laion/Anim-E-Ghibli-Backgrounds-FLUX-9B

Vivarium Warm-Painted Location Backgrounds — FLUX 9B

This dataset contains illustrated locations for animation, games, storyboards and visual development. Scenes are composed as backgrounds: rooms and outdoor places are empty or unoccupied, with people and animals excluded. The warm, hand-painted Ghibli-inspired direction matches the companion character dataset.

Sample locations

These are examples from the published TAR archives, showing the range from everyday rooms and shops to forest and nightlife scenes.

Living room	Bookshop	Coastal forest	Converted warehouse club

Choose a collection

The default configuration combines the available families. Choose base for the original 1,093 scenes, city for homes, shops and services, drama for story-oriented settings, or nature for the new forests, caves, utility rooms, transitions, abandoned sites, labs, industry and rural locations. The current public snapshot contains 1,693/2,693 planned backgrounds: 1,093 base, 400 city, 200 drama and 0/1,000 nature-extension images.

Images are 1,376 × 768 RGB PNGs stored in TAR archives. Each archive sample includes its image, a JSON record with the prompt and generation metadata, and a text member with the exact prompt. This keeps thousands of assets organized without a flat directory of individual files.

from datasets import load_dataset

locations = load_dataset("laion/Anim-E-Ghibli-Backgrounds-FLUX-9B", "default", split="train", streaming=True)
sample = next(iter(locations))
print(sample["json"]["primary_class"], sample["json"]["description"])

For browsing, see the dataset map, the nature extension guide, and the city guide. Each includes class statistics, searchable CSVs and exact prompt plans. The images use undistilled FLUX.2 Klein Base 9B, the rank-64 background LoRA, 50 steps, real CFG 4 and BF16, without quantization or distillation.

Original Anim-E-derived collection details

Wide, empty illustrated locations in the user-selected ghibli-warm-painted direction. The planned release contains 1,093 images at 1,376 × 768, all generated with undistilled FLUX.2-klein-base-9 B, a rank 64 background LoRA, 50 steps, real CFG4, and BF16. There is no quantization, distilled checkpoint, reference-image conditioning, or denoising approximation cache.

This is an original synthetic image release derived from a small canonical inventory of Anim-E locations. Anim-E itself is a prompt dataset, not an image collection that was repackaged here.

Release status

Complete: 1,093 / 1,093 images generated and 1,093 uploaded.

Images are generated directly into WebDataset TAR containers, uploaded incrementally, and checked against remote SHA256 and byte size. There are no separately uploaded individual image files. Shard manifest, all completed inference records, progress, and measured performance provide the snapshot corresponding to this dataset-card update. While generation is active, newly uploaded shards can be newer than that snapshot; the final update reconciles everything.

Coverage and distribution

Component	Scene descriptions	Images per scene	Images
All canonical Anim-E locations	231	3	693
Systematic supplementary descriptions	400	1	400
Total	1,093 distinct production prompts	1	1,093

The 231 canonical scenes come from the documented Anim-E inventory. Repeated procedural organization, quality, lighting, mood, and design-style clauses were removed before canonicalization, and Sunny/Rainy/Snowy prefixes were collapsed. This reduces 1,262,142 source location prompt rows to 231 scene descriptions. The source contains 90 contemporary, 60 science-fiction, and 81 fantasy/historical canonical scenes: this release allocates 270,180, and 243 images to those respective source worlds. It does not preserve the source row frequencies; every canonical scene receives the same three-image allocation.

For each canonical location the three prompts request (1) a broad establishing view from the main approach, (2) an alternate side view with a gentle diagonal, and (3) a view back across the setting from the opposite side. Each uses a different stable seed. These are independent text-to-image interpretations; they are not geometrically registered photographs of a fixed 3 D environment. Room architecture and object placement can differ between the three images.

The 400 supplements were prepared by a gpt-6-luna planning agent, checked for unique IDs and descriptions, and allocated deliberately to recurring settings. They are 400 variants across 37 room/street/plaza types, not a claim of 400 additional unrelated physical locations. They overlap some canonical room types intentionally to improve useful coverage.

Supplemental primary class	Images
Residential interiors	100
Education spaces	88
Medical spaces	75
Retail spaces	37
Community and civic interiors	40
Street scenes	35
Squares and plazas	25

Residential coverage includes 30 living rooms, 25 kitchens, 15 adult bedrooms, 15 teen bedrooms, and 15 children’s bedrooms. Education adds 28 classrooms, 16 school corridors, 16 university campuses, 12 university hallways, 10 auditoria/lecture halls, and 6 seminar rooms. Medical settings distinguish 9 hospital patient rooms, 10 wards, 6 hospital lobbies, 12 operating rooms, 15 medical practices, 12 clinics/urgent-care settings, and 11 diagnostic/laboratory settings. There are 22 supermarkets. Street/plaza variants cover residential, commercial, historic, alley, waterfront, rural, market, neighborhood, civic, transit, and event-square settings.

