px; box-sizing: border-box;
+        }
+        .history-audio-icon { font-size: 24px; line-height: 1; }
+        #audio-wrapper {
+            position: absolute; inset: 0; width: 100%; height: 100%; min-height: 0;
+            display: none; flex-direction: column; align-items: stretch; justify-content: stretch; z-index: 2;
+            background: #070707;
+            background-image: radial-gradient(rgba(255,255,255,0.035) 1px, transparent 1px);
+            background-size: 18px 18px;
+        }
+        .audio-preview-art {
+            position: absolute; inset: 0 0 54px 0; display: flex; flex-direction: column;
+            align-items: center; justify-content: center; gap: 12px; padding: 24px;
+            color: var(--text); cursor: pointer; user-select: none; z-index: 2;
+        }
+        .audio-preview-icon {
+            width: 86px; height: 86px; display: flex; align-items: center; justify-content: center;
+            border: 1px solid var(--border); border-radius: 50%; background: rgba(255,255,255,0.04);
+            font-size: 46px; font-weight: 800; color: var(--accent);
+        }
+        #audio-preview-title {
+            max-width: min(620px, 86%); font-size: 13px; font-weight: 800; color: var(--text);
+            overflow: hidden; text-overflow: ellipsis; white-space: nowrap; text-align: center;
+        }
+        #audio-wrapper .plyr {
+            position: absolute; left: 0; right: 0; bottom: 0; top: auto;
+            width: 100%; height: auto !important; min-height: 0; z-index: 3;
+            border-radius: 0;
+        }
+        #audio-wrapper .plyr--audio .plyr__controls {
+            border-radius: 0; border-top: 1px solid var(--border); background: rgba(12,12,14,0.96);
+        }
+        #res-audio { width: 100%; }
+        .preview-download-btn {
+            position: absolute;
+            top: 14px;
+            right: 14px;
+            z-index: 6;
+            display: inline-flex;
+            align-items: center;
+            gap: 8px;
+            padding: 8px 12px;
+            border-radius: 999px;
+            border: 1px solid rgba(255,255,255,0.16);
+            background: rgba(10,10,12,0.72);
+            color: #fff;
+            font-size: 11px;
+            font-weight: 800;
+            letter-spacing: 0.02em;
+            cursor: pointer;
+            backdrop-filter: blur(10px);
+            box-shadow: 0 8px 24px rgba(0,0,0,0.24);
+            transition: transform 0.18s ease, background 0.18s ease, border-color 0.18s ease, box-shadow 0.18s ease;
+        }
+        .preview-download-btn:hover {
+            transform: translateY(-1px);
+            background: rgba(18,18,22,0.9);
+            border-color: rgba(92,214,143,0.4);
+            box-shadow: 0 10px 28px rgba(0,0,0,0.3);
+        }
+        .preview-download-btn:active {
+            transform: translateY(0);
+            background: rgba(26,26,30,0.96);
+        }
+        .preview-download-btn-icon {
+            width: 24px;
+            height: 24px;
+            display: inline-flex;
+            align-items: center;
+            justify-content: center;
+            border-radius: 999px;
+            background: rgba(255,255,255,0.08);
+            color: var(--accent);
+            flex-shrink: 0;
+        }
+        .preview-download-btn-text {
+            line-height: 1;
+        }
+        .preview-replay-actions {
+            position: absolute;
+            top: 14px;
+            left: 14px;
+            z-index: 80;
+            display: inline-flex;
+            gap: 8px;
+            align-items: center;
+            pointer-events: auto;
+        }
+        .preview-replay-actions button {
+            height: 34px;
+            padding: 0 12px;
+            border-radius: 999px;
+            border: 1px solid rgba(255,255,255,0.16);
+            background: rgba(10,10,12,0.72);
+            color: #fff;
+            font-size: 11px;
+            font-weight: 800;
+            cursor: pointer;
+            backdrop-filter: blur(10px);
+            pointer-events: auto;
+        }
+        .preview-replay-actions button:hover {
+            border-color: rgba(92,214,143,0.42);
+            background: rgba(18,32,24,0.9);
+            color: var(--accent);
+        }
+        .preview-replay-actions.is-unavailable button {
+            opacity: 0.42;
+            color: var(--text-dim);
+            border-color: rgba(255,255,255,0.08);
+            background: rgba(10,10,12,0.54);
+        }
+        .seed-panel {
+            margin-top: 10px;
+            padding: 10px;
+            border-radius: 10px;
+            border: 1px solid var(--border-2);
+            background: rgba(255,255,255,0.025);
+        }
+        .seed-panel-head {
+            display: flex;
+            align-items: center;
+            margin-bottom: 8px;
+            color: var(--text-dim);
+            font-size: 10px;
+            font-weight: 800;
+            letter-spacing: 0.6px;
+            text-transform: uppercase;
+        }
+        .seed-control {
+            display: grid;
+            grid-template-columns: minmax(0, 1fr) 112px;
+            gap: 8px;
+            align-items: stretch;
+        }
+        .seed-input-shell {
+            height: 36px;
+            display: flex;
+            align-items: center;
+            border-radius: 8px;
+            background: var(--panel-2);
+            border: 1px solid var(--border);
+            overflow: hidden;
+            transition: border-color 0.15s, box-shadow 0.15s, opacity 0.15s;
+        }
+        .seed-input-shell:focus-within {
+            border-color: var(--accent);
+            box-shadow: 0 0 0 2px var(--accent-ring);
+        }
+        #seed-value {
+            height: 100%;
+            width: 100%;
+            margin: 0;
+            border: 0;
+            border-radius: 0;
+            background: transparent;
+            box-shadow: none;
+            color: var(--text);
+            font-family: ui-monospace, "SF Mono", monospace;
+            font-size: 12px;
+            font-weight: 700;
+            font-variant-numeric: tabular-nums;
+            padding: 0 11px;
+        }
+        #seed-value:focus { box-shadow: none; }
+        .seed-input-shell.is-random #seed-value {
+            color: var(--text-dim);
+        }
+        .seed-mode-tabs {
+            height: 36px;
+            display: grid;
+            grid-template-columns: 1fr 1fr;
+            gap: 3px;
+            padding: 3px;
+            border-radius: 8px;
+            border: 1px solid var(--border);
+            background: rgba(255,255,255,0.04);
+        }
+        .seed-mode-option {
+            margin: 0;
+            height: 28px;
+            display: inline-flex;
+            align-items: center;
+            justify-content: center;
+            border-radius: 6px;
+            color: var(--text-dim);
+            cursor: pointer;
+            font-size: 10.5px;
+            font-weight: 800;
+            letter-spacing: 0;
+            text-transform: none;
+            transition: background 0.15s, color 0.15s;
+        }
+        .seed-mode-option input { display: none; }
+        .seed-mode-option.is-active {
+            background: var(--accent);
+            color: #fff;
+        }
+        .seed-mode-option:not(.is-active):hover {
+            background: rgba(255,255,255,0.06);
+            color: var(--text-sub);
+        }
+        /* 解码/加载完成前避免视频黑块猛闪，与卡片底色一致；就绪后淡入 */
+        .history-card .history-thumb-media {
+            opacity: 0;
+            transition: opacity 0.28s ease;
+        }
+        .history-card .history-thumb-media.history-thumb-ready {
+            opacity: 1;
+        }
+        .history-type-badge {
+            position: absolute; top: 5px; left: 5px; font-size: 8px; padding: 1px 5px; border-radius: 3px;
+            background: rgba(0,0,0,0.8); color: var(--text-sub); border: 1px solid rgba(255,255,255,0.06);
+            z-index: 2; font-weight: 700; letter-spacing: 0.4px;
+        }
+        .history-delete-btn {
+            position: absolute; top: 5px; right: 5px; width: 20px; height: 20px;
+            border-radius: 50%; border: none; background: rgba(255,50,50,0.8); color: #fff;
+            font-size: 10px; cursor: pointer; z-index: 3; display: flex; align-items: center; justify-content: center;
+            opacity: 0; transition: opacity 0.2s;
+        }
+        .history-card:hover .history-delete-btn { opacity: 1; }
+        .history-delete-btn:hover { background: rgba(255,0,0,0.9); }
+        .vram-bar { width: 160px; height: 5px; background: rgba(255,255,255,0.08); border-radius: 999px; overflow: hidden; display: inline-block; vertical-align: middle; }
+        .vram-used { height: 100%; background: var(--accent); width: 0%; transition: width 0.5s; }
+
+        /* 智能多帧：工作流模式卡片式单选 */
+        .smart-param-mode-label {
+            font-size: 10px;
+            color: var(--text-dim);
+            font-weight: 700;
+            margin-bottom: 8px;
+            letter-spacing: 0.04em;
+            text-transform: uppercase;
+        }
+        .smart-param-modes {
+            display: flex;
+            flex-direction: row;
+            align-items: stretch;
+            gap: 0;
+            padding: 3px;
+            margin-bottom: 12px;
+            background: var(--panel-2);
+            border-radius: 8px;
+            border: 1px solid var(--border);
+        }
+        .smart-param-mode-opt {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            flex: 1;
+            min-width: 0;
+            gap: 0;
+            margin: 0;
+            padding: 6px 8px;
+            border-radius: 6px;
+            border: none;
+            background: transparent;
+            cursor: pointer;
+            transition: background 0.15s, color 0.15s;
+            position: relative;
+        }
+        .smart-param-mode-opt:hover:not(:has(input:checked)) {
+            background: rgba(255, 255, 255, 0.05);
+        }
+        .smart-param-mode-opt input[type="radio"] {
+            position: absolute;
+            opacity: 0;
+            width: 0;
+            height: 0;
+            margin: 0;
+        }
+        .smart-param-mode-opt:has(input:checked) {
+            background: var(--accent);
+            box-shadow: none;
+        }
+        .smart-param-mode-opt:has(input:checked) .smart-param-mode-title {
+            color: #fff;
+        }
+        .smart-param-mode-title {
+            font-size: 11px;
+            font-weight: 600;
+            color: var(--text-sub);
+            text-align: center;
+            line-height: 1.25;
+            flex: none;
+            min-width: 0;
+        }
+        /* 单次多关键帧：时间轴面板 */
+        .batch-kf-panel {
+            background: var(--item);
+            border-radius: 10px;
+            padding: 12px 14px;
+            margin-bottom: 10px;
+            border: 1px solid var(--border);
+        }
+        .batch-kf-panel-hd {
+            display: flex;
+            flex-wrap: wrap;
+            align-items: center;
+            justify-content: space-between;
+            gap: 10px;
+            margin-bottom: 8px;
+        }
+        .batch-kf-panel-title {
+            font-size: 12px;
+            font-weight: 700;
+            color: var(--text);
+        }
+        .batch-kf-total-pill {
+            font-size: 11px;
+            color: var(--text-sub);
+            background: var(--panel-2);
+            border: 1px solid var(--border);
+            border-radius: 999px;
+            padding: 6px 12px;
+            white-space: nowrap;
+        }
+        .batch-kf-total-pill strong {
+            color: var(--accent);
+            font-weight: 800;
+            font-variant-numeric: tabular-nums;
+            margin: 0 2px;
+        }
+        .batch-kf-total-unit {
+            font-size: 10px;
+            color: var(--text-dim);
+        }
+        .batch-kf-panel-hint {
+            font-size: 10px;
+            color: var(--text-dim);
+            line-height: 1.5;
+            margin: 0 0 12px;
+        }
+        .batch-kf-timeline-col {
+            display: flex;
+            flex-direction: column;
+            gap: 0;
+        }
+        .batch-kf-kcard {
+            border-radius: 10px;
+            border: 1px solid var(--border);
+            background: rgba(255, 255, 255, 0.03);
+            padding: 10px 12px;
+        }
+        .batch-kf-kcard-head {
+            display: flex;
+            align-items: center;
+            gap: 12px;
+            margin-bottom: 10px;
+        }
+        .batch-kf-kthumb {
+            width: 48px;
+            height: 48px;
+            border-radius: 8px;
+            object-fit: cover;
+            flex-shrink: 0;
+            border: 1px solid var(--border);
+        }
+        .batch-kf-kcard-titles {
+            display: flex;
+            flex-direction: column;
+            gap: 4px;
+            min-width: 0;
+        }
+        .batch-kf-ktitle {
+            font-size: 12px;
+            font-weight: 700;
+            color: var(--text);
+        }
+        .batch-kf-anchor {
+            font-size: 11px;
+            color: var(--accent);
+            font-variant-numeric: tabular-nums;
+            font-weight: 600;
+        }
+        .batch-kf-kcard-ctrl {
+            display: flex;
+            flex-wrap: wrap;
+            align-items: center;
+            gap: 12px;
+        }
+        .batch-kf-klabel {
+            font-size: 10px;
+            color: var(--text-dim);
+            display: flex;
+            align-items: center;
+            gap: 8px;
+        }
+        .batch-kf-klabel input[type="number"] {
+            width: 72px;
+            padding: 6px 8px;
+            font-size: 12px;
+            border-radius: 6px;
+            border: 1px solid var(--border);
+            background: var(--panel);
+            color: var(--text);
+        }
+        /* 关键帧之间：细时间轴 + 单行紧凑间隔输入 */
+        .batch-kf-gap {
+            display: flex;
+            align-items: stretch;
+            gap: 8px;
+            padding: 0 0 6px;
+            margin: 0 0 0 10px;
+        }
+        .batch-kf-gap-rail {
+            width: 2px;
+            flex-shrink: 0;
+            border-radius: 2px;
+            background: linear-gradient(
+                180deg,
+                rgba(255, 255, 255, 0.06),
+                var(--accent-dim),
+                rgba(255, 255, 255, 0.04)
+            );
+            min-height: 22px;
+            align-self: stretch;
+        }
+        .batch-kf-gap-inner {
+            display: flex;
+            align-items: center;
+            gap: 8px;
+            flex: 1;
+            min-width: 0;
+            padding: 2px 0 4px;
+        }
+        .batch-kf-gap-ix {
+            font-size: 10px;
+            font-weight: 600;
+            color: var(--text-dim);
+            font-variant-numeric: tabular-nums;
+            letter-spacing: -0.02em;
+            flex-shrink: 0;
+        }
+        .batch-kf-seg-field {
+            display: inline-flex;
+            align-items: center;
+            gap: 3px;
+            margin: 0;
+            cursor: text;
+        }
+        .batch-kf-seg-input {
+            width: 46px;
+            min-width: 0;
+            padding: 2px 5px;
+            font-size: 11px;
+            font-weight: 600;
+            line-height: 1.3;
+            border-radius: 4px;
+            border: 1px solid var(--border);
+            background: rgba(0, 0, 0, 0.2);
+            color: var(--text);
+            font-variant-numeric: tabular-nums;
+        }
+        .batch-kf-seg-input:hover {
+            border-color: rgba(255, 255, 255, 0.12);
+        }
+        .batch-kf-seg-input:focus {
+            outline: none;
+            border-color: var(--accent);
+            box-shadow: 0 0 0 1px var(--accent-ring);
+        }
+        .batch-kf-gap-unit {
+            font-size: 10px;
+            color: var(--text-dim);
+            font-weight: 500;
+            flex-shrink: 0;
+        }
+
+        .sub-mode-toggle { display: flex; background: var(--panel-2); border-radius: 7px; padding: 3px; border: 1px solid var(--border); }
+        .sub-mode-btn { flex: 1; padding: 6px 0; border-radius: 5px; border: none; background: transparent; font-size: 11.5px; color: var(--text-dim); font-weight: 600; cursor: pointer; transition: background 0.15s, color 0.15s; }
+        .sub-mode-btn.active { background: var(--accent); color: #fff; }
+        .sub-mode-btn:hover:not(.active) { background: rgba(255,255,255,0.05); color: var(--text-sub); }
+
+        .vid-section { display: none; margin-top: 12px; }
+        .vid-section.active-section { display: block; animation: fadeIn 0.25s ease; }
+        @keyframes fadeIn { from { opacity: 0; transform: translateY(4px); } to { opacity: 1; transform: translateY(0); } }
+
+        /* Status indicator */
+        @keyframes breathe-orange {
+            0%,100% { box-shadow: 0 0 4px #FF9F0A; opacity: 0.7; }
+            50%      { box-shadow: 0 0 10px #FF9F0A; opacity: 1; }
+        }
+        .indicator-busy    { background: #FF9F0A !important; animation: breathe-orange 1.6s infinite ease-in-out !important; box-shadow: none !important; transition: all 0.3s; }
+        .indicator-ready   { background: #30D158 !important; box-shadow: 0 0 8px rgba(48,209,88,0.6) !important; animation: none !important; transition: all 0.3s; }
+        .indicator-offline { background: #636366 !important; box-shadow: none !important; animation: none !important; transition: all 0.3s; }
+
+        .res-preview-tag { font-size: 11px; color: var(--accent); margin-bottom: 10px; font-family: ui-monospace, monospace; }
+        .top-status { display: flex; justify-content: space-between; font-size: 12px; color: var(--text-dim); margin-bottom: 8px; align-items: center; }
+        .checkbox-container { display: flex; align-items: center; gap: 8px; cursor: pointer; background: rgba(255,255,255,0.02); padding: 10px; border-radius: 8px; border: 1px solid var(--border-2); }
+        .checkbox-container input { width: 15px; height: 15px; accent-color: var(--accent); cursor: pointer; margin: 0; }
+        .checkbox-container label { margin-bottom: 0; cursor: pointer; text-transform: none; color: var(--text); }
+        .flex-row { display: flex; gap: 10px; }
+        .flex-1 { flex: 1; min-width: 0; }
+
+        @media (max-width: 1024px) {
+            body { flex-direction: column; overflow-y: auto; }
+            .sidebar { width: 100%; min-width: 100%; border-right: none; border-bottom: 1px solid var(--border); height: auto; overflow: visible; }
+            .workspace { height: auto; min-height: 100vh; overflow: visible; }
+        }
+:root {
+    --plyr-color-main: #3F51B5;
+    --plyr-video-control-background-hover: rgba(255,255,255,0.1);
+    --plyr-control-radius: 6px;
+    --plyr-player-width: 100%;
+}
+.plyr {
+    border-radius: 8px;
+    overflow: hidden;
+    width: 100%;
+    height: 100%;
+}
+.plyr--video .plyr__controls {
+    background: linear-gradient(rgba(0,0,0,0), rgba(0,0,0,0.8));
+    padding: 20px 15px 15px 15px;
+}

