Winnougan/MiniMax-H3-PTD-LoRAs-Quantized-ComfyUI

MiniMax-H3 PTD Ref2VA LoRAs: Pruned for ComfyUI

Smaller, rank-reduced versions of the MiniMax-H3 PTD Ref2VA few-step accelerator LoRAs, made to be lighter to download and load in ComfyUI.

All the real work belongs to the original authors. These are derivative files. Original repo and full documentation: taobao-content-ai/MiniMax-H3-PTD-LoRAs

Same license and terms as the original apply. Check the original repo.

---

What are these?

PTD (Parametric Trajectory Distillation) turns MiniMax-H3's Ref2VA task (reference image → video + audio) into a few-step generator.

Instead of one shared acceleration LoRA, PTD trains one LoRA per denoising step ("per-hop"):

• 4-step file: 4 hops, 4 transformer evaluations
• 8-step file: 8 hops, 8 transformer evaluations

Each step uses its own adapter plus its own output head.

What I changed

The originals are rank 128 and big. I used per-hop SVD rank reduction to cut each hop's LoRA down to rank 64 or 32.

• Every hop and every module is truncated on its own, so hops are never mixed
• lora_alpha is set equal to the new rank, so alpha/rank stays 1.0, same as the originals
• proj_out and audio_proj_out (the per-step output heads) are untouched
• Still BF16. This is rank pruning, not quantization.
• The metadata (ptd_num_hops, pdd_num_steps, lora_rank, lora_alpha, etc.) is kept and updated

Files

File	Hops	Rank	Approx. size	Energy retained (mean / min)
MiniMax-H3-Ref2VA-PTD-4Step_r64.safetensors	4	64	~2.6 GB	0.989 / 0.921
MiniMax-H3-Ref2VA-PTD-4Step_r32.safetensors	4	32	~1.3 GB	0.975 / 0.814
MiniMax-H3-Ref2VA-PTD-8Step_r64.safetensors	8	64	5.2 GB	0.985 / 0.865
MiniMax-H3-Ref2VA-PTD-8Step_r32.safetensors	8	32	~2.6 GB	0.965 / 0.724

Originals are rank 128: the 8-step is 10.35 GB.

Energy retained is how much of each LoRA's weight signal survives the rank cut. The mean is across all hop-modules, and the min is the worst single module. It's a rough guide, not a quality score.

Which one should I pick?

• Want the safest option? Use a r64 file.
• Short on disk or VRAM? Try r32, and compare against r64 before trusting it.
• 8-step r32 has the weakest worst-case module (min 0.72), so test it first.
• Few-step distilled models are touchy. If you see drift in motion, lip sync or identity, step up to r64 or use the original.

How to use (ComfyUI)

1. Drop the .safetensors into ComfyUI/models/loras/
2. Install the ComfyUI nodes from the original repo (comfyui_node/ComfyUI-MiniMaxH3-PTD/), plus ComfyUI-PTD, which they depend on
3. Use the original, unpruned Ref2VA transformer as the base. Pruned, Turbo, PDD and quantized base models are not valid for these adapters.
4. Load the LoRA with the PTD loader node at weight 1.0, and set the step count to match the hop count (4 or 8)

Full setup and the example workflow are in the original repo's ComfyUI README.

Notes

• The nodes read rank and alpha from the file metadata. If a loader hardcodes rank 128, the strength will be wrong.
• Pruning is lossy. The original files are the reference.
• These pruned files are community-made and not endorsed by the original authors.

Credits

• Original LoRAs: taobao-content-ai/MiniMax-H3-PTD-LoRAs
• PTD: Lan Feng et al., Parametric Trajectory Distillation for Few-Step Video Generation, arXiv:2610.11498 (2026). Project page
• PDD: Neta Shaul et al., Parallel Decoding Distillation for Fast Image and Video Generation, arXiv:2607.26004 (2026)
• Base model: MiniMaxAI/MiniMax-H3
• Pruning: Winnougan