The supplemental variation axes are architecture, layout, era/context, cultural setting, season, lighting, and scale. Most are contemporary everyday settings, with some historical buildings and preserved-era settings. These are editorial coverage choices, not demographic sampling weights or measured prevalence. Locations have no people, so clothing, ethnicity, age and gender are not person-image labels in this release. Adult/teen/children’s bedrooms refer to room function and furnishings.

The complete class list, CSV class list, machine-readable taxonomy, supplemental design rationale, and full prompt plan expose all allocations. Canonical topic names are the fifteen analytical groups from the Anim-E overview; supplemental primary classes are seven functional groups. They intentionally retain their source-family distinction rather than pretending to be an exhaustive universal ontology.

Empty-scene prompting and cleanup

Some source descriptions contain chefs, guests, students, medical staff, soldiers, robots, ghosts or creatures. Those active cast clauses are removed or replaced with physical setting details before generation. For example, the university lecture hall keeps wooden desks and blank blackboards while omitting students; the emergency department keeps clean vacant beds and medical equipment while omitting nurses and patients; fantasy battlefields show abandoned fortifications and debris rather than combatants. Human-shaped decorative objects are avoided in the supplementary descriptions. Text-prone props use blank boards, abstract interface panels, plain shop packaging and unlettered facades where relevant.

Both original versus cleaned canonical descriptions and original canonical source records are included. Original descriptions are provenance; the prompt field is what actually entered the image generator. A pilot review removed a generic packaging/wall-art instruction because it was causing irrelevant cartons and pictures in otherwise unrelated rooms.

The approved style is supplied by the shared template in tools/vivarium_prompt_templates.py:

vvar_bg, {description}. {style} Wide establishing view of an empty location. No people, no humans, no characters, no text, no watermark.

Here,{style} expands to:

Studio Ghibli-inspired hand-painted background art, rich warm watercolor-like shading, gentle luminous colors, atmospheric depth, delicate expressive architectural detail, a welcoming lived-in storybook atmosphere.

The real CFG negative prompt is:

people, human figures, characters, crowds, faces, readable text, lettering, logos, watermarks

This negative prompt supplements the positive template specifically to discourage figures and lettering. It is not a style-changing negative prompt. Generated images can still contain incidental shape or letter-like artifacts, and a prompt alone cannot guarantee every object is absent. The final quality report states what was inspected; do not interpret the intended empty-scene policy as an automated guarantee for every pixel.

Prompting examples

Use an empty environment description, then let the shared template add the style. Keep any boards or signs blank instead of describing text and asking the model to remove it later.

from tools.vivarium_prompt_templates import build_prompt

living_room = build_prompt("backgrounds", "An empty family living room with a low wooden table, a comfortable sofa, a brick fireplace, and tall windows looking into a leafy garden")
teen_room = build_prompt("backgrounds", "An empty teenager’s attic bedroom with a single bed, a study desk, a guitar on a stand, closed notebooks, and a broad dormer window overlooking tiled roofs")
lecture_hall = build_prompt("backgrounds", "An empty university lecture hall with tiered wooden benches, a vacant lectern, completely blank blackboards, and soft light entering tall side windows")
hospital = build_prompt("backgrounds", "An empty hospital patient room with a clean adjustable bed, a plain bedside cabinet, a privacy curtain, and softly lit abstract medical monitor displays")

The recommended appearance remains warm hand-painted animation scenery. Photorealistic/modern-realistic prompting, the 4 B model, and fast distilled variants were not selected for this release.

TAR structure and sample metadata

Each sample has three adjacent members with the same stable key:

base-021-v1.png
base-021-v1.json
base-021-v1.txt

Canonical IDs usebase-NNN-v1/v2/v3; supplementary IDs usebg-supp-NNNN. Each image is a full-resolution RGB PNG encoded directly from the generated PIL image into memory, then written to a TAR. No lossy image compression is applied. Containers use ordinary uncompressed TAR; PNGs already provide lossless compression. The first reviewed pilot shard on each worker targets 2 images; subsequent shards target 24 images, and final partial shards contain the remaining samples. Some shards contain fewer after rejected images were removed and regenerated under the same sample IDs. Shards may mix classes because a shared queue distributes work across eight GPUs.