LTX2.3-1.0.4-new/UI/index.html

@@ -0,0 +1,604 @@
+<!DOCTYPE html>
+<html lang="zh-CN">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>LTX-2 | Multi-GPU Cinematic Studio</title>
+    <link rel="stylesheet" href="index.css?v=en-tabs-1">
+        <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/plyr/3.7.8/plyr.css" />
+</head>
+<body>
+
+<aside class="sidebar">
+    <div class="sidebar-header">
+        <div style="display: flex; align-items: center; justify-content: space-between; margin-bottom: 12px;">
+            <div style="display: flex; align-items: center; gap: 10px;">
+                <div id="sys-indicator" class="indicator-ready" style="width: 12px; height: 12px; border-radius: 50%;"></div>
+                <span style="font-weight: 800; font-size: 18px;">LTX-2 STUDIO</span>
+            </div>
+            <div style="display: flex; gap: 8px; align-items: center;">
+                <button id="clearGpuBtn" onclick="clearGpu()" class="btn-outline" data-i18n="clearVram">释放显存</button>
+                <button type="button" id="lang-toggle-btn" class="lang-toggle">EN</button>
+            </div>
+        </div>
+        
+        <div class="top-status" style="margin-bottom: 5px;">
+            <div style="display: flex; align-items: center; gap: 8px;">
+                <span id="sys-status" style="font-weight:bold; color: var(--text-dim); font-size: 12px;" data-i18n="sysScanning">正在扫描 GPU...</span>
+            </div>
+            
+            <button type="button" onclick="const el = document.getElementById('sys-settings'); el.style.display = el.style.display === 'none' ? 'block' : 'none';" class="btn-icon" data-i18n-title="settingsTitle" title="系统高级设置">
+                <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="3"></circle><path d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 0 1 0 2.83 2 2 0 0 1-2.83 0l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-2 2 2 2 0 0 1-2-2v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 0 1-2.83 0 2 2 0 0 1 0-2.83l.06-.06a1.65 1.65 0 0 0 .33-1.82 1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1-2-2 2 2 0 0 1 2-2h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 0 1 0-2.83 2 2 0 0 1 2.83 0l.06.06a1.65 1.65 0 0 0 1.82.33H9a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 2-2 2 2 0 0 1 2 2v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 0 1 2.83 0 2 2 0 0 1 0 2.83l-.06.06a1.65 1.65 0 0 0-.33 1.82V9a1.65 1.65 0 0 0 1.51 1H21a2 2 0 0 1 2 2 2 2 0 0 1-2 2h-.09a1.65 1.65 0 0 0-1.51 1z"></path></svg>
+            </button>
+
+        </div>
+        
+        <div style="font-size: 11px; color: var(--text-dim); margin-bottom: 20px; display: flex; align-items: center; width: 100%;">
+            <div class="vram-bar" style="width: 120px; min-width: 120px; margin-top: 0; margin-right: 12px;"><div class="vram-used" id="vram-fill"></div></div> 
+            <span id="vram-text" style="font-variant-numeric: tabular-nums; flex-shrink: 0; text-align: right;">0/32 GB</span>
+            <span id="gpu-name" style="display: none;"></span> <!-- Hidden globally to avoid duplicate -->
+        </div>
+        
+        <div id="sys-settings" style="display: none; padding: 14px; background: rgba(0,0,0,0.4) !important; border-radius: 12px; border: 1px solid rgba(255,255,255,0.1); margin-bottom: 15px; box-shadow: 0 4px 15px rgba(0,0,0,0.5); backdrop-filter: blur(10px);">
+            <div style="font-size: 13px; font-weight: bold; margin-bottom: 12px; color: #fff;" data-i18n="advancedSettings">高级设置</div>
+            
+            <label style="font-size: 11px; margin-bottom: 6px;" data-i18n="deviceSelect">工作设备选择</label>
+            <select id="gpu-selector" onchange="switchGpu(this.value)" style="margin-bottom: 12px; font-size: 11px; padding: 6px;">
+                <option value="" data-i18n="gpuDetecting">正在检测 GPU...</option>
+            </select>
+
+            <label style="font-size: 11px; margin-bottom: 6px; margin-top: 12px;" data-i18n="vramLimitLabel">可用最高显存上限 (GB, 0为全开优先显存)</label>
+            <div style="display: flex; gap: 6px; margin-bottom: 9px; align-items: stretch;">
+                <input type="number" id="vram-limit-input" value="0" data-i18n-placeholder="vramLimitPh" placeholder="例如: 12 (0表示无限制)" style="flex: 1; height: 28px; box-sizing: border-box; font-size: 12px; padding: 0 10px;">
+                <button onclick="saveVramLimit()" style="font-size: 12px; padding: 0 10px; height: 28px; box-sizing: border-box; white-space: nowrap; background: #333; border: 1px solid #555; color: #fff; border-radius: 7px; cursor: pointer;" data-i18n="saveLabel">保存</button>
+            </div>
+            <div id="vram-limit-status" style="font-size: 10px; color: var(--text-dim);"></div>
+
+            <label style="font-size: 11px; margin-bottom: 6px; margin-top: 12px;" data-i18n="modelCheckpointLabel">视频模型（蒸馏版）</label>
+            <div style="display: flex; gap: 6px; margin-bottom: 9px; align-items: stretch;">
+                <select id="model-checkpoint-select" onchange="saveSelectedModelCheckpoint()" style="flex: 1; height: 28px; box-sizing: border-box; font-size: 12px; padding: 0 8px;">
+                    <option value="" data-i18n="modelCheckpointDefault">默认官方蒸馏模型</option>
+                </select>
+                <button onclick="loadModelCheckpoints()" style="font-size: 12px; padding: 0 10px; height: 28px; box-sizing: border-box; white-space: nowrap; background: #333; border: 1px solid #555; color: #fff; border-radius: 7px; cursor: pointer;" data-i18n="refresh">刷新</button>
+            </div>
+            <div id="model-checkpoint-status" style="font-size: 10px; color: var(--text-dim); margin-bottom: 6px;" data-i18n="modelCheckpointHint">推荐使用 distilled-fp8；仅显示 LTX 2.3 22B 蒸馏模型，避开 dev 模型。</div>
+            
+            
+            <label style="font-size: 11px; margin-bottom: 6px; margin-top: 12px;" data-i18n="loraFolderPath">LoRA 文件夹路径（可选）</label>
+            <div style="display: flex; gap: 6px; margin-bottom: 9px; align-items: stretch;">
+                <input type="text" id="lora-dir-input" placeholder="留空使用 模型目录\\loras" data-i18n-placeholder="loraFolderPathPlaceholder" style="flex: 1; height: 28px; box-sizing: border-box; font-size: 12px; padding: 0 10px;">
+                <button onclick="saveLoraDir()" style="font-size: 12px; padding: 0 10px; height: 28px; box-sizing: border-box; white-space: nowrap; background: #333; border: 1px solid #555; color: #fff; border-radius: 7px; cursor: pointer;" data-i18n="saveLabel">保存</button>
+            </div>
+            <div id="lora-placement-hint" style="font-size: 10px; color: var(--text-dim); line-height: 1.45; margin-bottom: 6px;" data-i18n="loraPlacementHint">将 LoRA 文件放到当前模型目录下的 loras 文件夹。</div>
+            <div id="lora-dir-status" style="font-size: 10px; color: var(--text-dim);"></div>
+        </div>
+
+        <div id="queue-panel" style="padding: 12px; background: rgba(255,255,255,0.03); border: 1px solid rgba(255,255,255,0.08); border-radius: 12px; margin-bottom: 14px;">
+            <div style="display:flex; align-items:center; justify-content:space-between; gap:8px; margin-bottom:8px;">
+                <div style="font-size:12px; font-weight:800; color:var(--text-main);" data-i18n="queueTitle">任务队列</div>
+                <div id="queue-summary" style="font-size:10px; color:var(--text-dim);" data-i18n="queueIdle">空闲</div>
+            </div>
+            <div id="queue-list" style="display:flex; flex-direction:column; gap:6px;"></div>
+        </div>
+    </div>
+
+    <div class="sidebar-section" id="main-tabs-section">
+        <div class="tabs">
+            <div id="tab-video" class="tab" onclick="switchMode('video')">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right: 6px;"><rect x="2" y="2" width="20" height="20" rx="2.18" ry="2.18"></rect><line x1="7" y1="2" x2="7" y2="22"></line><line x1="17" y1="2" x2="17" y2="22"></line><line x1="2" y1="12" x2="22" y2="12"></line><line x1="2" y1="7" x2="7" y2="7"></line><line x1="2" y1="17" x2="7" y2="17"></line><line x1="17" y1="17" x2="22" y2="17"></line><line x1="17" y1="7" x2="22" y2="7"></line></svg>
+                <span data-i18n="tabVideo">视频生成</span>
+            </div>
+            <div id="tab-batch" class="tab" onclick="switchMode('batch')">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right: 6px;"><rect x="3" y="3" width="7" height="7"></rect><rect x="14" y="3" width="7" height="7"></rect><rect x="14" y="14" width="7" height="7"></rect><rect x="3" y="14" width="7" height="7"></rect></svg>
+                <span data-i18n="tabBatch">智能多帧</span>
+            </div>
+            <div id="tab-motion" class="tab" onclick="switchMode('motion')">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right: 6px;"><path d="M3 12h5l3-8 4 16 3-8h3"></path></svg>
+                <span data-i18n="tabMotion">视频迁移</span>
+            </div>
+            <div id="tab-image" class="tab" onclick="switchMode('image')">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right: 6px;"><rect x="3" y="3" width="18" height="18" rx="2" ry="2"></rect><circle cx="8.5" cy="8.5" r="1.5"></circle><polyline points="21 15 16 10 5 21"></polyline></svg>
+                <span data-i18n="tabImage">图像生成</span>
+            </div>
+            <div id="tab-tts" class="tab" onclick="switchMode('tts')">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="margin-right: 6px;"><path d="M12 1a3 3 0 0 0-3 3v8a3 3 0 0 0 6 0V4a3 3 0 0 0-3-3z"></path><path d="M19 10v2a7 7 0 0 1-14 0v-2"></path><line x1="12" y1="19" x2="12" y2="23"></line><line x1="8" y1="23" x2="16" y2="23"></line></svg>
+                <span data-i18n="tabTts">TTS 语音</span>
+            </div>
+        </div>
+        
+        <div id="prompt-container">
+            <label data-i18n="promptLabel">视觉描述词 (Prompt)</label>
+            <textarea id="prompt" data-i18n-placeholder="promptPlaceholder" placeholder="在此输入视觉描述词 (Prompt)..." style="height: 90px; margin-bottom: 0;"></textarea>
+        </div>
+
+        <div id="seed-settings" class="seed-panel">
+            <div class="seed-panel-head">
+                <span data-i18n="seedLabel">随机种子 (Seed)</span>
+            </div>
+            <div class="seed-control">
+                <div class="seed-input-shell">
+                    <input type="number" id="seed-value" min="1" max="2147483647" value="42" disabled>
+                </div>
+                <div class="seed-mode-tabs" role="radiogroup" aria-label="Seed mode">
+                    <label class="seed-mode-option is-active">
+                        <input type="radio" name="seed-mode" value="random" checked onchange="updateSeedModeUI()">
+                        <span data-i18n="seedRandom">随机</span>
+                    </label>
+                    <label class="seed-mode-option">
+                        <input type="radio" name="seed-mode" value="fixed" onchange="updateSeedModeUI()">
+                        <span data-i18n="seedFixed">固定</span>
+                    </label>
+                </div>
+            </div>
+        </div>
+    </div>
+
+    <!-- 视频模式选项 -->
+    <div class="sidebar-section" id="video-opts" style="display:none">
+        <div class="setting-group">
+            <div class="group-title" data-i18n="basicEngine">基础画面 / Basic EngineSpecs</div>
+            <div class="flex-row">
+                <div class="flex-1">
+                    <label data-i18n="qualityLevel">清晰度级别</label>
+                    <select id="vid-quality" onchange="updateResPreview()">
+                        <option value="1080">1080P Full HD</option>
+                        <option value="720" selected>720P Standard</option>
+                        <option value="540">540P Preview</option>
+                    </select>
+                </div>
+                <div class="flex-1">
+                    <label data-i18n="aspectRatio">画幅比例</label>
+                    <select id="vid-ratio" onchange="updateResPreview()">
+                        <option value="16:9" data-i18n="ratio169">16:9 电影宽幅</option>
+                        <option value="9:16" data-i18n="ratio916">9:16 移动竖屏</option>
+                        <option value="1:1" data-i18n="ratio11">1:1 方形</option>
+                        <option value="4:3" data-i18n="ratio43">4:3 经典横幅</option>
+                        <option value="3:4" data-i18n="ratio34">3:4 经典竖幅</option>
+                        <option value="21:9" data-i18n="ratio219">21:9 超宽银幕</option>
+                        <option value="9:21" data-i18n="ratio921">9:21 超长竖屏</option>
+                        <option value="ref" data-i18n="ratioRef">跟随参考图</option>
+                        <option value="custom" data-i18n="ratioCustom">自定义尺寸</option>
+                    </select>
+                </div>
+            </div>
+            <div id="vid-custom-size" class="flex-row" style="display:none; margin-top: -2px; margin-bottom: 10px;">
+                <div class="flex-1"><label data-i18n="width">宽度</label><input type="number" id="vid-custom-w" value="1280" min="64" step="64" onchange="updateResPreview()"></div>
+                <div class="flex-1"><label data-i18n="height">高度</label><input type="number" id="vid-custom-h" value="704" min="64" step="64" onchange="updateResPreview()"></div>
+            </div>
+            <div id="res-preview" class="res-preview-tag" style="margin-top: -5px; margin-bottom: 12px;">最终发送: 1280x704</div>
+
+            <div class="flex-row">
+                <div class="flex-1">
+                    <label data-i18n="fpsLabel">帧率 (FPS)</label>
+                    <select id="vid-fps">
+                        <option value="24" selected>24 FPS</option>
+                        <option value="25">25 FPS</option>
+                        <option value="30">30 FPS</option>
+                        <option value="48">48 FPS</option>
+                        <option value="60">60 FPS</option>
+                    </select>
+                </div>
+                <div class="flex-1">
+                    <label data-i18n="durationLabel">时长 (秒)</label>
+                    <input type="number" id="vid-duration" value="5" min="1" max="30" step="1">
+                </div>
+            </div>
+
+            <label style="margin-top: 12px;" data-i18n="cameraMotion">镜头运动方式</label>
+            <select id="vid-motion">
+                <option value="static" selected data-i18n="motionStatic">Static (静止机位)</option>
+                <option value="dolly_in" data-i18n="motionDollyIn">Dolly In (推近)</option>
+                <option value="dolly_out" data-i18n="motionDollyOut">Dolly Out (拉远)</option>
+                <option value="dolly_left" data-i18n="motionDollyLeft">Dolly Left (向左)</option>
+                <option value="dolly_right" data-i18n="motionDollyRight">Dolly Right (向右)</option>
+                <option value="jib_up" data-i18n="motionJibUp">Jib Up (升臂)</option>
+                <option value="jib_down" data-i18n="motionJibDown">Jib Down (降臂)</option>
+                <option value="focus_shift" data-i18n="motionFocus">Focus Shift (焦点)</option>
+            </select>
+            <div class="checkbox-container" style="margin-top: 8px;">
+                <input type="checkbox" id="vid-audio" checked>
+                <label for="vid-audio" data-i18n="audioGen">生成 AI 环境音 (Audio Gen)</label>
+            </div>
+
+            <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 8px;">
+                <label data-i18n="selectLora" style="margin: 0;">选择 LoRA</label>
+                <button type="button" onclick="addLoraSelection('loras-container')" style="background:none; border:none; color:var(--accent); cursor:pointer; font-size:16px; padding:0 4px;" title="添加 LoRA">+</button>
+            </div>
+            <div id="loras-container" style="display: flex; flex-direction: column; gap: 8px; margin-bottom: 8px;"></div>
+        </div>
+
+        <!-- 生成媒介组 -->
+        <div class="setting-group" id="video-source-group">
+            <div class="group-title" data-i18n="genSource">生成媒介 / Generation Source</div>
+            
+            <div class="flex-row" style="margin-bottom: 10px;">
+                <div class="flex-1">
+                    <label data-i18n="startFrame">起始帧 (首帧)</label>
+                    <div class="upload-zone" id="start-frame-drop-zone" onclick="document.getElementById('start-frame-input').click()">
+                        <div class="clear-img-overlay" id="clear-start-frame-overlay" onclick="event.stopPropagation(); clearFrame('start')">×</div>
+                        <div id="start-frame-placeholder">
+                            <div class="upload-icon">🖼️</div>
+                            <div class="upload-text" data-i18n="uploadStart">上传首帧</div>
+                        </div>
+                        <img id="start-frame-preview" class="preview-thumb">
+                        <input type="file" id="start-frame-input" accept="image/*" style="display:none" onchange="handleFrameUpload(this.files[0], 'start')">
+                    </div>
+                    <input type="hidden" id="start-frame-path">
+                </div>
+                <div class="flex-1">
+                    <label data-i18n="endFrame">结束帧 (尾帧)</label>
+                    <div class="upload-zone" id="end-frame-drop-zone" onclick="document.getElementById('end-frame-input').click()">
+                        <div class="clear-img-overlay" id="clear-end-frame-overlay" onclick="event.stopPropagation(); clearFrame('end')">×</div>
+                        <div id="end-frame-placeholder">
+                            <div class="upload-icon">🏁</div>
+                            <div class="upload-text" data-i18n="uploadEnd">上传尾帧 (可选)</div>
+                        </div>
+                        <img id="end-frame-preview" class="preview-thumb">
+                        <input type="file" id="end-frame-input" accept="image/*" style="display:none" onchange="handleFrameUpload(this.files[0], 'end')">
+                    </div>
+                    <input type="hidden" id="end-frame-path">
+                </div>
+            </div>
+
+            <div class="flex-row">
+                <div class="flex-1">
+                    <label data-i18n="refAudio">参考音频 (A2V)</label>
+                    <div class="upload-zone" id="audio-drop-zone" onclick="document.getElementById('vid-audio-input').click()">
+                        <div class="clear-img-overlay" id="clear-audio-overlay" onclick="event.stopPropagation(); clearUploadedAudio()">×</div>
+                        <div id="audio-upload-placeholder">
+                            <div class="upload-icon">🎵</div>
+                            <div class="upload-text" data-i18n="uploadAudio">点击上传音频</div>
+                        </div>
+                        <div id="audio-upload-status" style="display:none;">
+                            <div class="upload-icon" style="color:var(--accent); opacity:1;">✔️</div>
+                            <div id="audio-filename-status" class="upload-text"></div>
+                        </div>
+                        <input type="file" id="vid-audio-input" accept="audio/*" style="display:none" onchange="handleAudioUpload(this.files[0])">
+                    </div>
+                    <input type="hidden" id="uploaded-audio-path">
+                </div>
+            </div>
+            <div style="font-size: 10px; color: var(--text-dim); text-align: center; margin-top: 5px;" data-i18n="sourceHint">
+                💡 若仅上传首帧 = 图生视频/音视频；若同时上传首尾帧 = 首尾插帧。
+            </div>
+        </div>
+    </div>
+
+    <!-- 图像模式选项 -->
+    <div class="sidebar-section" id="motion-opts" style="display:none">
+        <div class="setting-group">
+            <div>
+                <label data-i18n="motionTransferModeLabel">迁移类型</label>
+                <div style="display:grid;grid-template-columns:1fr 1fr 1fr;gap:6px;margin-top:6px;">
+                    <button type="button" id="motion-mode-action" class="btn-outline" onclick="setVideoTransferMode('action')" style="height:32px;background:var(--accent);color:#fff;border-color:var(--accent);" data-i18n="motionModeAction">动作迁移</button>
+                    <button type="button" id="motion-mode-camera" class="btn-outline" onclick="setVideoTransferMode('camera')" style="height:32px;" data-i18n="motionModeCamera">运镜迁移</button>
+                    <button type="button" id="motion-mode-repaint" class="btn-outline" onclick="setVideoTransferMode('repaint')" style="height:32px;" data-i18n="motionModeRepaint">视频重绘</button>
+                </div>
+            </div>
+
+            <div id="motion-action-control-section" style="margin-top: 12px;">
+                <label data-i18n="motionControlType">控制类型</label>
+                <select id="motion-conditioning-type">
+                    <option value="pose" selected data-i18n="motionControlPose">Pose 姿态</option>
+                    <option value="canny" data-i18n="motionControlCanny">Canny 轮廓</option>
+                    <option value="depth" data-i18n="motionControlDepth">Depth 深度</option>
+                    <option value="video" hidden>Video</option>
+                </select>
+            </div>
+
+            <label style="margin-top: 12px;" data-i18n="motionRefVideoLabel">参考动作视频</label>
+            <div class="upload-zone" id="motion-video-drop-zone" onclick="document.getElementById('motion-video-input').click()" style="height: 96px;">
+                <div class="clear-img-overlay" id="clear-motion-video-overlay" onclick="event.stopPropagation(); clearMotionVideo()">×</div>
+                <div id="motion-video-placeholder">
+                    <div class="upload-icon">🎬</div>
+                    <div data-i18n="motionVideoUploadText">点击或拖拽视频</div>
+                    <small data-i18n="motionVideoUploadHint">提取动作/轮廓作为控制</small>
+                </div>
+                <div id="motion-video-status" style="display:none; padding: 10px; text-align:center;">
+                    <div style="font-size: 22px;">🎞️</div>
+                    <div id="motion-video-name" style="font-size: 11px; color: var(--text-sub); word-break: break-all;"></div>
+                </div>
+            </div>
+            <input type="file" id="motion-video-input" accept="video/*" hidden onchange="handleMotionVideoUpload(this.files[0])">
+            <input type="hidden" id="motion-video-path">
+
+            <div id="motion-image-section">
+                <label style="margin-top: 12px;" data-i18n="motionTargetImageLabel">目标主体图</label>
+                <div class="upload-zone" id="motion-image-drop-zone" onclick="document.getElementById('motion-image-input').click()" style="height: 130px;">
+                    <div class="clear-img-overlay" id="clear-motion-image-overlay" onclick="event.stopPropagation(); clearMotionImage()">×</div>
+                    <div id="motion-image-placeholder">
+                        <div class="upload-icon">🖼️</div>
+                        <div data-i18n="motionImageUploadText">点击或拖拽图片</div>
+                        <small data-i18n="motionImageUploadHint">作为主体/首帧引导</small>
+                    </div>
+                    <img id="motion-image-preview" style="display:none; width:100%; height:100%; object-fit:cover; border-radius:10px;">
+                </div>
+                <input type="file" id="motion-image-input" accept="image/*" hidden onchange="handleMotionImageUpload(this.files[0])">
+                <input type="hidden" id="motion-image-path">
+            </div>
+
+            <div style="margin-top: 12px;">
+                <label style="display:flex;justify-content:space-between;align-items:center;">
+                    <span id="motion-strength-label" data-i18n="motionControlStrength">控制强度</span>
+                    <span id="motion-strength-val" style="color:var(--accent);font-size:11px;font-weight:800;">1</span>
+                </label>
+                <input type="range" id="motion-strength" value="1" min="0" max="2" step="0.05" style="width:100%;" oninput="document.getElementById('motion-strength-val').textContent = Number(this.value).toFixed(2).replace(/\.?0+$/, '')">
+            </div>
+        </div>
+    </div>
+
+    <div id="image-opts" class="sidebar-section" style="display:none">
+        <label data-i18n="imgPreset">预设分辨率 (Presets)</label>
+        <select id="img-res-preset" onchange="applyImgPreset(this.value)">
+            <option value="1024x1024" data-i18n="imgOptSquare">1:1 Square (1024x1024)</option>
+            <option value="1280x720" data-i18n="imgOptLand">16:9 Landscape (1280x720)</option>
+            <option value="720x1280" data-i18n="imgOptPort">9:16 Portrait (720x1280)</option>
+            <option value="custom" data-i18n="imgOptCustom">Custom 自定义...</option>
+        </select>
+
+        <div id="img-custom-res" class="flex-row" style="margin-top: 10px;">
+            <div class="flex-1"><label data-i18n="width">宽度</label><input type="number" id="img-w" value="1024" onchange="updateImgResPreview()"></div>
+            <div class="flex-1"><label data-i18n="height">高度</label><input type="number" id="img-h" value="1024" onchange="updateImgResPreview()"></div>
+        </div>
+        <div id="img-res-preview" class="res-preview-tag">最终发送: 1024x1024</div>
+        
+        <div class="label-group" style="margin-top: 15px;">
+            <label data-i18n="samplingSteps">采样步数 (Steps)</label>
+            <span class="val-badge" id="stepsVal">28</span>
+        </div>
+        <div class="slider-container">
+            <input type="range" id="img-steps" min="1" max="50" value="28" oninput="document.getElementById('stepsVal').innerText=this.value">
+        </div>
+    </div>
+
+    <!-- 智能多帧模式 -->
+    <div class="sidebar-section" id="batch-opts" style="display:none">
+        <div class="setting-group">
+            <div class="group-title" data-i18n="smartMultiFrameGroup">智能多帧</div>
+            <div class="smart-param-mode-label" data-i18n="workflowModeLabel">工作流模式（点击切换）</div>
+            <div class="smart-param-modes" role="radiogroup" aria-label="工作流模式">
+                <label class="smart-param-mode-opt">
+                    <input type="radio" name="batch-workflow" value="single" checked onchange="onBatchWorkflowChange()">
+                    <span class="smart-param-mode-title" data-i18n="wfSingle">单次多关键帧</span>
+                </label>
+                <label class="smart-param-mode-opt">
+                    <input type="radio" name="batch-workflow" value="segments" onchange="onBatchWorkflowChange()">
+                    <span class="smart-param-mode-title" data-i18n="wfSegments">分段拼接</span>
+                </label>
+            </div>
+            
+            <label data-i18n="uploadImages">上传图片</label>
+            <div class="upload-zone" id="batch-images-drop-zone" onclick="document.getElementById('batch-images-input').click()" style="min-height: 72px; margin-bottom: 0;">
+                <div id="batch-images-placeholder">
+                    <div class="upload-icon">📁</div>
+                    <div class="upload-text" data-i18n="uploadMulti1">点击或拖入多张图片</div>
+                    <div class="upload-hint" data-i18n="uploadMulti2">支持一次选多张，可多次添加</div>
+                </div>
+                <input type="file" id="batch-images-input" accept="image/*" multiple style="display:none" onchange="handleBatchImagesUpload(this.files, true)">
+            </div>
+            <input type="hidden" id="batch-images-path">
+
+            <div class="batch-thumb-strip-wrap" id="batch-thumb-strip-wrap" style="display: none;">
+                <div class="batch-thumb-strip-head">
+                    <span class="batch-thumb-strip-title" data-i18n="batchStripTitle">已选图片 · 顺序 = 播放先后</span>
+                    <span class="batch-thumb-strip-hint" data-i18n="batchStripHint">在缩略图上按住拖动排序；松手落入虚线框位置</span>
+                </div>
+                <div class="batch-images-container" id="batch-images-container"></div>
+            </div>
+
+            <div style="font-size: 10px; color: var(--text-dim); margin-bottom: 12px; margin-top: 10px; line-height: 1.45;" data-i18n-html="batchFfmpegHint">
+                💡 <strong>分段模式</strong>：2 张 = 1 段；3 张 = 2 段再拼接。<strong>单次模式</strong>：几张图就几个 latent 锚点，一条视频出片。<br>
+                多段需 <code style="font-size:9px;">ffmpeg</code>：装好后加 PATH，或设环境变量 <code style="font-size:9px;">LTX_FFMPEG_PATH</code>，或在 <code style="font-size:9px;">%LOCALAPPDATA%\LTXDesktop\ffmpeg_path.txt</code> 第一行写 ffmpeg.exe 完整路径。
+            </div>
+
+            <label style="margin-top: 8px;" data-i18n="bgmLabel">成片配乐（可选，统一音轨）</label>
+            <div class="upload-zone" id="batch-audio-drop-zone" onclick="document.getElementById('batch-audio-input').click()" style="min-height: 44px; margin-bottom: 8px; position: relative;">
+                <div class="clear-img-overlay" id="clear-batch-audio-overlay" onclick="event.stopPropagation(); clearBatchBackgroundAudio()" style="display: none;">×</div>
+                <div id="batch-audio-placeholder">
+                    <div class="upload-text" style="font-size: 11px;" data-i18n="bgmUploadHint">上传一条完整 BGM（生成完成后会替换整段成片的音轨）</div>
+                </div>
+                <div id="batch-audio-status" style="display: none; font-size: 11px; color: var(--accent);"></div>
+                <input type="file" id="batch-audio-input" accept="audio/*" style="display:none" onchange="handleBatchBackgroundAudioUpload(this.files[0])">
+            </div>
+            <input type="hidden" id="batch-background-audio-path">
+
+            <div id="batch-segments-container" style="margin-top: 15px;"></div>
+        </div>
+
+        <div class="setting-group">
+            <div class="group-title" data-i18n="basicEngine">基础画面 / Basic EngineSpecs</div>
+            <div class="flex-row">
+                <div class="flex-1">
+                    <label data-i18n="qualityLevel">清晰度级别</label>
+                    <select id="batch-quality" onchange="updateBatchResPreview()">
+                        <option value="1080">1080P Full HD</option>
+                        <option value="720" selected>720P Standard</option>
+                        <option value="540">540P Preview</option>
+                    </select>
+                </div>
+                <div class="flex-1">
+                    <label data-i18n="aspectRatio">画幅比例</label>
+                    <select id="batch-ratio" onchange="updateBatchResPreview()">
+                        <option value="16:9" data-i18n="ratio169">16:9 电影宽幅</option>
+                        <option value="9:16" data-i18n="ratio916">9:16 移动竖屏</option>
+                        <option value="1:1" data-i18n="ratio11">1:1 方形</option>
+                        <option value="4:3" data-i18n="ratio43">4:3 经典横幅</option>
+                        <option value="3:4" data-i18n="ratio34">3:4 经典竖幅</option>
+                        <option value="21:9" data-i18n="ratio219">21:9 超宽银幕</option>
+                        <option value="9:21" data-i18n="ratio921">9:21 超长竖屏</option>
+                        <option value="ref" data-i18n="ratioRef">跟随参考图</option>
+                        <option value="custom" data-i18n="ratioCustom">自定义尺寸</option>
+                    </select>
+                </div>
+            </div>
+            <div id="batch-custom-size" class="flex-row" style="display:none; margin-top: -2px; margin-bottom: 10px;">
+                <div class="flex-1"><label data-i18n="width">宽度</label><input type="number" id="batch-custom-w" value="1280" min="64" step="64" onchange="updateBatchResPreview()"></div>
+                <div class="flex-1"><label data-i18n="height">高度</label><input type="number" id="batch-custom-h" value="704" min="64" step="64" onchange="updateBatchResPreview()"></div>
+            </div>
+            <div id="batch-res-preview" class="res-preview-tag" style="margin-top: -5px; margin-bottom: 12px;">最终发送: 1280x704</div>
+
+            <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 8px;">
+                <label data-i18n="selectLora" style="margin: 0;">选择 LoRA</label>
+                <button type="button" onclick="addLoraSelection('batch-loras-container')" style="background:none; border:none; color:var(--accent); cursor:pointer; font-size:16px; padding:0 4px;" title="添加 LoRA">+</button>
+            </div>
+            <div id="batch-loras-container" style="display: flex; flex-direction: column; gap: 8px; margin-bottom: 8px;"></div>
+        </div>
+    </div>
+
+    <!-- TTS 语音合成面板 -->
+    <div id="tts-opts" class="sidebar-section" style="display:none; padding-top: 0;">
+        <!-- 状态栏 -->
+        <div id="tts-status-bar" style="padding: 8px 12px; border-radius: 8px; background: rgba(255,255,255,0.04); border: 1px solid var(--border); margin-bottom: 14px; font-size: 11px; color: var(--text-dim);" data-i18n="ttsStatusBarDetecting">
+            🔍 正在检测 TTS 模型...
+        </div>
+
+        <!-- 合成文本 -->
+        <div class="setting-group">
+            <div class="group-title" data-i18n="ttsTextTitle">合成文本 / Text</div>
+            <label style="font-size:11px; color:var(--text-dim); margin-bottom:5px; display:block;" data-i18n-html="ttsTextHint">
+                支持在文本开头加英文括号描述声音，例如：<code style="font-size:10px;">(年轻女声，温柔甜美)</code>
+            </label>
+            <textarea id="tts-text" data-i18n-placeholder="ttsTextPlaceholder" placeholder="输入要合成的文本内容..." style="width:100%; height:90px; padding:8px; font-size:12px; box-sizing:border-box; resize:vertical; border-radius:8px; border:1px solid var(--border); background:var(--item); color:var(--text);"></textarea>
+        </div>
+
+        <!-- 合成模式 -->
+        <div class="setting-group">
+            <div class="group-title" data-i18n="ttsModeTitle">合成模式 / Mode</div>
+            <select id="tts-mode" onchange="onTtsModeChange()" style="margin-bottom:8px;">
+                <option value="text_only" data-i18n="ttsModeTextOnly">🗣️ 文字转语音（含声音设计）</option>
+                <option value="clone" data-i18n="ttsModeClone">🎙️ 声音克隆</option>
+                <option value="ultimate_clone" data-i18n="ttsModeUltimate">⭐ 终极克隆（最高还原度）</option>
+            </select>
+
+            <!-- 声音克隆：参考音频 -->
+            <div id="tts-ref-section" style="display:none;">
+                <label style="font-size:11px; margin-bottom:4px; display:block;" data-i18n="ttsRefLabel">📎 参考音频（Reference）</label>
+                <div class="upload-zone" id="tts-ref-drop" onclick="document.getElementById('tts-ref-input').click()" style="min-height:48px; margin-bottom:6px; position:relative;">
+                    <div class="clear-img-overlay" id="tts-ref-clear" onclick="event.stopPropagation(); clearTtsRef()" style="display:none;">×</div>
+                    <div id="tts-ref-placeholder">
+                        <div class="upload-icon">🎵</div>
+                        <div class="upload-text" style="font-size:11px;" data-i18n="ttsRefUploadHint">点击上传参考音频 (.wav / .mp3)</div>
+                    </div>
+                    <div id="tts-ref-status" style="display:none; font-size:11px; color:var(--accent); padding:8px; text-align:center;"></div>
+                    <input type="file" id="tts-ref-input" accept="audio/*" style="display:none" onchange="handleTtsRefUpload(this.files[0])">
+                </div>
+            </div>
+
+            <!-- 终极克隆额外选项 -->
+            <div id="tts-ultimate-section" style="display:none;">
+                <label style="font-size:11px; margin-bottom:4px; display:block;" data-i18n="ttsUltimateLabel">📝 参考音频对应的文本转录（可选，能显著提升相似度）</label>
+                <textarea id="tts-prompt-text" data-i18n-placeholder="ttsUltimatePlaceholder" placeholder="与参考音频完全一致的文本内容..." style="width:100%; height:60px; padding:6px 8px; font-size:11px; box-sizing:border-box; resize:vertical; border-radius:6px; border:1px solid var(--border); background:var(--item); color:var(--text); margin-bottom:6px;"></textarea>
+            </div>
+        </div>
+
+        <!-- 参数调节 -->
+        <div class="setting-group">
+            <div class="group-title" data-i18n="ttsParamsTitle">高级参数 / Parameters</div>
+            <div class="flex-row" style="gap:12px;">
+                <div class="flex-1">
+                    <div class="label-group">
+                        <label style="font-size:11px;" data-i18n="ttsCfgLabel">CFG 强度</label>
+                        <span class="val-badge" id="ttsCfgVal">2.0</span>
+                    </div>
+                    <div class="slider-container">
+                        <input type="range" id="tts-cfg" min="0.5" max="5.0" step="0.5" value="2.0"
+                               oninput="document.getElementById('ttsCfgVal').textContent=this.value">
+                    </div>
+                </div>
+                <div class="flex-1">
+                    <div class="label-group">
+                        <label style="font-size:11px;" data-i18n="ttsStepsLabel">推理步数</label>
+                        <span class="val-badge" id="ttsStepsVal">10</span>
+                    </div>
+                    <div class="slider-container">
+                        <input type="range" id="tts-steps" min="5" max="50" step="5" value="10"
+                               oninput="document.getElementById('ttsStepsVal').textContent=this.value">
+                    </div>
+                </div>
+            </div>
+        </div>
+
+        <!-- 输出结果播放区 -->
+        <div id="tts-result-section" style="display:none;" class="setting-group">
+            <div class="group-title" data-i18n="ttsResultTitle">生成结果 / Output</div>
+            <audio id="tts-audio-player" controls style="width:100%; border-radius:8px; margin-bottom:6px;"></audio>
+            <a id="tts-download-link" href="#" download style="font-size:11px; color:var(--accent); text-decoration:none;" data-i18n="ttsDownload">⬇️ 下载音频</a>
+        </div>
+
+        <div style="padding: 0 0 10px 0;">
+            <button class="btn-primary" id="tts-gen-btn" onclick="runTts()" data-i18n="ttsGenBtn">🎙️ 开始生成语音</button>
+        </div>
+    </div>
+
+    <div style="padding: 0 30px 30px 30px;">
+        <button class="btn-primary" id="mainBtn" onclick="run()" data-i18n="mainRender">开始渲染</button>
+    </div>
+</aside>
+
+<main class="workspace">
+    <section class="viewer" id="viewer-section">
+        <div class="monitor" id="viewer">
+            <button id="preview-download-btn" type="button" onclick="downloadCurrentPreviewAsset()" class="preview-download-btn" style="display:none;">
+                <span class="preview-download-btn-icon" aria-hidden="true">
+                    <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round">
+                        <path d="M12 3v12"></path>
+                        <path d="M7 10l5 5 5-5"></path>
+                        <path d="M5 21h14"></path>
+                    </svg>
+                </span>
+                <span class="preview-download-btn-text" data-i18n="downloadLabel">下载</span>
+            </button>
+            <div id="preview-replay-actions" class="preview-replay-actions" style="display:none;">
+                <button type="button" onclick="loadCurrentPreviewSeed()" data-i18n="previewLoadSeed">载入种子</button>
+                <button type="button" onclick="loadCurrentPreviewParams()" data-i18n="previewLoadParams">载入参数</button>
+            </div>
+            <div id="loading-txt" data-i18n="waitingTask">等待分配渲染任务...</div>
+            <img id="res-img" src="">
+            <div id="video-wrapper" style="width:100%; height:100%; display:none; max-height:100%; align-items:center; justify-content:center;">
+                <video id="res-video" autoplay loop playsinline></video>
+            </div>
+            <div id="audio-wrapper">
+                <div class="audio-preview-art" id="audio-preview-art" role="button" tabindex="0">
+                    <div class="audio-preview-icon">♪</div>
+                    <div id="audio-preview-title">TTS Audio</div>
+                </div>
+                <audio id="res-audio" controls></audio>
+            </div>
+            <div class="progress-container"><div id="progress-fill"></div></div>
+        </div>
+    </section>
+
+    <!-- Drag Handle -->
+    <div id="resize-handle" style="
+        height: 5px; background: transparent; cursor: row-resize;
+        flex-shrink: 0; position: relative; z-index: 50;
+        display: flex; align-items: center; justify-content: center;
+    " data-i18n-title="resizeHandleTitle" title="拖动调整面板高度">
+        <div style="width: 40px; height: 3px; background: var(--border); border-radius: 999px; pointer-events: none;"></div>
+    </div>
+
+    <section class="library" id="library-section">
+        <div style="display: flex; justify-content: space-between; margin-bottom: 15px; align-items: center; border-bottom: 1px solid var(--border); padding-bottom: 10px;">
+            <div style="display: flex; gap: 20px;">
+                <span id="tab-history" style="font-size: 11px; font-weight: 800; color: var(--accent); cursor: pointer; border-bottom: 2px solid var(--accent); padding-bottom: 11px; margin-bottom: -11px;" onclick="switchLibTab('history')" data-i18n="libHistory">历史资产 / ASSETS</span>
+                <span id="tab-log" style="font-size: 11px; font-weight: 800; color: var(--text-dim); cursor: pointer; border-bottom: 2px solid transparent; padding-bottom: 11px; margin-bottom: -11px;" onclick="switchLibTab('log')" data-i18n="libLog">系统日志 / LOGS</span>
+            </div>
+            <button type="button" onclick="fetchHistory(currentHistoryPage)" style="background: var(--item); border: 1px solid var(--border); border-radius: 6px; color: var(--text-dim); font-size: 11px; padding: 4px 10px; cursor: pointer;" data-i18n="refresh">刷新</button>
+        </div>
+        
+        <div id="log-container" style="display: none; flex: 1; flex-direction: column;">
+            <div id="log" data-i18n="logReady">> LTX-2 Studio Ready. Expecting commands...</div>
+        </div>
+        
+        <div id="history-wrapper">
+            <div id="history-container"></div>
+        </div>
+        <div id="pagination-bar" style="display:none;"></div>
+    </section>
+</main>
+<script src="https://cdnjs.cloudflare.com/ajax/libs/plyr/3.7.8/plyr.min.js"></script>
+    <script src="i18n.js?v=en-tabs-1"></script>
+    <script src="index.js?v=model-switch-1"></script>
+
+</body>
+</html>