JSON fields include stable ID, source family, scene description, full prompt, negative prompt, seed, class/subcategory, source world, variation axes, original source description, source revision, planning agent, image SHA256, UTC generation time, storage member names, and a completeinference block. Inference records identify the base and LoRA repositories and revisions, rank 64/alpha 64, selected training step 2000, LoRA strength 1.0,50 steps, CFG4,1376×768, BF16, sequence length 256, the scheduler class and configuration, real CFG batching, LoRA/QKV fusion, and regional torch compilation. Runtime records include GPU, index, batch size, library versions, model load time, CUDA-synchronized generation time, and whether first-run compilation was included.

The prompt and JSON are in the same container as the image, so samples remain usable when downloaded independently of the repository-level indexes. .txt is exactly the JSONprompt. Plan validation confirms 1,093 unique prompts and seeds and no text-encoder truncation.

Streaming and loading

Using Hugging Face Datasets:

from datasets import load_dataset

dataset = load_dataset("laion/Anim-E-Ghibli-Backgrounds-FLUX-9B", split="train", streaming=True)
sample = next(iter(dataset))
image = sample["png"]
metadata = sample["json"]
prompt = sample["txt"]
print(metadata["id"], image.size, prompt)

Onlydata/*.tar is selected by the dataset-card configuration; documentation/index CSV and JSON files are not treated as train examples. See the official WebDataset loading guide for WebDataset support.

For a lightweight reader that downloads no full dataset and extracts no individual images:

pip install pillow
python tools/read_webdataset.py train-g0-00000.tar --limit 2
# A local TAR or an explicit HTTPS URL works too.

Python WebDataset users can pass resolved shard URLs towebdataset.WebDataset(urls).decode("pil"). The three sample fields arepng,json, andtxt. Shard URLs can be built from metadata/shards.json; pin a dataset revision for a reproducible snapshot.

Reproducing inference

Base model: black-forest-labs/FLUX.2-klein-base-9 B, revision32773329fbe7e81a90ef971740e8ba4b0364ecf3.

Background adapter: laion/FLUX.2-klein-base-9 B-vivarium-loras, subfolderbackgrounds, revision5bd654514cc35a1c0d0de878c3a514019896a44e, SHA256253dd23291a64514f7cb107714f2c20118efef930781d620a83417b441198982.

Install the pinned dependencies, accept base-model access on Hugging Face, and provide your own token via the standard HF environment/login mechanism. The token used for this release is not included in any file.

pip install -r requirements.txt
python tools/infer_background_tar.py --id base-021-v1 --output reproduced.tar
python tools/infer_background_tar.py --description "An empty sunlit pottery workshop with wooden shelves, clay bowls, a vacant potter’s wheel and a window overlooking a garden" --seed 28008 --output pottery.tar

The helper reproduces the published prompt and seed or supplies the default warm-painted template for a custom description. Output is again a paired TAR, never a separately saved PNG. An 80 GB CUDA GPU is the measured production profile. --base-path permits an existing local model snapshot;--no-compile skips compilation but keeps 50 steps, CFG4, BF16 and the same adapter. Exact pixels can vary across CUDA/PyTorch versions and attention kernels even with the same seed. The LoRA repository also contains straightforward inference instructions.

Eight independent pipelines were used for aggregate throughput, with one image per GPU at a time. Each fuses the LoRA and QKV projections, caches the repeated negative text embedding, and compiles repeated transformer blocks. True conditional/unconditional CFG evaluations are batched without changing the guidance equation. No second GPU is consumed for an individual sample. Generation timings include positive text encoding and VAE decoding; loading, first-compile warmup, PNG compression, TAR writes and upload are separated in the reports.

Quality and limitations

296 distinct production sample IDs were visually reviewed in contact sheets as of this snapshot. Coverage includes 231 of 231 canonical scene categories. Nine targeted replacements removed prominent human posters/portraits from teen-bedroom, museum and emissary-suite scenes; their final images were inspected again. Small incidental lettering, decorative reliefs, anatomical diagrams and tiny illustrations can remain. Full-resolution decode, member pairing, dimensions, prompt matching and SHA256 checks are recorded separately in the final integrity report. Visual review is distinct from integrity verification and does not certify unreviewed images.

These are illustrative settings, not literal architectural plans or geographically verified depictions. Clinical equipment, scientific instruments, and historical details may be imaginative. The wide empty foreground is intended for later character compositing. Source-world labels and supplemental cultural contexts describe prompts and are not certifications of factual accuracy. Variants can resemble each other because they share a source scene and the same style LoRA.

Source and generation terms

The source prompt descriptions come from LAION Anim-E at revision0c939cef5a586292950724b6500f10ffc5bfa24f; source attribution and its Apache-2.0 notice are preserved in the provenance. The 9 B generation model is provided under the FLUX Non-Commercial License v 2.1. See terms and provenance and the complete model terms, including the provisions concerning outputs/generated data. This dataset contains generated images and metadata, not model weights. The Apache-2.0 notice of the source prompt dataset does not change the generation model’s terms.