LTX2.3-1.0.4-new/main.py

@@ -0,0 +1,266 @@
+import os
+import sys
+import subprocess
+import threading
+import time
+import socket
+import logging
+from fastapi import FastAPI
+from fastapi.responses import FileResponse
+from fastapi.staticfiles import StaticFiles
+import uvicorn
+
+# ============================================================
+# 配置区 (动态路径适配与补丁挂载)
+# ============================================================
+def resolve_ltx_path():
+    import glob, tempfile, subprocess
+    sc_dir = os.path.join(os.getcwd(), "LTX_Shortcut")
+    os.makedirs(sc_dir, exist_ok=True)
+    lnk_files = glob.glob(os.path.join(sc_dir, "*.lnk"))
+    if not lnk_files:
+        print("\033[91m[ERROR] 未在 LTX_Shortcut 文件夹中找到快捷方式！\n请打开程序目录下的 LTX_Shortcut 文件夹，并将官方 LTX Desktop 的快捷方式复制进去后重试。\033[0m")
+        sys.exit(1)
+        
+    lnk_path = lnk_files[0]
+    # 使用 VBScript 解析快捷方式，兼容所有 Windows 系统
+    vbs_code = f'''Set sh = CreateObject("WScript.Shell")\nSet obj = sh.CreateShortcut("{os.path.abspath(lnk_path)}")\nWScript.Echo obj.TargetPath'''
+    fd, vbs_path = tempfile.mkstemp(suffix='.vbs')
+    with os.fdopen(fd, 'w') as f:
+        f.write(vbs_code)
+    try:
+        out = subprocess.check_output(['cscript', '//nologo', vbs_path], stderr=subprocess.STDOUT)
+        target_exe = out.decode('ansi').strip()
+    finally:
+        os.remove(vbs_path)
+        
+    if not target_exe or not os.path.exists(target_exe):
+        # 如果快捷方式解析失败，或者解析出来的是朋友电脑的路径（当前电脑不存在），自动全盘搜索默认路径
+        default_paths = [
+            os.path.join(os.environ.get("LOCALAPPDATA", ""), r"Programs\LTX Desktop\LTX Desktop.exe"),
+            r"C:\Program Files\LTX Desktop\LTX Desktop.exe",
+            r"D:\Program Files\LTX Desktop\LTX Desktop.exe",
+            r"E:\Program Files\LTX Desktop\LTX Desktop.exe"
+        ]
+        found = False
+        for p in default_paths:
+            if os.path.exists(p):
+                target_exe = p
+                print(f"\033[96m[INFO] 自动检测到 LTX 原版安装路径: {p}\033[0m")
+                found = True
+                break
+        
+        if not found:
+            print(f"\033[91m[ERROR] 未能找到原版 LTX Desktop 的安装路径！\033[0m")
+            print("请清理 LTX_Shortcut 文件夹，并将您当前电脑上真正的原版快捷方式重贴复制进去。")
+            sys.exit(1)
+        
+    return os.path.dirname(target_exe)
+
+USER_PROFILE = os.path.expanduser("~")
+PYTHON_EXE = os.path.join(USER_PROFILE, r"AppData\Local\LTXDesktop\python\python.exe")
+DATA_DIR = os.path.join(USER_PROFILE, r"AppData\Local\LTXDesktop")
+
+# 1. 动态获取主安装路径
+LTX_INSTALL_DIR = resolve_ltx_path()
+BACKEND_DIR = os.path.join(LTX_INSTALL_DIR, r"resources\backend")
+UI_FILE_NAME = "UI/index.html"
+
+# 环境致命检测：如果官方 Python 还没解压释放，立刻强制中断整个程序
+if not os.path.exists(PYTHON_EXE):
+    print(f"\n\033[1;41m [致命错误] 您的电脑上尚未配置好 LTX 的官方渲染核心框架！ \033[0m")
+    print(f"\033[93m此应用仅是 UI 图形控制台，必需依赖原版软件环境才能生成。在 ({PYTHON_EXE}) 未找到运行引擎。\n")
+    print(">> 解决方案：\n1. 请先在您的电脑上正常安装【LTX Desktop 官方原版软件】。")
+    print("2. 必需：双击打开运行一次原版软件！（运行后原版软件会在后台自动释放环境）")
+    print("3. 把原版软件的快捷方式复制到本文档的 LTX_Shortcut 文件夹里面。")
+    print("4. 全部完成后，再重新启动本 run.bat 脚本即可！\033[0m\n")
+    os._exit(1)
+
+# 2. 从目录读取改动过的 Python 文件 (热修复拦截器)
+PATCHES_DIR = os.path.join(os.getcwd(), "patches")
+os.makedirs(PATCHES_DIR, exist_ok=True)
+
+# 3. 默认输出定向至程序根目录
+LOCAL_OUTPUTS = os.path.join(os.getcwd(), "outputs")
+os.makedirs(LOCAL_OUTPUTS, exist_ok=True)
+
+# 强制注入自定义输出录至 LTX 缓存数据中
+os.makedirs(DATA_DIR, exist_ok=True)
+with open(os.path.join(DATA_DIR, "custom_dir.txt"), 'w', encoding='utf-8') as f:
+    f.write(LOCAL_OUTPUTS)
+
+os.environ["LTX_APP_DATA_DIR"] = DATA_DIR
+
+# 将 patches 目录优先级提升，做到 Python 无损替换
+os.environ["PYTHONPATH"] = f"{PATCHES_DIR};{BACKEND_DIR}"
+
+def get_lan_ip():
+    try:
+        host_name = socket.gethostname()
+        _, _, ip_list = socket.gethostbyname_ex(host_name)
+        
+        candidates = []
+        for ip in ip_list:
+            if ip.startswith("192.168."):
+                return ip
+            elif ip.startswith("10.") or (ip.startswith("172.") and 16 <= int(ip.split('.')[1]) <= 31):
+                candidates.append(ip)
+                
+        if candidates:
+            return candidates[0]
+            
+        # Fallback to the default socket routing approach if no obvious LAN IP found
+        s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+        s.connect(("8.8.8.8", 80))
+        ip = s.getsockname()[0]
+        s.close()
+        return ip
+    except:
+        return "127.0.0.1"
+
+LAN_IP = get_lan_ip()
+
+# ============================================================
+# 服务启动逻辑
+# ============================================================
+def check_port_in_use(port):
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        return s.connect_ex(('127.0.0.1', port)) == 0
+
+def launch_backend():
+    """启动核心引擎 - 监听 0.0.0.0 确保局域网可调"""
+    if check_port_in_use(3000):
+        print(f"\n\033[1;41m [致命错误] 3000 端口已被占用，无法启动核心引擎！ \033[0m")
+        print("\033[93m>> 绝大多数情况下，这是因为【官方原版 LTX Desktop】正在您的电脑后台运行。\033[0m")
+        print(">> 冲突会导致显存爆炸。请检查右下角系统托盘图标，右键完全退出官方软件。")
+        print(">> 退出后重新双击 run.bat 启动本程序！\n")
+        os._exit(1)
+
+    print(f"\033[96m[CORE] 核心引擎正在启动...\033[0m")
+    # 只开启重要级别的 Python 应用层日志，去除无用的 HTTP 刷屏
+    import logging as _logging
+    _logging.basicConfig(
+        level=_logging.INFO,
+        format="[%(asctime)s] %(levelname)s %(name)s: %(message)s",
+        datefmt="%H:%M:%S",
+        force=True
+    )
+    
+    # 构建绝对无损的环境拦截器：防止其他电脑被 cwd 劫持加载原版文件
+    launcher_code = f"""
+import sys
+import os
+
+patch_dir = r"{PATCHES_DIR}"
+backend_dir = r"{BACKEND_DIR}"
+
+# 防御性清除：强行剥离所有的默认 backend_dir 引用
+sys.path = [p for p in sys.path if p and os.path.normpath(p) != os.path.normpath(backend_dir)]
+sys.path = [p for p in sys.path if p and p != "." and p != ""]
+
+# 绝对插队注入：优先搜索 PATCHES_DIR
+sys.path.insert(0, patch_dir)
+sys.path.insert(1, backend_dir)
+
+import uvicorn
+from ltx2_server import app
+
+if __name__ == '__main__':
+    uvicorn.run(app, host="0.0.0.0", port=3000, log_level="info", access_log=False)
+"""
+    launcher_path = os.path.join(PATCHES_DIR, "launcher.py")
+    with open(launcher_path, "w", encoding="utf-8") as f:
+        f.write(launcher_code)
+
+    cmd = [PYTHON_EXE, launcher_path]
+    env = os.environ.copy()
+    result = subprocess.run(cmd, cwd=BACKEND_DIR, env=env)
+    if result.returncode != 0:
+        print(f"\n\033[1;41m [致命错误] 核心引擎异常崩溃退出！ (Exit Code: {result.returncode})\033[0m")
+        print(">> 请检查上述终端报错信息。确认显卡驱动是否正常。")
+        os._exit(1)
+
+ui_app = FastAPI()
+# 已移除存在安全隐患的静态资源挂载目录
+
+UI_NO_CACHE_HEADERS = {"Cache-Control": "no-store, max-age=0"}
+
+@ui_app.get("/")
+async def serve_index():
+    return FileResponse(os.path.join(os.getcwd(), UI_FILE_NAME), headers=UI_NO_CACHE_HEADERS)
+
+@ui_app.get("/index.css")
+async def serve_css():
+    return FileResponse(os.path.join(os.getcwd(), "UI/index.css"), headers=UI_NO_CACHE_HEADERS)
+
+@ui_app.get("/index.js")
+async def serve_js():
+    return FileResponse(os.path.join(os.getcwd(), "UI/index.js"), headers=UI_NO_CACHE_HEADERS)
+
+
+@ui_app.get("/i18n.js")
+async def serve_i18n():
+    return FileResponse(os.path.join(os.getcwd(), "UI/i18n.js"), headers=UI_NO_CACHE_HEADERS)
+
+
+def launch_ui_server():
+    print(f"\033[92m[UI] 工作站已就绪！\033[0m")
+    print(f"\033[92m[LOCAL] 本机访问: http://127.0.0.1:4000\033[0m")
+    print(f"\033[93m[WIFI]  局域网访问: http://{LAN_IP}:4000\033[0m")
+    
+    # 彻底压制 WinError 10054 (客户端强制断开) 的底层警告报错
+    if sys.platform == 'win32':
+        # Uvicorn 内部会拉起循环，所以只能通过底层 Logging Filter 拦截控制台噪音
+        class UvicornAsyncioNoiseFilter(logging.Filter):
+            """压掉客户端断开、Win Proactor 管道收尾等无害 asyncio 控制台刷屏。"""
+
+            def filter(self, record):
+                if record.name != "asyncio":
+                    return True
+                msg = record.getMessage()
+                if "_call_connection_lost" in msg or "_ProactorBasePipeTransport" in msg:
+                    return False
+                if hasattr(record, "exc_info") and record.exc_info:
+                    exc_type, exc_value, _ = record.exc_info
+                    if isinstance(exc_value, ConnectionResetError) and getattr(
+                        exc_value, "winerror", None
+                    ) == 10054:
+                        return False
+                if "10054" in msg and "ConnectionResetError" in msg:
+                    return False
+                return True
+
+        logging.getLogger("asyncio").addFilter(UvicornAsyncioNoiseFilter())
+        
+    uvicorn.run(ui_app, host="0.0.0.0", port=4000, log_level="warning", access_log=False)
+
+if __name__ == "__main__":
+    os.system('cls' if os.name == 'nt' else 'clear')
+    print("\033[1;97;44m LTX-2 CINEMATIC WORKSTATION | NETWORK ENABLED \033[0m\n")
+    
+    threading.Thread(target=launch_backend, daemon=True).start()
+    
+    # 强制校验 3000 端口是否存活
+    print("\033[93m[SYS] 正在等待内部核心 3000 端口启动...\033[0m")
+    backend_ready = False
+    for _ in range(30):
+        try:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+                if s.connect_ex(('127.0.0.1', 3000)) == 0:
+                    backend_ready = True
+                    break
+        except Exception:
+            pass
+        time.sleep(1)
+        
+    if backend_ready:
+        print("\033[92m[SYS] 3000 端口已通过连通性握手验证！后端装载成功。\033[0m")
+    else:
+        print("\033[1;41m [崩坏警告] 等待 30 秒后，3000 端口依然无法连通！ \033[0m")
+        print(">> Uvicorn 可能在后台陷入了死锁，或者被防火墙拦截，前端大概率将无法连接到后端！")
+        print(">> 请检查上方是否有 Python 报错。\n")
+        
+    try:
+        launch_ui_server()
+    except KeyboardInterrupt:
+        sys.exit(0)

LTX2.3-1.0.4-new/patches/API模式问题修复说明.md

@@ -0,0 +1,41 @@
+# LTX 本地显卡模式修复
+
+## 问题描述
+系统强制使用 FAL API 生成图片，即使本地有 GPU 可用。
+
+## 原因
+LTX 强制要求 GPU 有 31GB VRAM 才会使用本地显卡，低于此值会强制走 API 模式。
+
+## 修复方法
+
+### 方法一：自动替换（推荐）
+运行程序后，patches 目录中的文件会自动替换原版文件。
+
+### 方法二：手动替换
+
+#### 1. 修改 VRAM 阈值
+- **原文件**: `C:\Program Files\LTX Desktop\resources\backend\runtime_config\runtime_policy.py`
+- **找到** (第16行):
+  ```python
+  return vram_gb < 31
+  ```
+- **改为**:
+  ```python
+  return vram_gb < 6
+  ```
+
+#### 2. 清空无效 API Key
+- **原文件**: `C:\Users\Administrator\AppData\Local\LTXDesktop\settings.json`
+- **找到**:
+  ```json
+  "fal_api_key": "12123",
+  ```
+- **改为**:
+  ```json
+  "fal_api_key": "",
+  ```
+
+## 说明
+- VRAM 阈值改为 6GB，意味着 6GB 及以上显存都会使用本地显卡
+- 清空 fal_api_key 避免系统误判为已配置 API
+- 修改后重启程序即可生效

LTX2.3-1.0.4-new/patches/__pycache__/app_factory.cpython-313.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a39240a5806a7be8c147f3fbd45dc7ea3f8c0a6a3550741d90423d4345335ee
+size 143860

LTX2.3-1.0.4-new/patches/api_types.py

@@ -0,0 +1,403 @@
+"""Pydantic request/response models and TypedDicts for ltx2_server."""
+
+from __future__ import annotations
+
+from typing import Literal, NamedTuple, TypeAlias, TypedDict
+from typing import Annotated
+
+from pydantic import BaseModel, Field, StringConstraints
+
+NonEmptyPrompt = Annotated[str, StringConstraints(strip_whitespace=True, min_length=1)]
+ModelFileType = Literal[
+    "checkpoint",
+    "upsampler",
+    "distilled_lora",
+    "ic_lora",
+    "depth_processor",
+    "person_detector",
+    "pose_processor",
+    "text_encoder",
+    "zit",
+]
+
+
+class ImageConditioningInput(NamedTuple):
+    """Image conditioning triplet used by all video pipelines."""
+
+    path: str
+    frame_idx: int
+    strength: float
+
+
+# ============================================================
+# TypedDicts for module-level state globals
+# ============================================================
+
+
+class GenerationState(TypedDict):
+    id: str | None
+    cancelled: bool
+    result: str | list[str] | None
+    error: str | None
+    status: str  # "idle" | "running" | "complete" | "cancelled" | "error"
+    phase: str
+    progress: int
+    current_step: int
+    total_steps: int
+
+
+JsonObject: TypeAlias = dict[str, object]
+VideoCameraMotion = Literal[
+    "none",
+    "dolly_in",
+    "dolly_out",
+    "dolly_left",
+    "dolly_right",
+    "jib_up",
+    "jib_down",
+    "static",
+    "focus_shift",
+]
+
+RetakeMode: TypeAlias = Literal[
+    "replace_audio_and_video", "replace_video", "replace_audio"
+]
+
+
+# ============================================================
+# Response Models
+# ============================================================
+
+
+class ModelStatusItem(BaseModel):
+    id: str
+    name: str
+    loaded: bool
+    downloaded: bool
+
+
+class GpuTelemetry(BaseModel):
+    name: str
+    vram: int
+    vramUsed: int
+
+
+class HealthResponse(BaseModel):
+    status: str
+    models_loaded: bool
+    active_model: str | None
+    gpu_info: GpuTelemetry
+    sage_attention: bool
+    models_status: list[ModelStatusItem]
+
+
+class GpuInfoResponse(BaseModel):
+    cuda_available: bool
+    mps_available: bool = False
+    gpu_available: bool = False
+    gpu_name: str | None
+    vram_gb: int | None
+    gpu_info: GpuTelemetry
+
+
+class RuntimePolicyResponse(BaseModel):
+    force_api_generations: bool
+
+
+class GenerationProgressResponse(BaseModel):
+    status: str
+    phase: str
+    progress: int
+    currentStep: int | None
+    totalSteps: int | None
+
+
+class ModelInfo(BaseModel):
+    id: str
+    name: str
+    description: str
+
+
+class ModelFileStatus(BaseModel):
+    id: ModelFileType
+    name: str
+    description: str
+    downloaded: bool
+    size: int
+    expected_size: int
+    required: bool = True
+    is_folder: bool = False
+    optional_reason: str | None = None
+
+
+class TextEncoderStatus(BaseModel):
+    downloaded: bool
+    size_bytes: int
+    size_gb: float
+    expected_size_gb: float
+
+
+class ModelsStatusResponse(BaseModel):
+    models: list[ModelFileStatus]
+    all_downloaded: bool
+    total_size: int
+    downloaded_size: int
+    total_size_gb: float
+    downloaded_size_gb: float
+    models_path: str
+    has_api_key: bool
+    text_encoder_status: TextEncoderStatus
+    use_local_text_encoder: bool
+
+
+class DownloadProgressRunningResponse(BaseModel):
+    status: Literal["downloading"]
+    current_downloading_file: ModelFileType | None
+    current_file_progress: float
+    total_progress: float
+    total_downloaded_bytes: int
+    expected_total_bytes: int
+    completed_files: set[ModelFileType]
+    all_files: set[ModelFileType]
+    error: None = None
+    speed_bytes_per_sec: float
+
+
+class DownloadProgressCompleteResponse(BaseModel):
+    status: Literal["complete"]
+
+
+class DownloadProgressErrorResponse(BaseModel):
+    status: Literal["error"]
+    error: str
+
+
+DownloadProgressResponse: TypeAlias = (
+    DownloadProgressRunningResponse
+    | DownloadProgressCompleteResponse
+    | DownloadProgressErrorResponse
+)
+
+
+class SuggestGapPromptResponse(BaseModel):
+    status: str = "success"
+    suggested_prompt: str
+
+
+class GenerateVideoCompleteResponse(BaseModel):
+    status: Literal["complete"]
+    video_path: str
+
+
+class GenerateVideoCancelledResponse(BaseModel):
+    status: Literal["cancelled"]
+
+
+GenerateVideoResponse: TypeAlias = (
+    GenerateVideoCompleteResponse | GenerateVideoCancelledResponse
+)
+
+
+class GenerateImageCompleteResponse(BaseModel):
+    status: Literal["complete"]
+    image_paths: list[str]
+
+
+class GenerateImageCancelledResponse(BaseModel):
+    status: Literal["cancelled"]
+
+
+GenerateImageResponse: TypeAlias = (
+    GenerateImageCompleteResponse | GenerateImageCancelledResponse
+)
+
+
+class CancelCancellingResponse(BaseModel):
+    status: Literal["cancelling"]
+    id: str
+
+
+class CancelNoActiveGenerationResponse(BaseModel):
+    status: Literal["no_active_generation"]
+
+
+CancelResponse: TypeAlias = CancelCancellingResponse | CancelNoActiveGenerationResponse
+
+
+class RetakeVideoResponse(BaseModel):
+    status: Literal["complete"]
+    video_path: str
+
+
+class RetakePayloadResponse(BaseModel):
+    status: Literal["complete"]
+    result: JsonObject
+
+
+class RetakeCancelledResponse(BaseModel):
+    status: Literal["cancelled"]
+
+
+RetakeResponse: TypeAlias = (
+    RetakeVideoResponse | RetakePayloadResponse | RetakeCancelledResponse
+)
+
+
+class IcLoraExtractResponse(BaseModel):
+    conditioning: str
+    original: str
+    conditioning_type: Literal["canny", "depth", "pose", "video"]
+    frame_time: float
+
+
+class IcLoraGenerateCompleteResponse(BaseModel):
+    status: Literal["complete"]
+    video_path: str
+
+
+class IcLoraGenerateCancelledResponse(BaseModel):
+    status: Literal["cancelled"]
+
+
+IcLoraGenerateResponse: TypeAlias = (
+    IcLoraGenerateCompleteResponse | IcLoraGenerateCancelledResponse
+)
+
+
+class ModelDownloadStartResponse(BaseModel):
+    status: Literal["started"]
+    message: str
+    sessionId: str
+
+
+class TextEncoderDownloadStartedResponse(BaseModel):
+    status: Literal["started"]
+    message: str
+    sessionId: str
+
+
+class TextEncoderAlreadyDownloadedResponse(BaseModel):
+    status: Literal["already_downloaded"]
+    message: str
+
+
+TextEncoderDownloadResponse: TypeAlias = (
+    TextEncoderDownloadStartedResponse | TextEncoderAlreadyDownloadedResponse
+)
+
+
+class StatusResponse(BaseModel):
+    status: str
+
+
+class ErrorResponse(BaseModel):
+    error: str
+    message: str | None = None
+
+
+# ============================================================
+# Request Models
+# ============================================================
+
+
+class GenerateVideoRequest(BaseModel):
+    prompt: NonEmptyPrompt
+    resolution: str = "512p"
+    model: str = "fast"
+    cameraMotion: VideoCameraMotion = "none"
+    negativePrompt: str = ""
+    duration: str = "2"
+    fps: str = "24"
+    audio: str = "false"
+    imagePath: str | None = None
+    audioPath: str | None = None
+    startFramePath: str | None = None
+    endFramePath: str | None = None
+    # 多张图单次推理：latent 时间轴多锚点（Comfy LTXVAddGuideMulti 思路）；≥2 路径时优先于首尾帧
+    keyframePaths: list[str] | None = None
+    # 与 keyframePaths 等长、0.1–1.0；不传则按 Comfy 类工作流自动降低中间帧强度，减轻闪烁
+    keyframeStrengths: list[float] | None = None
+    # 与 keyframePaths 等长，单位秒，落在 [0, 整段时长]；全提供时按时间映射 latent，否则仍自动均分
+    keyframeTimes: list[float] | None = None
+    aspectRatio: str = "16:9"
+    customWidth: int | None = None
+    customHeight: int | None = None
+    modelPath: str | None = None
+    loraPath: str | None = None
+    loraStrength: float = 2.0
+    loraPaths: list[str] | None = None
+    loraStrengths: list[float] | None = None
+    seed: int | None = None
+
+
+class GenerateImageRequest(BaseModel):
+    prompt: NonEmptyPrompt
+    width: int = 1024
+    height: int = 1024
+    numSteps: int = 4
+    numImages: int = 1
+    seed: int | None = None
+
+
+def _default_model_types() -> set[ModelFileType]:
+    return set()
+
+
+class ModelDownloadRequest(BaseModel):
+    modelTypes: set[ModelFileType] = Field(default_factory=_default_model_types)
+
+
+class RequiredModelsResponse(BaseModel):
+    modelTypes: list[ModelFileType]
+
+
+class SuggestGapPromptRequest(BaseModel):
+    beforePrompt: str = ""
+    afterPrompt: str = ""
+    beforeFrame: str | None = None
+    afterFrame: str | None = None
+    gapDuration: float = 5
+    mode: str = "t2v"
+    inputImage: str | None = None
+
+
+class RetakeRequest(BaseModel):
+    video_path: str
+    start_time: float = 0
+    duration: float = 0
+    prompt: str = ""
+    mode: str = "replace_video_only"
+    width: int | None = None
+    height: int | None = None
+
+
+class IcLoraExtractRequest(BaseModel):
+    video_path: str
+    conditioning_type: Literal["canny", "depth", "pose", "video"] = "canny"
+    frame_time: float = 0
+
+
+class IcLoraImageInput(BaseModel):
+    path: str
+    frame: int = 0
+    strength: float = 1.0
+
+
+def _default_ic_lora_images() -> list[IcLoraImageInput]:
+    return []
+
+
+class IcLoraGenerateRequest(BaseModel):
+    video_path: str
+    conditioning_type: Literal["canny", "depth", "pose", "video"]
+    prompt: NonEmptyPrompt
+    conditioning_strength: float = 1.0
+    num_inference_steps: int = 30
+    cfg_guidance_scale: float = 1.0
+    negative_prompt: str = ""
+    images: list[IcLoraImageInput] = Field(default_factory=_default_ic_lora_images)
+    ic_lora_path: str | None = None
+    seed: int | None = None
+
+
+ConditioningType: TypeAlias = Literal["canny", "depth", "pose", "video"]

LTX2.3-1.0.4-new/patches/app_settings_patch.py

@@ -0,0 +1,22 @@
+"""运行时补丁：给 AppSettings 添加 lora_dir 字段（如果不存在）。"""
+
+import sys
+import os
+
+
+def patch_app_settings():
+    try:
+        from state.app_settings import AppSettings
+        from pydantic import Field
+
+        if "lora_dir" not in AppSettings.model_fields:
+            AppSettings.model_fields["lora_dir"] = Field(
+                default="", validation_alias="loraDir", serialization_alias="loraDir"
+            )
+            AppSettings.model_rebuild(_force=True)
+            print("[PATCH] AppSettings patched: added lora_dir field")
+    except Exception as e:
+        print(f"[PATCH] AppSettings patch failed: {e}")
+
+
+patch_app_settings()

LTX2.3-1.0.4-new/patches/handlers/video_generation_handler.py

@@ -0,0 +1,882 @@
+"""Video generation orchestration handler."""
+
+from __future__ import annotations
+
+import logging
+import os
+import tempfile
+import time
+import uuid
+from datetime import datetime
+from pathlib import Path
+from threading import RLock
+from typing import TYPE_CHECKING
+
+from PIL import Image
+
+from api_types import (
+    GenerateVideoRequest,
+    GenerateVideoResponse,
+    ImageConditioningInput,
+    VideoCameraMotion,
+)
+from _routes._errors import HTTPError
+from handlers.base import StateHandlerBase
+from handlers.generation_handler import GenerationHandler
+from handlers.pipelines_handler import PipelinesHandler
+from handlers.text_handler import TextHandler
+from runtime_config.model_download_specs import resolve_model_path
+from server_utils.media_validation import (
+    normalize_optional_path,
+    validate_audio_file,
+    validate_image_file,
+)
+from services.interfaces import LTXAPIClient
+from state.app_state_types import AppState
+from state.app_settings import should_video_generate_with_ltx_api
+
+if TYPE_CHECKING:
+    from runtime_config.runtime_config import RuntimeConfig
+
+logger = logging.getLogger(__name__)
+
+FORCED_API_MODEL_MAP: dict[str, str] = {
+    "fast": "ltx-2-3-fast",
+    "pro": "ltx-2-3-pro",
+}
+FORCED_API_RESOLUTION_MAP: dict[str, dict[str, str]] = {
+    "1080p": {"16:9": "1920x1080", "9:16": "1080x1920"},
+    "1440p": {"16:9": "2560x1440", "9:16": "1440x2560"},
+    "2160p": {"16:9": "3840x2160", "9:16": "2160x3840"},
+}
+A2V_FORCED_API_RESOLUTION = "1920x1080"
+FORCED_API_ALLOWED_ASPECT_RATIOS = {"16:9", "9:16"}
+FORCED_API_ALLOWED_FPS = {24, 25, 48, 50}
+
+
+def _get_allowed_durations(model_id: str, resolution_label: str, fps: int) -> set[int]:
+    if model_id == "ltx-2-3-fast" and resolution_label == "1080p" and fps in {24, 25}:
+        return {6, 8, 10, 12, 14, 16, 18, 20}
+    return {6, 8, 10}
+
+
+class VideoGenerationHandler(StateHandlerBase):
+    def __init__(
+        self,
+        state: AppState,
+        lock: RLock,
+        generation_handler: GenerationHandler,
+        pipelines_handler: PipelinesHandler,
+        text_handler: TextHandler,
+        ltx_api_client: LTXAPIClient,
+        config: RuntimeConfig,
+    ) -> None:
+        super().__init__(state, lock, config)
+        self._generation = generation_handler
+        self._pipelines = pipelines_handler
+        self._text = text_handler
+        self._ltx_api_client = ltx_api_client
+
+    def generate(self, req: GenerateVideoRequest) -> GenerateVideoResponse:
+        if should_video_generate_with_ltx_api(
+            force_api_generations=self.config.force_api_generations,
+            settings=self.state.app_settings,
+        ):
+            return self._generate_forced_api(req)
+
+        if self._generation.is_generation_running():
+            raise HTTPError(409, "Generation already in progress")
+
+        resolution = req.resolution
+
+        duration = int(float(req.duration))
+        fps = int(float(req.fps))
+
+        audio_path = normalize_optional_path(req.audioPath)
+        if audio_path:
+            return self._generate_a2v(req, duration, fps, audio_path=audio_path)
+
+        logger.info("Resolution %s - using fast pipeline", resolution)
+
+        RESOLUTION_MAP_16_9: dict[str, tuple[int, int]] = {
+            "540p": (1024, 576),
+            "720p": (1280, 704),
+            "1080p": (1920, 1088),
+        }
+
+        def get_16_9_size(res: str) -> tuple[int, int]:
+            return RESOLUTION_MAP_16_9.get(res, (1280, 704))
+
+        def get_9_16_size(res: str) -> tuple[int, int]:
+            w, h = get_16_9_size(res)
+            return h, w
+
+        match req.aspectRatio:
+            case "9:16":
+                width, height = get_9_16_size(resolution)
+            case "16:9":
+                width, height = get_16_9_size(resolution)
+
+        num_frames = self._compute_num_frames(duration, fps)
+
+        image = None
+        image_path = normalize_optional_path(req.imagePath)
+        if image_path:
+            image = self._prepare_image(image_path, width, height)
+            logger.info("Image: %s -> %sx%s", image_path, width, height)
+
+        generation_id = self._make_generation_id()
+        seed = self._resolve_seed()
+
+        logger.info(
+            f"Request loraPath: '{req.loraPath}', loraStrength: {req.loraStrength}, inferenceSteps: {req.inferenceSteps}"
+        )
+
+        # 尝试支持自定义步数（实验性）
+        inference_steps = req.inferenceSteps
+        logger.info(f"Using inference steps: {inference_steps}")
+
+        loras = []
+        try:
+            import os
+            from ltx_core.loader import LoraPathStrengthAndSDOps
+            from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
+
+            # Handle legacy single LoRA
+            if req.loraPath and req.loraPath.strip():
+                lora_path = req.loraPath.strip()
+                if os.path.exists(lora_path):
+                    loras.append(
+                        LoraPathStrengthAndSDOps(
+                            path=lora_path,
+                            strength=req.loraStrength,
+                            sd_ops=LTXV_LORA_COMFY_RENAMING_MAP,
+                        )
+                    )
+            
+            # Handle multiple LoRAs
+            if req.loraPaths and req.loraStrengths:
+                for lp, ls in zip(req.loraPaths, req.loraStrengths):
+                    if lp and lp.strip():
+                        p = lp.strip()
+                        if os.path.exists(p):
+                            # Avoid duplicates if single LoRA was also in paths
+                            if not any(x.path == p for x in loras):
+                                loras.append(
+                                    LoraPathStrengthAndSDOps(
+                                        path=p,
+                                        strength=float(ls),
+                                        sd_ops=LTXV_LORA_COMFY_RENAMING_MAP,
+                                    )
+                                )
+                                logger.info(f"Multi-LoRA prepared: {p} with strength {ls}")
+                        else:
+                            logger.warning(f"Multi-LoRA file not found: {p}")
+        except Exception as e:
+            logger.warning(f"Failed to load LoRAs: {e}")
+            import traceback
+            logger.warning(f"LoRA traceback: {traceback.format_exc()}")
+            loras = []
+
+        if not loras:
+            loras = None
+
+        if loras is not None:
+            sig_list = []
+            for item in sorted(loras, key=lambda x: x.path):
+                sig_list.extend([item.path, round(float(item.strength), 4)])
+            desired_sig = ("fast", tuple(sig_list))
+        else:
+            desired_sig = ("fast", "", 0.0)
+
+        try:
+            if loras is not None:
+                # 强制卸载并重新加载带LoRA的pipeline
+                logger.info("Unloading pipeline for LoRA...")
+                from keep_models_runtime import force_unload_gpu_pipeline
+
+                force_unload_gpu_pipeline(self._pipelines)
+
+                # 强制垃圾回收
+                import gc
+
+                gc.collect()
+                # 释放 CUDA 缓存，降低 LoRA 首次构建的显存峰值/碎片风险
+                try:
+                    import torch
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+                        torch.cuda.ipc_collect()
+                except Exception:
+                    pass
+
+                gemma_root = self._pipelines._text_handler.resolve_gemma_root()
+                from runtime_config.model_download_specs import resolve_model_path
+                from services.fast_video_pipeline.ltx_fast_video_pipeline import (
+                    LTXFastVideoPipeline,
+                )
+
+                checkpoint_path = str(
+                    resolve_model_path(
+                        self._pipelines.models_dir,
+                        self._pipelines.config.model_download_specs,
+                        "checkpoint",
+                    )
+                )
+                upsampler_path = str(
+                    resolve_model_path(
+                        self._pipelines.models_dir,
+                        self._pipelines.config.model_download_specs,
+                        "upsampler",
+                    )
+                )
+
+                logger.info(
+                    f"Creating pipeline with LoRA: {loras}, steps: {inference_steps}"
+                )
+                from lora_injection import (
+                    _lora_init_kwargs,
+                    inject_loras_into_fast_pipeline,
+                )
+
+                lora_kw = _lora_init_kwargs(LTXFastVideoPipeline, loras)
+                pipeline = LTXFastVideoPipeline(
+                    checkpoint_path,
+                    gemma_root,
+                    upsampler_path,
+                    self._pipelines.config.device,
+                    **lora_kw,
+                )
+                n_inj = inject_loras_into_fast_pipeline(pipeline, loras)
+                if hasattr(pipeline, "pipeline") and hasattr(
+                    pipeline.pipeline, "model_ledger"
+                ):
+                    try:
+                        pipeline.pipeline.model_ledger.loras = tuple(loras)
+                    except Exception:
+                        pass
+                logger.info(
+                    "LoRA 注入: init_kw=%s, 注入点=%s, model_ledger.loras=%s",
+                    list(lora_kw.keys()),
+                    n_inj,
+                    getattr(
+                        getattr(pipeline.pipeline, "model_ledger", None),
+                        "loras",
+                        None,
+                    ),
+                )
+
+                from state.app_state_types import (
+                    VideoPipelineState,
+                    VideoPipelineWarmth,
+                    GpuSlot,
+                )
+
+                state = VideoPipelineState(
+                    pipeline=pipeline,
+                    warmth=VideoPipelineWarmth.COLD,
+                    is_compiled=False,
+                )
+
+                self._pipelines.state.gpu_slot = GpuSlot(
+                    active_pipeline=state, generation=None
+                )
+                logger.info("Pipeline with LoRA loaded successfully")
+            else:
+                # 无论有没有LoRA，都尝试使用自定义步数重新加载pipeline
+                logger.info(f"Loading pipeline with {inference_steps} steps")
+                from keep_models_runtime import force_unload_gpu_pipeline
+
+                force_unload_gpu_pipeline(self._pipelines)
+
+                import gc
+
+                gc.collect()
+
+                gemma_root = self._pipelines._text_handler.resolve_gemma_root()
+                from runtime_config.model_download_specs import resolve_model_path
+                from services.fast_video_pipeline.ltx_fast_video_pipeline import (
+                    LTXFastVideoPipeline,
+                )
+
+                checkpoint_path = str(
+                    resolve_model_path(
+                        self._pipelines.models_dir,
+                        self._pipelines.config.model_download_specs,
+                        "checkpoint",
+                    )
+                )
+                upsampler_path = str(
+                    resolve_model_path(
+                        self._pipelines.models_dir,
+                        self._pipelines.config.model_download_specs,
+                        "upsampler",
+                    )
+                )
+
+                pipeline = LTXFastVideoPipeline(
+                    checkpoint_path,
+                    gemma_root,
+                    upsampler_path,
+                    self._pipelines.config.device,
+                )
+
+                from state.app_state_types import (
+                    VideoPipelineState,
+                    VideoPipelineWarmth,
+                    GpuSlot,
+                )
+
+                state = VideoPipelineState(
+                    pipeline=pipeline,
+                    warmth=VideoPipelineWarmth.COLD,
+                    is_compiled=False,
+                )
+
+                self._pipelines.state.gpu_slot = GpuSlot(
+                    active_pipeline=state, generation=None
+                )
+
+            self._pipelines._pipeline_signature = desired_sig
+
+            self._generation.start_generation(generation_id)
+
+            output_path = self.generate_video(
+                prompt=req.prompt,
+                image=image,
+                height=height,
+                width=width,
+                num_frames=num_frames,
+                fps=fps,
+                seed=seed,
+                camera_motion=req.cameraMotion,
+                negative_prompt=req.negativePrompt,
+            )
+
+            self._generation.complete_generation(output_path)
+            return GenerateVideoResponse(status="complete", video_path=output_path)
+
+        except Exception as e:
+            self._generation.fail_generation(str(e))
+            if "cancelled" in str(e).lower():
+                logger.info("Generation cancelled by user")
+                return GenerateVideoResponse(status="cancelled")
+
+            raise HTTPError(500, str(e)) from e
+
+    def generate_video(
+        self,
+        prompt: str,
+        image: Image.Image | None,
+        height: int,
+        width: int,
+        num_frames: int,
+        fps: float,
+        seed: int,
+        camera_motion: VideoCameraMotion,
+        negative_prompt: str,
+    ) -> str:
+        t_total_start = time.perf_counter()
+        gen_mode = "i2v" if image is not None else "t2v"
+        logger.info(
+            "[%s] Generation started (model=fast, %dx%d, %d frames, %d fps)",
+            gen_mode,
+            width,
+            height,
+            num_frames,
+            int(fps),
+        )
+
+        if self._generation.is_generation_cancelled():
+            raise RuntimeError("Generation was cancelled")
+
+        if not resolve_model_path(
+            self.models_dir, self.config.model_download_specs, "checkpoint"
+        ).exists():
+            raise RuntimeError(
+                "Models not downloaded. Please download the AI models first using the Model Status menu."
+            )
+
+        total_steps = 8
+
+        self._generation.update_progress("loading_model", 5, 0, total_steps)
+        t_load_start = time.perf_counter()
+        pipeline_state = self._pipelines.load_gpu_pipeline("fast", should_warm=False)
+        t_load_end = time.perf_counter()
+        logger.info("[%s] Pipeline load: %.2fs", gen_mode, t_load_end - t_load_start)
+
+        self._generation.update_progress("encoding_text", 10, 0, total_steps)
+
+        enhanced_prompt = prompt + self.config.camera_motion_prompts.get(
+            camera_motion, ""
+        )
+
+        images: list[ImageConditioningInput] = []
+        temp_image_path: str | None = None
+        if image is not None:
+            temp_image_path = tempfile.NamedTemporaryFile(
+                suffix=".png", delete=False
+            ).name
+            image.save(temp_image_path)
+            images = [
+                ImageConditioningInput(path=temp_image_path, frame_idx=0, strength=1.0)
+            ]
+
+        output_path = self._make_output_path()
+
+        try:
+            settings = self.state.app_settings
+            use_api_encoding = not self._text.should_use_local_encoding()
+            if image is not None:
+                enhance = use_api_encoding and settings.prompt_enhancer_enabled_i2v
+            else:
+                enhance = use_api_encoding and settings.prompt_enhancer_enabled_t2v
+
+            encoding_method = "api" if use_api_encoding else "local"
+            t_text_start = time.perf_counter()
+            self._text.prepare_text_encoding(enhanced_prompt, enhance_prompt=enhance)
+            t_text_end = time.perf_counter()
+            logger.info(
+                "[%s] Text encoding (%s): %.2fs",
+                gen_mode,
+                encoding_method,
+                t_text_end - t_text_start,
+            )
+
+            self._generation.update_progress("inference", 15, 0, total_steps)
+
+            height = round(height / 64) * 64
+            width = round(width / 64) * 64
+
+            t_inference_start = time.perf_counter()
+            pipeline_state.pipeline.generate(
+                prompt=enhanced_prompt,
+                seed=seed,
+                height=height,
+                width=width,
+                num_frames=num_frames,
+                frame_rate=fps,
+                images=images,
+                output_path=str(output_path),
+            )
+            t_inference_end = time.perf_counter()
+            logger.info(
+                "[%s] Inference: %.2fs", gen_mode, t_inference_end - t_inference_start
+            )
+
+            if self._generation.is_generation_cancelled():
+                if output_path.exists():
+                    output_path.unlink()
+                raise RuntimeError("Generation was cancelled")
+
+            t_total_end = time.perf_counter()
+            logger.info(
+                "[%s] Total generation: %.2fs (load=%.2fs, text=%.2fs, inference=%.2fs)",
+                gen_mode,
+                t_total_end - t_total_start,
+                t_load_end - t_load_start,
+                t_text_end - t_text_start,
+                t_inference_end - t_inference_start,
+            )
+
+            self._generation.update_progress("complete", 100, total_steps, total_steps)
+            return str(output_path)
+        finally:
+            self._text.clear_api_embeddings()
+            if temp_image_path and os.path.exists(temp_image_path):
+                os.unlink(temp_image_path)
+
+    def _generate_a2v(
+        self, req: GenerateVideoRequest, duration: int, fps: int, *, audio_path: str
+    ) -> GenerateVideoResponse:
+        if req.model != "pro":
+            logger.warning(
+                "A2V local requested with model=%s; A2V always uses pro pipeline",
+                req.model,
+            )
+        validated_audio_path = validate_audio_file(audio_path)
+        audio_path_str = str(validated_audio_path)
+
+        # 支持竖屏和横屏
+        RESOLUTION_MAP: dict[str, tuple[int, int]] = {
+            "540p": (1024, 576),
+            "720p": (1280, 704),
+            "1080p": (1920, 1088),
+        }
+
+        base_w, base_h = RESOLUTION_MAP.get(req.resolution, (1280, 704))
+
+        # 根据 aspectRatio 调整分辨率
+        if req.aspectRatio == "9:16":
+            width, height = base_h, base_w  # 竖屏
+        else:
+            width, height = base_w, base_h  # 横屏
+
+        num_frames = self._compute_num_frames(duration, fps)
+
+        image = None
+        temp_image_path: str | None = None
+        image_path = normalize_optional_path(req.imagePath)
+        if image_path:
+            image = self._prepare_image(image_path, width, height)
+
+        # 获取首尾帧
+        start_frame_path = normalize_optional_path(getattr(req, "startFramePath", None))
+        end_frame_path = normalize_optional_path(getattr(req, "endFramePath", None))
+
+        seed = self._resolve_seed()
+
+        generation_id = self._make_generation_id()
+
+        temp_image_paths: list[str] = []
+        try:
+            a2v_state = self._pipelines.load_a2v_pipeline()
+            self._generation.start_generation(generation_id)
+
+            enhanced_prompt = req.prompt + self.config.camera_motion_prompts.get(
+                req.cameraMotion, ""
+            )
+            neg = (
+                req.negativePrompt
+                if req.negativePrompt
+                else self.config.default_negative_prompt
+            )
+
+            images: list[ImageConditioningInput] = []
+            temp_image_paths: list[str] = []
+
+            # 首帧
+            if start_frame_path:
+                start_img = self._prepare_image(start_frame_path, width, height)
+                temp_start_path = tempfile.NamedTemporaryFile(
+                    suffix=".png", delete=False
+                ).name
+                start_img.save(temp_start_path)
+                temp_image_paths.append(temp_start_path)
+                images.append(
+                    ImageConditioningInput(
+                        path=temp_start_path, frame_idx=0, strength=1.0
+                    )
+                )
+
+            # 中间图片（如果有）
+            if image is not None and not start_frame_path:
+                temp_image_path = tempfile.NamedTemporaryFile(
+                    suffix=".png", delete=False
+                ).name
+                image.save(temp_image_path)
+                temp_image_paths.append(temp_image_path)
+                images.append(
+                    ImageConditioningInput(
+                        path=temp_image_path, frame_idx=0, strength=1.0
+                    )
+                )
+
+            # 尾帧
+            if end_frame_path:
+                last_latent_idx = (num_frames - 1) // 8 + 1 - 1
+                end_img = self._prepare_image(end_frame_path, width, height)
+                temp_end_path = tempfile.NamedTemporaryFile(
+                    suffix=".png", delete=False
+                ).name
+                end_img.save(temp_end_path)
+                temp_image_paths.append(temp_end_path)
+                images.append(
+                    ImageConditioningInput(
+                        path=temp_end_path, frame_idx=last_latent_idx, strength=1.0
+                    )
+                )
+
+            output_path = self._make_output_path()
+
+            total_steps = 11  # distilled: 8 steps (stage 1) + 3 steps (stage 2)
+
+            a2v_settings = self.state.app_settings
+            a2v_use_api = not self._text.should_use_local_encoding()
+            if image is not None:
+                a2v_enhance = a2v_use_api and a2v_settings.prompt_enhancer_enabled_i2v
+            else:
+                a2v_enhance = a2v_use_api and a2v_settings.prompt_enhancer_enabled_t2v
+
+            self._generation.update_progress("loading_model", 5, 0, total_steps)
+            self._generation.update_progress("encoding_text", 10, 0, total_steps)
+            self._text.prepare_text_encoding(
+                enhanced_prompt, enhance_prompt=a2v_enhance
+            )
+            self._generation.update_progress("inference", 15, 0, total_steps)
+
+            a2v_state.pipeline.generate(
+                prompt=enhanced_prompt,
+                negative_prompt=neg,
+                seed=seed,
+                height=height,
+                width=width,
+                num_frames=num_frames,
+                frame_rate=fps,
+                num_inference_steps=total_steps,
+                images=images,
+                audio_path=audio_path_str,
+                audio_start_time=0.0,
+                audio_max_duration=None,
+                output_path=str(output_path),
+            )
+
+            if self._generation.is_generation_cancelled():
+                if output_path.exists():
+                    output_path.unlink()
+                raise RuntimeError("Generation was cancelled")
+
+            self._generation.update_progress("complete", 100, total_steps, total_steps)
+            self._generation.complete_generation(str(output_path))
+            return GenerateVideoResponse(status="complete", video_path=str(output_path))
+
+        except Exception as e:
+            self._generation.fail_generation(str(e))
+            if "cancelled" in str(e).lower():
+                logger.info("Generation cancelled by user")
+                return GenerateVideoResponse(status="cancelled")
+            raise HTTPError(500, str(e)) from e
+        finally:
+            self._text.clear_api_embeddings()
+            # 清理所有临时图片
+            for tmp_path in temp_image_paths:
+                if tmp_path and os.path.exists(tmp_path):
+                    try:
+                        os.unlink(tmp_path)
+                    except Exception:
+                        pass
+            if temp_image_path and os.path.exists(temp_image_path):
+                try:
+                    os.unlink(temp_image_path)
+                except Exception:
+                    pass
+
+    def _prepare_image(self, image_path: str, width: int, height: int) -> Image.Image:
+        validated_path = validate_image_file(image_path)
+        try:
+            img = Image.open(validated_path).convert("RGB")
+        except Exception:
+            raise HTTPError(400, f"Invalid image file: {image_path}") from None
+        img_w, img_h = img.size
+        target_ratio = width / height
+        img_ratio = img_w / img_h
+        if img_ratio > target_ratio:
+            new_h = height
+            new_w = int(img_w * (height / img_h))
+        else:
+            new_w = width
+            new_h = int(img_h * (width / img_w))
+        resized = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
+        left = (new_w - width) // 2
+        top = (new_h - height) // 2
+        return resized.crop((left, top, left + width, top + height))
+
+    @staticmethod
+    def _make_generation_id() -> str:
+        return uuid.uuid4().hex[:8]
+
+    @staticmethod
+    def _compute_num_frames(duration: int, fps: int) -> int:
+        n = ((duration * fps) // 8) * 8 + 1
+        return max(n, 9)
+
+    def _resolve_seed(self) -> int:
+        settings = self.state.app_settings
+        if settings.seed_locked:
+            logger.info("Using locked seed: %s", settings.locked_seed)
+            return settings.locked_seed
+        return int(time.time()) % 2147483647
+
+    def _make_output_path(self) -> Path:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        return (
+            self.config.outputs_dir
+            / f"ltx2_video_{timestamp}_{self._make_generation_id()}.mp4"
+        )
+
+    def _generate_forced_api(self, req: GenerateVideoRequest) -> GenerateVideoResponse:
+        if self._generation.is_generation_running():
+            raise HTTPError(409, "Generation already in progress")
+
+        generation_id = self._make_generation_id()
+        self._generation.start_api_generation(generation_id)
+
+        audio_path = normalize_optional_path(req.audioPath)
+        image_path = normalize_optional_path(req.imagePath)
+        has_input_audio = bool(audio_path)
+        has_input_image = bool(image_path)
+
+        try:
+            self._generation.update_progress("validating_request", 5, None, None)
+
+            api_key = self.state.app_settings.ltx_api_key.strip()
+            logger.info(
+                "Forced API generation route selected (key_present=%s)", bool(api_key)
+            )
+            if not api_key:
+                raise HTTPError(400, "PRO_API_KEY_REQUIRED")
+
+            requested_model = req.model.strip().lower()
+            api_model_id = FORCED_API_MODEL_MAP.get(requested_model)
+            if api_model_id is None:
+                raise HTTPError(400, "INVALID_FORCED_API_MODEL")
+
+            resolution_label = req.resolution
+            resolution_by_aspect = FORCED_API_RESOLUTION_MAP.get(resolution_label)
+            if resolution_by_aspect is None:
+                raise HTTPError(400, "INVALID_FORCED_API_RESOLUTION")
+
+            aspect_ratio = req.aspectRatio.strip()
+            if aspect_ratio not in FORCED_API_ALLOWED_ASPECT_RATIOS:
+                raise HTTPError(400, "INVALID_FORCED_API_ASPECT_RATIO")
+
+            api_resolution = resolution_by_aspect[aspect_ratio]
+
+            prompt = req.prompt
+
+            if self._generation.is_generation_cancelled():
+                raise RuntimeError("Generation was cancelled")
+
+            if has_input_audio:
+                if requested_model != "pro":
+                    logger.warning(
+                        "A2V requested with model=%s; overriding to 'pro'",
+                        requested_model,
+                    )
+                api_model_id = FORCED_API_MODEL_MAP["pro"]
+                if api_resolution != A2V_FORCED_API_RESOLUTION:
+                    logger.warning(
+                        "A2V requested with resolution=%s; overriding to '%s'",
+                        api_resolution,
+                        A2V_FORCED_API_RESOLUTION,
+                    )
+                api_resolution = A2V_FORCED_API_RESOLUTION
+                validated_audio_path = validate_audio_file(audio_path)
+                validated_image_path: Path | None = None
+                if image_path is not None:
+                    validated_image_path = validate_image_file(image_path)
+
+                self._generation.update_progress("uploading_audio", 20, None, None)
+                audio_uri = self._ltx_api_client.upload_file(
+                    api_key=api_key,
+                    file_path=str(validated_audio_path),
+                )
+                image_uri: str | None = None
+                if validated_image_path is not None:
+                    self._generation.update_progress("uploading_image", 35, None, None)
+                    image_uri = self._ltx_api_client.upload_file(
+                        api_key=api_key,
+                        file_path=str(validated_image_path),
+                    )
+                self._generation.update_progress("inference", 55, None, None)
+                video_bytes = self._ltx_api_client.generate_audio_to_video(
+                    api_key=api_key,
+                    prompt=prompt,
+                    audio_uri=audio_uri,
+                    image_uri=image_uri,
+                    model=api_model_id,
+                    resolution=api_resolution,
+                )
+                self._generation.update_progress("downloading_output", 85, None, None)
+            elif has_input_image:
+                validated_image_path = validate_image_file(image_path)
+
+                duration = self._parse_forced_numeric_field(
+                    req.duration, "INVALID_FORCED_API_DURATION"
+                )
+                fps = self._parse_forced_numeric_field(
+                    req.fps, "INVALID_FORCED_API_FPS"
+                )
+                if fps not in FORCED_API_ALLOWED_FPS:
+                    raise HTTPError(400, "INVALID_FORCED_API_FPS")
+                if duration not in _get_allowed_durations(
+                    api_model_id, resolution_label, fps
+                ):
+                    raise HTTPError(400, "INVALID_FORCED_API_DURATION")
+
+                generate_audio = self._parse_audio_flag(req.audio)
+                self._generation.update_progress("uploading_image", 20, None, None)
+                image_uri = self._ltx_api_client.upload_file(
+                    api_key=api_key,
+                    file_path=str(validated_image_path),
+                )
+                self._generation.update_progress("inference", 55, None, None)
+                video_bytes = self._ltx_api_client.generate_image_to_video(
+                    api_key=api_key,
+                    prompt=prompt,
+                    image_uri=image_uri,
+                    model=api_model_id,
+                    resolution=api_resolution,
+                    duration=float(duration),
+                    fps=float(fps),
+                    generate_audio=generate_audio,
+                    camera_motion=req.cameraMotion,
+                )
+                self._generation.update_progress("downloading_output", 85, None, None)
+            else:
+                duration = self._parse_forced_numeric_field(
+                    req.duration, "INVALID_FORCED_API_DURATION"
+                )
+                fps = self._parse_forced_numeric_field(
+                    req.fps, "INVALID_FORCED_API_FPS"
+                )
+                if fps not in FORCED_API_ALLOWED_FPS:
+                    raise HTTPError(400, "INVALID_FORCED_API_FPS")
+                if duration not in _get_allowed_durations(
+                    api_model_id, resolution_label, fps
+                ):
+                    raise HTTPError(400, "INVALID_FORCED_API_DURATION")
+
+                generate_audio = self._parse_audio_flag(req.audio)
+                self._generation.update_progress("inference", 55, None, None)
+                video_bytes = self._ltx_api_client.generate_text_to_video(
+                    api_key=api_key,
+                    prompt=prompt,
+                    model=api_model_id,
+                    resolution=api_resolution,
+                    duration=float(duration),
+                    fps=float(fps),
+                    generate_audio=generate_audio,
+                    camera_motion=req.cameraMotion,
+                )
+                self._generation.update_progress("downloading_output", 85, None, None)
+
+            if self._generation.is_generation_cancelled():
+                raise RuntimeError("Generation was cancelled")
+
+            output_path = self._write_forced_api_video(video_bytes)
+            if self._generation.is_generation_cancelled():
+                output_path.unlink(missing_ok=True)
+                raise RuntimeError("Generation was cancelled")
+
+            self._generation.update_progress("complete", 100, None, None)
+            self._generation.complete_generation(str(output_path))
+            return GenerateVideoResponse(status="complete", video_path=str(output_path))
+        except HTTPError as e:
+            self._generation.fail_generation(e.detail)
+            raise
+        except Exception as e:
+            self._generation.fail_generation(str(e))
+            if "cancelled" in str(e).lower():
+                logger.info("Generation cancelled by user")
+                return GenerateVideoResponse(status="cancelled")
+            raise HTTPError(500, str(e)) from e
+
+    def _write_forced_api_video(self, video_bytes: bytes) -> Path:
+        output_path = self._make_output_path()
+        output_path.write_bytes(video_bytes)
+        return output_path
+
+    @staticmethod
+    def _parse_forced_numeric_field(raw_value: str, error_detail: str) -> int:
+        try:
+            return int(float(raw_value))
+        except (TypeError, ValueError):
+            raise HTTPError(400, error_detail) from None
+
+    @staticmethod
+    def _parse_audio_flag(audio_value: str | bool) -> bool:
+        if isinstance(audio_value, bool):
+            return audio_value
+        normalized = audio_value.strip().lower()
+        return normalized in {"1", "true", "yes", "on"}

LTX2.3-1.0.4-new/patches/keep_models_runtime.py

@@ -0,0 +1,16 @@
+"""仅提供强制卸载 GPU 管线。「保持模型加载」功能已移除。"""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+def force_unload_gpu_pipeline(pipelines: Any) -> None:
+    """释放推理管线占用的显存（切换 GPU、清理、LoRA 重建等场景）。"""
+    try:
+        pipelines.unload_gpu_pipeline()
+    except Exception:
+        try:
+            type(pipelines).unload_gpu_pipeline(pipelines)
+        except Exception:
+            pass

LTX2.3-1.0.4-new/patches/launcher.py

@@ -0,0 +1,20 @@
+
+import sys
+import os
+
+patch_dir = r"C:\Users\1-xuanran\Desktop\LTX队列\patches"
+backend_dir = r"C:\Program Files\LTX Desktop\resources\backend"
+
+# 防御性清除：强行剥离所有的默认 backend_dir 引用
+sys.path = [p for p in sys.path if p and os.path.normpath(p) != os.path.normpath(backend_dir)]
+sys.path = [p for p in sys.path if p and p != "." and p != ""]
+
+# 绝对插队注入：优先搜索 PATCHES_DIR
+sys.path.insert(0, patch_dir)
+sys.path.insert(1, backend_dir)
+
+import uvicorn
+from ltx2_server import app
+
+if __name__ == '__main__':
+    uvicorn.run(app, host="0.0.0.0", port=3000, log_level="info", access_log=False)

LTX2.3-1.0.4-new/patches/lora_build_hook.py

@@ -0,0 +1,172 @@
+"""
+在 SingleGPUModelBuilder.build() 时合并「当前请求」的用户 LoRA。
+
+桌面版 Fast 管线往往只在 model_ledger 上挂 loras，真正 load 权重时仍用
+初始化时的空 loras Builder；此处对 DiT/Transformer 的 Builder 在 build 前注入。
+"""
+
+from __future__ import annotations
+
+import contextvars
+import logging
+from dataclasses import replace
+from typing import Any
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+# 当前 HTTP 请求/生成任务中要额外融合的 LoRA（LoraPathStrengthAndSDOps 元组）
+_pending_user_loras: contextvars.ContextVar[tuple[Any, ...] | None] = contextvars.ContextVar(
+    "ltx_pending_user_loras", default=None
+)
+
+_HOOK_INSTALLED = False
+_FP8_LORA_PATCH_INSTALLED = False
+
+
+def pending_loras_token(loras: tuple[Any, ...] | None):
+    """返回 contextvar Token，供 finally reset；loras 为 None 表示本任务不用额外 LoRA。"""
+    return _pending_user_loras.set(loras)
+
+
+def reset_pending_loras(token: contextvars.Token | None) -> None:
+    if token is not None:
+        _pending_user_loras.reset(token)
+
+
+def _get_pending() -> tuple[Any, ...] | None:
+    return _pending_user_loras.get()
+
+
+def _is_ltx_diffusion_transformer_builder(builder: Any) -> bool:
+    """避免给 Gemma / VAE / Upsampler 的 Builder 误加视频 LoRA。"""
+    cfg = getattr(builder, "model_class_configurator", None)
+    if cfg is None:
+        return False
+    name = getattr(cfg, "__name__", "") or ""
+    # 排除明显非 DiT 的
+    for bad in (
+        "Gemma",
+        "VideoEncoder",
+        "VideoDecoder",
+        "AudioEncoder",
+        "AudioDecoder",
+        "Vocoder",
+        "EmbeddingsProcessor",
+        "LatentUpsampler",
+    ):
+        if bad in name:
+            return False
+    try:
+        from ltx_core.model.transformer import LTXModelConfigurator
+
+        if isinstance(cfg, type):
+            try:
+                if issubclass(cfg, LTXModelConfigurator):
+                    return True
+            except TypeError:
+                pass
+        if cfg is LTXModelConfigurator:
+            return True
+    except ImportError:
+        pass
+    # 兜底：LTX 主 transformer 配置器命名习惯（排除已列出的 VAE/Gemma）
+    return "LTX" in name and "ModelConfigurator" in name
+
+
+def _install_fp8_lora_fusion_patch() -> None:
+    """Make LTX's scaled-FP8 LoRA fusion tolerant of checkpoint layout variants."""
+    global _FP8_LORA_PATCH_INSTALLED
+    if _FP8_LORA_PATCH_INSTALLED:
+        return
+    try:
+        import ltx_core.loader.fuse_loras as fuse_mod
+    except ImportError:
+        return
+
+    _orig_scaled = getattr(fuse_mod, "_fuse_delta_with_scaled_fp8", None)
+    if _orig_scaled is None:
+        return
+
+    def _quantize_preserve_layout(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        tensor_fp32 = tensor.to(torch.float32)
+        fp8_min = torch.finfo(torch.float8_e4m3fn).min
+        fp8_max = torch.finfo(torch.float8_e4m3fn).max
+        max_abs = torch.amax(torch.abs(tensor_fp32))
+        if max_abs == 0:
+            max_abs = torch.ones((), dtype=torch.float32, device=tensor_fp32.device)
+        scale = fp8_max / max_abs
+        quantized = torch.clamp(tensor_fp32 * scale, min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+        return quantized, scale.reciprocal()
+
+    def _patched_scaled(deltas: torch.Tensor, weight: torch.Tensor, key: str, scale_key: str, model_sd: Any) -> dict[str, torch.Tensor]:
+        weight_scale = model_sd.sd[scale_key].to(device=weight.device)
+        delta = deltas.to(device=weight.device, dtype=torch.float32)
+        weight_fp32 = weight.to(torch.float32)
+
+        # Standard LTX scaled-FP8 layout: checkpoint stores (in, out), LoRA delta is (out, in).
+        normal_layout = weight_fp32.t() * weight_scale
+        if normal_layout.shape == delta.shape:
+            new_weight = normal_layout + delta
+            new_fp8_weight, new_weight_scale = fuse_mod.quantize_weight_to_fp8_per_tensor(new_weight)
+            return {key: new_fp8_weight, scale_key: new_weight_scale}
+        if normal_layout.shape == delta.t().shape:
+            new_weight = normal_layout + delta.t()
+            new_fp8_weight, new_weight_scale = fuse_mod.quantize_weight_to_fp8_per_tensor(new_weight)
+            return {key: new_fp8_weight, scale_key: new_weight_scale}
+
+        # Some FP8 checkpoints already arrive in the module/storage layout.
+        storage_layout = weight_fp32 * weight_scale
+        if storage_layout.shape == delta.shape:
+            new_weight = storage_layout + delta
+            new_fp8_weight, new_weight_scale = _quantize_preserve_layout(new_weight)
+            return {key: new_fp8_weight, scale_key: new_weight_scale}
+        if storage_layout.shape == delta.t().shape:
+            new_weight = storage_layout + delta.t()
+            new_fp8_weight, new_weight_scale = _quantize_preserve_layout(new_weight)
+            return {key: new_fp8_weight, scale_key: new_weight_scale}
+
+        print(
+            "[PATCH] FP8 LoRA shape mismatch, skip layer: "
+            f"{key}, weight={tuple(weight.shape)}, delta={tuple(deltas.shape)}, "
+            f"normal={tuple(normal_layout.shape)}, storage={tuple(storage_layout.shape)}"
+        )
+        return {}
+
+    fuse_mod._fuse_delta_with_scaled_fp8 = _patched_scaled
+    _FP8_LORA_PATCH_INSTALLED = True
+    logger.info("lora_build_hook: 已挂载 scaled-FP8 LoRA 融合兼容补丁")
+
+
+def install_lora_build_hook() -> None:
+    global _HOOK_INSTALLED
+    _install_fp8_lora_fusion_patch()
+    if _HOOK_INSTALLED:
+        return
+    try:
+        from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
+    except ImportError:
+        logger.warning("lora_build_hook: 无法导入 SingleGPUModelBuilder，跳过")
+        return
+
+    _orig_build = SingleGPUModelBuilder.build
+
+    def build(self: Any, *args: Any, **kwargs: Any) -> Any:
+        extra = _get_pending()
+        if extra and _is_ltx_diffusion_transformer_builder(self):
+            have = {getattr(x, "path", None) for x in self.loras}
+            add = tuple(x for x in extra if getattr(x, "path", None) not in have)
+            if add:
+                merged = (*tuple(self.loras), *add)
+                self = replace(self, loras=merged)
+                logger.info(
+                    "lora_build_hook: 已向 DiT Builder 合并 %d 个用户 LoRA: %s",
+                    len(add),
+                    [getattr(x, "path", x) for x in add],
+                )
+        return _orig_build(self, *args, **kwargs)
+
+    SingleGPUModelBuilder.build = build  # type: ignore[method-assign]
+    _HOOK_INSTALLED = True
+    logger.info("lora_build_hook: 已挂载 SingleGPUModelBuilder.build")

LTX2.3-1.0.4-new/patches/lora_injection.py

@@ -0,0 +1,139 @@
+"""将用户 LoRA 注入 Fast 视频管线：兼容 ModelLedger 与 LTX-2 DiffusionStage/Builder。"""
+
+from __future__ import annotations
+
+import inspect
+import logging
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+
+def _lora_init_kwargs(
+    pipeline_cls: type, loras: list[Any] | tuple[Any, ...]
+) -> dict[str, Any]:
+    if not loras:
+        return {}
+    try:
+        sig = inspect.signature(pipeline_cls.__init__)
+        names = sig.parameters.keys()
+    except (TypeError, ValueError):
+        return {}
+    tup = tuple(loras)
+    for key in ("loras", "lora", "extra_loras", "user_loras"):
+        if key in names:
+            return {key: tup}
+    return {}
+
+
+def inject_loras_into_fast_pipeline(ltx_pipe: Any, loras: list[Any] | tuple[Any, ...]) -> int:
+    """在已构造的管线上尽量把 LoRA 写进会参与 build 的 Builder / ledger。返回成功写入的处数。"""
+    if not loras:
+        return 0
+    tup = tuple(loras)
+    patched = 0
+    visited: set[int] = set()
+
+    def visit(obj: Any, depth: int) -> None:
+        nonlocal patched
+        if obj is None or depth > 10:
+            return
+        oid = id(obj)
+        if oid in visited:
+            return
+        visited.add(oid)
+
+        # ModelLedger.loras（旧桌面）
+        ml = getattr(obj, "model_ledger", None)
+        if ml is not None:
+            try:
+                ml.loras = tup
+                patched += 1
+                logger.info("LoRA: 已设置 model_ledger.loras")
+            except Exception as e:
+                logger.debug("model_ledger.loras: %s", e)
+
+        # SingleGPUModelBuilder.with_loras（常见与变体属性名）
+        for holder in (obj, ml):
+            if holder is None:
+                continue
+            candidates: list[Any] = []
+            for attr in (
+                "_transformer_builder",
+                "transformer_builder",
+                "_model_builder",
+                "model_builder",
+            ):
+                tb = getattr(holder, attr, None)
+                if tb is not None:
+                    candidates.append((attr, tb))
+            try:
+                for attr in dir(holder):
+                    al = attr.lower()
+                    if "transformer" in al and "builder" in al and attr not in (
+                        "_transformer_builder",
+                        "transformer_builder",
+                    ):
+                        tb = getattr(holder, attr, None)
+                        if tb is not None:
+                            candidates.append((attr, tb))
+            except Exception:
+                pass
+            for attr, tb in candidates:
+                if hasattr(tb, "with_loras"):
+                    try:
+                        new_tb = tb.with_loras(tup)
+                        setattr(holder, attr, new_tb)
+                        patched += 1
+                        logger.info("LoRA: 已更新 %s.with_loras", attr)
+                    except Exception as e:
+                        logger.debug("with_loras %s: %s", attr, e)
+
+        # DiffusionStage（类名或 isinstance）
+        is_diffusion = type(obj).__name__ == "DiffusionStage"
+        if not is_diffusion:
+            try:
+                from ltx_pipelines.utils.blocks import DiffusionStage as _DS
+
+                is_diffusion = isinstance(obj, _DS)
+            except ImportError:
+                pass
+        if is_diffusion:
+            tb = getattr(obj, "_transformer_builder", None)
+            if tb is not None and hasattr(tb, "with_loras"):
+                try:
+                    obj._transformer_builder = tb.with_loras(tup)
+                    patched += 1
+                    logger.info("LoRA: 已写入 DiffusionStage._transformer_builder")
+                except Exception as e:
+                    logger.debug("DiffusionStage: %s", e)
+
+        # 常见嵌套属性
+        for name in (
+            "pipeline",
+            "inner",
+            "_inner",
+            "fast_pipeline",
+            "_pipeline",
+            "stage_1",
+            "stage_2",
+            "stage",
+            "_stage",
+            "stages",
+            "diffusion",
+            "_diffusion",
+        ):
+            try:
+                ch = getattr(obj, name, None)
+            except Exception:
+                continue
+            if ch is not None and ch is not obj:
+                visit(ch, depth + 1)
+
+        if isinstance(obj, (list, tuple)):
+            for item in obj[:8]:
+                visit(item, depth + 1)
+
+    root = getattr(ltx_pipe, "pipeline", ltx_pipe)
+    visit(root, 0)
+    return patched

LTX2.3-1.0.4-new/patches/low_vram_runtime.py

@@ -0,0 +1,264 @@
+"""低显存模式：尽量降峰值显存（以速度换显存）；效果取决于官方管线是否支持 offload。"""
+
+from __future__ import annotations
+
+import gc
+import logging
+import os
+import types
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger("ltx_low_vram")
+
+
+def _ltx_desktop_config_dir() -> Path:
+    p = (
+        Path(os.environ.get("LOCALAPPDATA", os.path.expanduser("~/AppData/Local")))
+        / "LTXDesktop"
+    )
+    p.mkdir(parents=True, exist_ok=True)
+    return p.resolve()
+
+
+def low_vram_pref_path() -> Path:
+    return _ltx_desktop_config_dir() / "low_vram_mode.pref"
+
+
+def read_low_vram_pref() -> bool:
+    f = low_vram_pref_path()
+    if not f.is_file():
+        return False
+    return f.read_text(encoding="utf-8").strip().lower() in ("1", "true", "yes", "on")
+
+
+def write_low_vram_pref(enabled: bool) -> None:
+    low_vram_pref_path().write_text(
+        "true\n" if enabled else "false\n", encoding="utf-8"
+    )
+
+
+def apply_low_vram_config_tweaks(handler: Any) -> None:
+    """在官方 RuntimeConfig 上尽量关闭 fast 超分等（若字段存在）。"""
+    cfg = getattr(handler, "config", None)
+    if cfg is None:
+        return
+    fm = getattr(cfg, "fast_model", None)
+    if fm is None:
+        return
+    try:
+        if hasattr(fm, "model_copy"):
+            updated = fm.model_copy(update={"use_upscaler": False})
+            setattr(cfg, "fast_model", updated)
+        elif hasattr(fm, "use_upscaler"):
+            setattr(fm, "use_upscaler", False)
+    except Exception as e:
+        logger.debug("low_vram: 无法关闭 fast_model.use_upscaler: %s", e)
+
+
+def restore_full_vram_config_tweaks(handler: Any) -> None:
+    """显存上限为 0 时恢复速度优先配置。"""
+    cfg = getattr(handler, "config", None)
+    if cfg is None:
+        return
+    fm = getattr(cfg, "fast_model", None)
+    if fm is None:
+        return
+    try:
+        if hasattr(fm, "model_copy"):
+            updated = fm.model_copy(update={"use_upscaler": True})
+            setattr(cfg, "fast_model", updated)
+        elif hasattr(fm, "use_upscaler"):
+            setattr(fm, "use_upscaler", True)
+    except Exception as e:
+        logger.debug("low_vram: 无法恢复 fast_model.use_upscaler: %s", e)
+
+
+def install_low_vram_on_pipelines(handler: Any) -> None:
+    """启动时读取偏好，挂到 pipelines 上供各补丁读取。"""
+    pl = handler.pipelines
+    low = read_low_vram_pref() and should_use_cpu_offload()
+    setattr(pl, "low_vram_mode", bool(low))
+    if low:
+        apply_low_vram_config_tweaks(handler)
+        logger.info(
+            "low_vram_mode: 已开启（尝试关闭 fast 超分；若显存仍高，多为权重常驻 GPU，需降分辨率/时长或 FP8 权重）"
+        )
+    else:
+        restore_full_vram_config_tweaks(handler)
+
+
+def install_low_vram_pipeline_hooks(pl: Any) -> None:
+    """在 load_gpu_pipeline / load_a2v 返回后尝试 Diffusers 式 CPU offload（无则静默）。"""
+    if getattr(pl, "_ltx_low_vram_hooks_installed", False):
+        return
+    pl._ltx_low_vram_hooks_installed = True
+
+    if hasattr(pl, "load_gpu_pipeline"):
+        _orig_gpu = pl.load_gpu_pipeline
+        pl._ltx_orig_load_gpu_for_low_vram = _orig_gpu
+
+        def _load_gpu_wrapped(self: Any, *a: Any, **kw: Any) -> Any:
+            r = _orig_gpu(*a, **kw)
+            if getattr(self, "low_vram_mode", False):
+                try_sequential_offload_on_pipeline_state(r)
+            return r
+
+        pl.load_gpu_pipeline = types.MethodType(_load_gpu_wrapped, pl)
+
+    if hasattr(pl, "load_a2v_pipeline"):
+        _orig_a2v = pl.load_a2v_pipeline
+        pl._ltx_orig_load_a2v_for_low_vram = _orig_a2v
+
+        def _load_a2v_wrapped(self: Any, *a: Any, **kw: Any) -> Any:
+            r = _orig_a2v(*a, **kw)
+            if getattr(self, "low_vram_mode", False):
+                try_sequential_offload_on_pipeline_state(r)
+            return r
+
+        pl.load_a2v_pipeline = types.MethodType(_load_a2v_wrapped, pl)
+
+    # Monkey patch: 接管 1.0.3 新增的底层 layer streaming 来实现完美的线性显存控制
+    if not getattr(pl, "_ltx_layer_streaming_patched", False):
+        pl._ltx_layer_streaming_patched = True
+        try:
+            def _patch_pipeline_class(cls_name, mod_name):
+                import importlib
+                try:
+                    mod = importlib.import_module(mod_name)
+                    pipeline_cls = getattr(mod, cls_name)
+                    _orig_call = pipeline_cls.__call__
+                    
+                    def _patched_call(self, *args, **kwargs):
+                        lim = get_vram_limit()
+                        if lim is not None:
+                            count = get_streaming_prefetch_count()
+                            kwargs["streaming_prefetch_count"] = count
+                            if count is None:
+                                logger.info(
+                                    "low_vram_mode: VRAM limit is unlimited/high. Disabled layer streaming."
+                                )
+                            else:
+                                logger.info(
+                                    "low_vram_mode: Dynamically tuned layer streaming prefetch count to %s for %sGB limit.",
+                                    count,
+                                    lim,
+                                )
+                                
+                        return _orig_call(self, *args, **kwargs)
+                        
+                    pipeline_cls.__call__ = _patched_call
+                    logger.info(f"low_vram_mode: Successfully patched {cls_name} to override streaming_prefetch_count")
+                except Exception as e:
+                    pass
+
+            _patch_pipeline_class("DistilledPipeline", "ltx_pipelines.distilled")
+            _patch_pipeline_class("TI2VidTwoStagesPipeline", "ltx_pipelines.ti2vid_two_stages")
+            _patch_pipeline_class("LTXRetakePipeline", "services.retake_pipeline.ltx_retake_pipeline")
+            _patch_pipeline_class("ICLoRAPipeline", "services.ic_lora_pipeline.ltx_ic_lora_pipeline")
+            _patch_pipeline_class("A2VPipeline", "services.a2v_pipeline.distilled_a2v_pipeline")
+        except Exception:
+            pass
+
+
+def get_vram_limit() -> float | None:
+    try:
+        import json
+        from pathlib import Path
+        settings_file = _ltx_desktop_config_dir() / "settings.json"
+        if settings_file.exists():
+            with open(settings_file, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            if "vram_limit" in data:
+                lim = data["vram_limit"]
+                if lim != "":
+                    return float(lim)
+    except Exception:
+        pass
+    return None
+
+
+def get_streaming_prefetch_count() -> int | None:
+    """把设置里的显存上限映射为 layer streaming 强度。
+
+    ``0`` 或留空表示纯 GPU / 速度优先，不启用层流式加载。
+    """
+    lim = get_vram_limit()
+    if lim is None or lim == 0:
+        return None
+    if lim <= 10.0:
+        return 1
+    if lim >= 25.0:
+        return None
+    extra_gb = float(lim) - 10.0
+    return max(1, min(32, 1 + round(extra_gb / 0.67)))
+
+
+def should_use_cpu_offload() -> bool:
+    """只有设置了大于 0 的显存上限时才启用 CPU/offload 慢速兼容路径。"""
+    lim = get_vram_limit()
+    return lim is not None and lim > 0
+
+
+def try_sequential_offload_on_pipeline_state(state: Any) -> None:
+    """按设定最高显存分配，爆显存后写入系统内存"""
+    if state is None:
+        return
+    if not should_use_cpu_offload():
+        logger.info(
+            "low_vram_mode: VRAM limit is 0/blank. Skip CPU offload for pure GPU speed."
+        )
+        return
+    root = getattr(state, "pipeline", state)
+    candidates: list[Any] = [root]
+    inner = getattr(root, "pipeline", None)
+    if inner is not None and inner is not root:
+        candidates.append(inner)
+        
+    # Capped-VRAM mode applies macro offload so T5/VAE can leave GPU while DiT runs.
+    # Pure GPU mode returns above and keeps the old fast path.
+    for obj in candidates:
+        for method_name in (
+            "enable_model_cpu_offload",
+            "enable_sequential_cpu_offload",
+        ):
+            fn = getattr(obj, method_name, None)
+            if callable(fn):
+                try:
+                    fn()
+                    logger.info(
+                        "low_vram_mode: 已对管线调用 %s()",
+                        method_name,
+                    )
+                    return
+                except Exception as e:
+                    logger.debug(
+                        "low_vram_mode: %s() 失败（可忽略）: %s",
+                        method_name,
+                        e,
+                    )
+
+
+def maybe_release_pipeline_after_task(handler: Any) -> None:
+    """单次生成结束后：低显存模式下强制卸载管线并回收缓存。"""
+    pl = getattr(handler, "pipelines", None) or getattr(handler, "_pipelines", None)
+    if pl is None or not getattr(pl, "low_vram_mode", False):
+        return
+    try:
+        from keep_models_runtime import force_unload_gpu_pipeline
+
+        force_unload_gpu_pipeline(pl)
+    except Exception as e:
+        logger.debug("low_vram_mode: 任务后卸载失败: %s", e)
+    try:
+        pl._pipeline_signature = None
+    except Exception:
+        pass
+    gc.collect()
+    try:
+        import torch
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    except Exception:
+        pass

LTX2.3-1.0.4-new/patches/ltx_dev_video_pipeline.py

@@ -0,0 +1,156 @@
+"""Patch-side wrapper for LTX dev checkpoints.
+
+The desktop Fast wrapper is built around ``DistilledPipeline``.  Dev checkpoints
+need the full TI2V two-stage pipeline; otherwise LoRA keys can match the wrong
+stage shape and fail during FP8 fusion.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from pathlib import Path
+from typing import Final
+
+import torch
+
+from api_types import ImageConditioningInput
+from services.ltx_pipeline_common import (
+    default_tiling_config,
+    encode_video_output,
+    video_chunks_number,
+)
+from services.services_utils import AudioOrNone, TilingConfigType, device_supports_fp8
+
+
+class LTXDevVideoPipeline:
+    pipeline_kind: Final = "dev"
+
+    def __init__(
+        self,
+        checkpoint_path: str,
+        gemma_root: str | None,
+        upsampler_path: str,
+        distilled_lora_path: str,
+        device: torch.device,
+        loras: list[object] | tuple[object, ...] | None = None,
+    ) -> None:
+        from ltx_core.loader import LoraPathStrengthAndSDOps
+        from ltx_core.quantization import QuantizationPolicy
+        from ltx_pipelines.ti2vid_two_stages import TI2VidTwoStagesPipeline
+        from ltx_pipelines.utils.constants import detect_params
+
+        self._checkpoint_path = checkpoint_path
+        self._device = device
+        self._params = detect_params(checkpoint_path)
+
+        quantization = None
+        if "fp8" in checkpoint_path.lower() and device_supports_fp8(device):
+            try:
+                quantization = QuantizationPolicy.fp8_scaled_mm()
+            except Exception as exc:
+                print(f"[PATCH] Dev FP8 scaled-mm 不可用，回退 fp8_cast: {exc}")
+                quantization = QuantizationPolicy.fp8_cast()
+
+        distilled_lora = []
+        checkpoint_name = Path(checkpoint_path).name.lower()
+        distilled_lora_name = Path(distilled_lora_path).name.lower() if distilled_lora_path else ""
+        incompatible_builtin_lora = (
+            "2.3" in checkpoint_name
+            and ("2-19b" in distilled_lora_name or "19b" in distilled_lora_name)
+        )
+        if incompatible_builtin_lora:
+            print(
+                "[PATCH] Dev two-stage: 跳过不匹配的内置 distilled LoRA "
+                f"({distilled_lora_name})，当前 checkpoint 是 {checkpoint_name}"
+            )
+        elif distilled_lora_path and Path(distilled_lora_path).is_file():
+            distilled_lora = [
+                LoraPathStrengthAndSDOps(
+                    path=distilled_lora_path,
+                    strength=1.0,
+                    sd_ops=None,
+                )
+            ]
+        elif distilled_lora_path:
+            print(
+                "[PATCH] Dev two-stage: distilled LoRA 不存在，跳过内置 stage-2 distilled LoRA: "
+                f"{distilled_lora_path}"
+            )
+
+        self.pipeline = TI2VidTwoStagesPipeline(
+            checkpoint_path=checkpoint_path,
+            distilled_lora=distilled_lora,
+            spatial_upsampler_path=upsampler_path,
+            gemma_root=gemma_root or "",
+            loras=tuple(loras or ()),
+            device=device,
+            quantization=quantization,
+        )
+
+    def _run_inference(
+        self,
+        prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        tiling_config: TilingConfigType,
+    ) -> tuple[torch.Tensor | Iterator[torch.Tensor], AudioOrNone]:
+        from ltx_pipelines.utils.args import ImageConditioningInput as _LtxImageInput
+        try:
+            from low_vram_runtime import get_streaming_prefetch_count
+
+            streaming_prefetch_count = get_streaming_prefetch_count()
+        except Exception:
+            streaming_prefetch_count = None
+
+        params = self._params
+        return self.pipeline(
+            prompt=prompt,
+            negative_prompt="",
+            seed=seed,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            frame_rate=frame_rate,
+            num_inference_steps=params.num_inference_steps,
+            video_guider_params=params.video_guider_params,
+            audio_guider_params=params.audio_guider_params,
+            images=[_LtxImageInput(img.path, img.frame_idx, img.strength) for img in images],
+            tiling_config=tiling_config,
+            streaming_prefetch_count=streaming_prefetch_count,
+        )
+
+    @torch.inference_mode()
+    def generate(
+        self,
+        prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        output_path: str,
+    ) -> None:
+        tiling_config = default_tiling_config()
+        video, audio = self._run_inference(
+            prompt=prompt,
+            seed=seed,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            frame_rate=frame_rate,
+            images=images,
+            tiling_config=tiling_config,
+        )
+        chunks = video_chunks_number(num_frames, tiling_config)
+        encode_video_output(
+            video=video,
+            audio=audio,
+            fps=int(frame_rate),
+            output_path=output_path,
+            video_chunks_number_value=chunks,
+        )

LTX2.3-1.0.4-new/patches/ltx_fp8_video_pipeline.py

@@ -0,0 +1,269 @@
+"""Fast pipeline wrapper for pre-quantized FP8 distilled checkpoints.
+
+The stock desktop wrapper uses ``QuantizationPolicy.fp8_cast()``, which is
+meant to cast BF16 checkpoints to FP8 while loading.  Pre-quantized FP8
+checkpoints already contain FP8 weights plus weight/input scales, so casting
+them again can produce valid-looking inference that decodes to black frames.
+
+This wrapper loads those checkpoints with the scaled-FP8 state-dict/module
+layout.  When TensorRT-LLM is not available, its FP8Linear forward falls back
+to a PyTorch dequantize-then-linear path, keeping FP8 storage while avoiding
+black output.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+import os
+from types import SimpleNamespace
+from typing import Any, Final, cast
+
+import torch
+from torch import nn
+
+from api_types import ImageConditioningInput
+from services.ltx_pipeline_common import (
+    default_tiling_config,
+    encode_video_output,
+    video_chunks_number,
+)
+from services.services_utils import AudioOrNone, TilingConfigType
+
+
+_FP8_FALLBACK_INSTALLED = False
+
+
+def _install_fp8linear_torch_fallback() -> None:
+    global _FP8_FALLBACK_INSTALLED
+    if _FP8_FALLBACK_INSTALLED:
+        return
+
+    from ltx_core.quantization.fp8_scaled_mm import FP8Linear
+
+    def _fallback_forward(self: Any, x: torch.Tensor) -> torch.Tensor:
+        weight_scale = self.weight_scale.to(dtype=x.dtype, device=x.device)
+        weight = (self.weight.to(dtype=x.dtype) * weight_scale).t().contiguous()
+        bias = self.bias
+        if bias is not None and bias.dtype != x.dtype:
+            bias = bias.to(dtype=x.dtype, device=x.device)
+        return torch.nn.functional.linear(x, weight, bias)
+
+    # TensorRT-LLM is not bundled in LTX Desktop, so use a deterministic PyTorch
+    # fallback for scaled FP8 checkpoints instead of silently ignoring scales.
+    FP8Linear.forward = _fallback_forward  # type: ignore[method-assign]
+    _FP8_FALLBACK_INSTALLED = True
+
+
+def _fp8_layer_names(checkpoint_path: str) -> frozenset[str]:
+    from safetensors import safe_open
+
+    names: set[str] = set()
+    with safe_open(checkpoint_path, framework="pt", device="cpu") as handle:
+        for key in handle.keys():
+            if not key.endswith(".weight_scale"):
+                continue
+            layer_name = key.removeprefix("model.diffusion_model.").removesuffix(
+                ".weight_scale"
+            )
+            if layer_name.startswith("transformer_blocks."):
+                names.add(layer_name)
+    return frozenset(names)
+
+
+def _scaled_fp8_quantization_policy(checkpoint_path: str) -> Any:
+    from ltx_core.loader.module_ops import ModuleOps
+    from ltx_core.loader.sd_ops import KeyValueOperationResult, SDOps
+    from ltx_core.model.transformer import LTXModel
+    from ltx_core.quantization.fp8_scaled_mm import FP8Linear
+
+    fp8_layers = _fp8_layer_names(checkpoint_path)
+
+    def transpose_fp8_weight(
+        key: str, value: torch.Tensor
+    ) -> list[KeyValueOperationResult]:
+        layer_name = key.removesuffix(".weight")
+        if layer_name in fp8_layers and value.dim() == 2:
+            return [KeyValueOperationResult(key, value.t())]
+        return [KeyValueOperationResult(key, value)]
+
+    def convert_fp8_layers(model: nn.Module) -> nn.Module:
+        if not isinstance(model, LTXModel):
+            return model
+        replacements: list[tuple[nn.Module, str, nn.Linear]] = []
+        for name, module in model.named_modules():
+            if name not in fp8_layers or not isinstance(module, nn.Linear):
+                continue
+            parent_name, attr_name = name.rsplit(".", 1)
+            replacements.append((model.get_submodule(parent_name), attr_name, module))
+        for parent, attr_name, linear in replacements:
+            setattr(
+                parent,
+                attr_name,
+                FP8Linear(
+                    in_features=linear.in_features,
+                    out_features=linear.out_features,
+                    bias=linear.bias is not None,
+                    device=linear.weight.device,
+                ),
+            )
+        return model
+
+    _install_fp8linear_torch_fallback()
+    return SimpleNamespace(
+        sd_ops=SDOps("fp8_selected_layers_transpose").with_kv_operation(
+            transpose_fp8_weight,
+            key_prefix="transformer_blocks.",
+            key_suffix=".weight",
+        ),
+        module_ops=(
+            ModuleOps(
+                name="fp8_prepare_selected_layers_for_loading",
+                matcher=lambda model: isinstance(model, LTXModel),
+                mutator=convert_fp8_layers,
+            ),
+        ),
+    )
+
+
+class LTXFp8VideoPipeline:
+    pipeline_kind: Final = "fast-fp8"
+
+    @staticmethod
+    def create(
+        checkpoint_path: str,
+        gemma_root: str | None,
+        upsampler_path: str,
+        device: torch.device,
+    ) -> "LTXFp8VideoPipeline":
+        return LTXFp8VideoPipeline(
+            checkpoint_path=checkpoint_path,
+            gemma_root=gemma_root,
+            upsampler_path=upsampler_path,
+            device=device,
+        )
+
+    def __init__(
+        self,
+        checkpoint_path: str,
+        gemma_root: str | None,
+        upsampler_path: str,
+        device: torch.device,
+        **_ignored: Any,
+    ) -> None:
+        from ltx_pipelines.distilled import DistilledPipeline
+
+        self._checkpoint_path = checkpoint_path
+        self._gemma_root = gemma_root
+        self._upsampler_path = upsampler_path
+        self._device = device
+        self._quantization = _scaled_fp8_quantization_policy(checkpoint_path)
+
+        self.pipeline = DistilledPipeline(
+            distilled_checkpoint_path=checkpoint_path,
+            gemma_root=cast(str, gemma_root),
+            spatial_upsampler_path=upsampler_path,
+            loras=[],
+            device=device,
+            quantization=self._quantization,
+        )
+
+    def _run_inference(
+        self,
+        prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        tiling_config: TilingConfigType,
+    ) -> tuple[torch.Tensor | Iterator[torch.Tensor], AudioOrNone]:
+        from ltx_pipelines.utils.args import ImageConditioningInput as _LtxImageInput
+
+        return self.pipeline(
+            prompt=prompt,
+            seed=seed,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            frame_rate=frame_rate,
+            images=[
+                _LtxImageInput(img.path, img.frame_idx, img.strength)
+                for img in images
+            ],
+            tiling_config=tiling_config,
+            streaming_prefetch_count=2,
+        )
+
+    @torch.inference_mode()
+    def generate(
+        self,
+        prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        output_path: str,
+    ) -> None:
+        tiling_config = default_tiling_config()
+        video, audio = self._run_inference(
+            prompt=prompt,
+            seed=seed,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            frame_rate=frame_rate,
+            images=images,
+            tiling_config=tiling_config,
+        )
+        chunks = video_chunks_number(num_frames, tiling_config)
+        encode_video_output(
+            video=video,
+            audio=audio,
+            fps=int(frame_rate),
+            output_path=output_path,
+            video_chunks_number_value=chunks,
+        )
+
+    @torch.inference_mode()
+    def warmup(self, output_path: str) -> None:
+        warmup_frames = 9
+        tiling_config = default_tiling_config()
+
+        try:
+            video, audio = self._run_inference(
+                prompt="test warmup",
+                seed=42,
+                height=256,
+                width=384,
+                num_frames=warmup_frames,
+                frame_rate=8,
+                images=[],
+                tiling_config=tiling_config,
+            )
+            chunks = video_chunks_number(warmup_frames, tiling_config)
+            encode_video_output(
+                video=video,
+                audio=audio,
+                fps=8,
+                output_path=output_path,
+                video_chunks_number_value=chunks,
+            )
+        finally:
+            if os.path.exists(output_path):
+                os.unlink(output_path)
+
+    def compile_transformer(self) -> None:
+        from ltx_pipelines.distilled import DistilledPipeline
+
+        self.pipeline = DistilledPipeline(
+            distilled_checkpoint_path=self._checkpoint_path,
+            gemma_root=cast(str, self._gemma_root),
+            spatial_upsampler_path=self._upsampler_path,
+            loras=[],
+            device=self._device,
+            quantization=self._quantization,
+            torch_compile=True,
+        )

LTX2.3-1.0.4-new/patches/runtime_policy.py

@@ -0,0 +1,21 @@
+"""Runtime policy decisions for forced API mode."""
+
+from __future__ import annotations
+
+
+def decide_force_api_generations(
+    system: str, cuda_available: bool, vram_gb: int | None
+) -> bool:
+    """Return whether API-only generation must be forced for this runtime."""
+    if system == "Darwin":
+        return True
+
+    if system in ("Windows", "Linux"):
+        if not cuda_available:
+            return True
+        if vram_gb is None:
+            return True
+        return vram_gb < 6
+
+    # Fail closed for non-target platforms unless explicitly relaxed.
+    return True

LTX2.3-1.0.4-new/patches/settings.json

@@ -0,0 +1,23 @@
+{
+  "use_torch_compile": false,
+  "load_on_startup": false,
+  "ltx_api_key": "",
+  "user_prefers_ltx_api_video_generations": false,
+  "fal_api_key": "",
+  "use_local_text_encoder": true,
+  "fast_model": {
+    "use_upscaler": true
+  },
+  "pro_model": {
+    "steps": 20,
+    "use_upscaler": true
+  },
+  "prompt_cache_size": 100,
+  "prompt_enhancer_enabled_t2v": true,
+  "prompt_enhancer_enabled_i2v": false,
+  "gemini_api_key": "",
+  "seed_locked": false,
+  "locked_seed": 42,
+  "models_dir": "",
+  "lora_dir": ""
+}

LTX2.3-1.0.4-new/patches/tts_worker.py

@@ -0,0 +1,222 @@
+"""Standalone TTS worker process for VoxCPM-based generation."""
+
+from __future__ import annotations
+
+# === [核心修复] 彻底封印 PyTorch 的所有动态编译机制 ===
+import os
+# 1. 禁用 Dynamo 编译器 (PyTorch 2.x)
+os.environ["TORCH_COMPILE_DISABLE"] = "1"
+# 2. 禁用 TorchScript JIT 编译器 (解决 nvrtc 报错)
+os.environ["PYTORCH_JIT"] = "0"
+# 3. 禁用底层算子融合器 NvFuser
+os.environ["NVFUSER_DISABLE"] = "1"
+
+import torch
+import torch._dynamo
+torch._dynamo.config.disable = True
+
+# 如果环境支持，强行在代码层关闭 nvfuser
+try:
+    if hasattr(torch._C, '_jit_set_nvfuser_enabled'):
+        torch._C._jit_set_nvfuser_enabled(False)
+except Exception:
+    pass
+# ==============================================================
+
+import argparse
+import json
+import tempfile
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+
+_MODEL_CACHE: dict[str, object] = {}
+
+
+
+
+def _to_1d_float32(audio: np.ndarray) -> np.ndarray:
+    arr = np.asarray(audio)
+    orig_dtype = arr.dtype
+
+    if arr.ndim == 0:
+        arr = arr.reshape(1)
+    elif arr.ndim == 2:
+        # Prefer channel-average while keeping the time axis.
+        if arr.shape[0] <= 8 and arr.shape[1] > arr.shape[0]:
+            arr = arr.mean(axis=0)
+        else:
+            arr = arr.mean(axis=1)
+    elif arr.ndim > 2:
+        arr = np.squeeze(arr)
+        if arr.ndim != 1:
+            arr = arr.reshape(-1)
+
+    if np.issubdtype(orig_dtype, np.integer):
+        scale = float(max(abs(np.iinfo(orig_dtype).min), np.iinfo(orig_dtype).max))
+        arr = arr.astype(np.float32) / max(scale, 1.0)
+    else:
+        arr = arr.astype(np.float32, copy=False)
+
+    arr = np.nan_to_num(arr, nan=0.0, posinf=0.0, neginf=0.0)
+    if arr.size == 0:
+        return np.zeros(1, dtype=np.float32)
+
+    # Remove obvious DC offset.
+    arr = arr - float(np.mean(arr))
+    return arr
+
+
+def _resample_linear(audio: np.ndarray, src_sr: int, dst_sr: int) -> np.ndarray:
+    if src_sr == dst_sr:
+        return audio
+    if audio.size <= 1:
+        return audio
+    dst_len = max(1, int(round(audio.size * float(dst_sr) / float(src_sr))))
+    x_old = np.arange(audio.size, dtype=np.float64)
+    x_new = np.linspace(0.0, float(audio.size - 1), dst_len, dtype=np.float64)
+    out = np.interp(x_new, x_old, audio.astype(np.float64))
+    return out.astype(np.float32)
+
+
+def _read_audio_any(path: str) -> tuple[np.ndarray, int]:
+    try:
+        data, sr = sf.read(path, always_2d=False)
+        return np.asarray(data), int(sr)
+    except Exception:
+        try:
+            import librosa
+        except Exception as exc:
+            raise RuntimeError(
+                "参考音频无法解码（建议上传 WAV，或安装 librosa 以支持更多格式）"
+            ) from exc
+        data, sr = librosa.load(path, sr=None, mono=False)
+        return np.asarray(data), int(sr)
+
+
+def _prepare_reference_audio(
+    path: str, out_dir: Path, target_sr: int, stem: str
+) -> str:
+    data, sr = _read_audio_any(path)
+    mono = _to_1d_float32(data)
+    mono = _resample_linear(mono, sr, target_sr)
+
+    peak = float(np.max(np.abs(mono))) if mono.size else 0.0
+    if peak > 0:
+        mono = mono / peak * 0.95
+
+    out_path = out_dir / f"{stem}.wav"
+    sf.write(str(out_path), mono, target_sr, subtype="PCM_16")
+    return str(out_path)
+
+
+def _normalize_generated_audio(wav: object) -> np.ndarray:
+    if hasattr(wav, "detach") and callable(getattr(wav, "detach")):
+        wav = wav.detach().cpu().numpy()
+    arr = _to_1d_float32(np.asarray(wav))
+
+    peak = float(np.max(np.abs(arr))) if arr.size else 0.0
+    if peak <= 1e-9:
+        return np.zeros(1, dtype=np.float32)
+
+    # Prevent clipping/noise if model output scale drifts.
+    if peak > 1.0:
+        arr = arr / peak
+    arr = np.clip(arr, -0.98, 0.98)
+    return arr.astype(np.float32)
+
+
+def _get_model(model_dir: str):
+    if model_dir not in _MODEL_CACHE:
+        from voxcpm import VoxCPM
+
+        _MODEL_CACHE[model_dir] = VoxCPM.from_pretrained(model_dir, load_denoiser=False)
+    return _MODEL_CACHE[model_dir]
+
+
+def run_generate(req: dict[str, object]) -> dict[str, object]:
+    text = str(req.get("text") or "").strip()
+    if not text:
+        raise RuntimeError("text 不能为空")
+
+    mode = str(req.get("mode") or "text_only").strip() or "text_only"
+    model_dir = str(req.get("model_dir") or "").strip()
+    output_dir = Path(str(req.get("output_dir") or ".")).resolve()
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    cfg_value = float(req.get("cfg_value") or 2.0)
+    inference_timesteps = int(req.get("inference_timesteps") or 10)
+
+    model = _get_model(model_dir)
+    sample_rate = int(getattr(getattr(model, "tts_model", None), "sample_rate", 24000))
+
+    ref_in = req.get("reference_wav_path")
+    prompt_in = req.get("prompt_wav_path")
+    prompt_text = str(req.get("prompt_text") or "")
+
+    temp_dir = Path(tempfile.mkdtemp(prefix="ltx_tts_"))
+    ref_ready = None
+    prompt_ready = None
+    try:
+        if isinstance(ref_in, str) and ref_in.strip():
+            ref_ready = _prepare_reference_audio(
+                ref_in.strip(), temp_dir, sample_rate, "reference"
+            )
+        if isinstance(prompt_in, str) and prompt_in.strip():
+            prompt_ready = _prepare_reference_audio(
+                prompt_in.strip(), temp_dir, sample_rate, "prompt"
+            )
+
+        if mode in {"clone", "ultimate_clone"} and not ref_ready:
+            raise RuntimeError("克隆模式必须提供参考音频")
+
+        gen_kwargs: dict[str, object] = {
+            "text": text,
+            "cfg_value": cfg_value,
+            "inference_timesteps": inference_timesteps,
+        }
+        if mode == "clone":
+            gen_kwargs["reference_wav_path"] = ref_ready
+        elif mode == "ultimate_clone":
+            gen_kwargs["reference_wav_path"] = ref_ready
+            if prompt_ready:
+                gen_kwargs["prompt_wav_path"] = prompt_ready
+            if prompt_text:
+                gen_kwargs["prompt_text"] = prompt_text
+
+        wav = model.generate(**gen_kwargs)
+        out = _normalize_generated_audio(wav)
+
+        import uuid
+
+        fname = f"tts_{uuid.uuid4().hex[:8]}.wav"
+        out_path = output_dir / fname
+        sf.write(str(out_path), out, sample_rate, subtype="PCM_16")
+        return {"status": "complete", "audio_path": fname, "sample_rate": sample_rate}
+    finally:
+        try:
+            for p in temp_dir.glob("*"):
+                try:
+                    p.unlink()
+                except Exception:
+                    pass
+            temp_dir.rmdir()
+        except Exception:
+            pass
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--request-json", required=True, help="Path to request json")
+    args = parser.parse_args()
+
+    req_path = Path(args.request_json)
+    req = json.loads(req_path.read_text(encoding="utf-8"))
+    result = run_generate(req)
+    print(json.dumps(result, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

LTX2.3-1.0.4-new/run.bat

@@ -0,0 +1,38 @@
+@echo off
+title LTX-2 Cinematic Workstation
+
+echo =========================================================
+echo    LTX-2 Cinematic UI Booting...
+echo =========================================================
+echo.
+
+set "LTX_PY=%USERPROFILE%\AppData\Local\LTXDesktop\python\python.exe"
+set "LTX_UI_URL=http://127.0.0.1:4000/"
+
+if exist "%LTX_PY%" (
+    echo [SUCCESS] LTX Bundled Python environment detected!
+    echo [INFO] Browser will open automatically when UI is ready...
+    start "" powershell -NoProfile -WindowStyle Hidden -Command "$ProgressPreference='SilentlyContinue'; $deadline=(Get-Date).AddSeconds(60); while((Get-Date) -lt $deadline){ try { Invoke-WebRequest -UseBasicParsing '%LTX_UI_URL%' -TimeoutSec 2 | Out-Null; Start-Process '%LTX_UI_URL%'; exit 0 } catch { Start-Sleep -Seconds 1 } }"
+    echo [INFO] Starting workspace natively...
+    echo ---------------------------------------------------------
+    "%LTX_PY%" main.py
+    pause
+    exit /b
+)
+
+python --version >nul 2>&1
+if %errorlevel% equ 0 (
+    echo [WARNING] LTX Bundled Python not found.
+    echo [INFO] Browser will open automatically when UI is ready...
+    start "" powershell -NoProfile -WindowStyle Hidden -Command "$ProgressPreference='SilentlyContinue'; $deadline=(Get-Date).AddSeconds(60); while((Get-Date) -lt $deadline){ try { Invoke-WebRequest -UseBasicParsing '%LTX_UI_URL%' -TimeoutSec 2 | Out-Null; Start-Process '%LTX_UI_URL%'; exit 0 } catch { Start-Sleep -Seconds 1 } }"
+    echo [INFO] Falling back to global Python environment...
+    echo ---------------------------------------------------------
+    python main.py
+    pause
+    exit /b
+)
+
+echo [ERROR] FATAL: No Python interpreter found on this system.
+echo [INFO] Please run install.bat to download and set up Python!
+echo.
+pause

LTX2.3-1.0.4-new/安装TTS环境.txt

@@ -0,0 +1,17 @@
+# 0. 进入 LTX Python 目录 (使用环境变量，自动适配任何用户的电脑)
+cd $env:LOCALAPPDATA\LTXDesktop\python
+
+# 1. 下载官方的 pip 安装脚本
+Invoke-WebRequest -Uri https://bootstrap.pypa.io/get-pip.py -OutFile get-pip.py
+
+# 2. 用 LTX 的 Python 运行这个脚本来修复/更新 pip
+.\python.exe get-pip.py
+
+# 3. 临时设置 MSVC 编译器的环境变量为 UTF-8 (以防又遇到刚才的 C++ 乱码)
+$env:CL="/utf-8"
+
+# 4. 指定用 LTX 的 Python 安装 editdistance
+.\python.exe -m pip install editdistance
+
+# 5. 【核心防御】指定用 LTX 的 Python 安装主包，并强制指定 GPU 镜像源防覆盖
+.\python.exe -m pip install voxcpm --extra-index-url https://download.pytorch.org/whl/cu118