Winnougan/Wan-Dancer-14B-INT8-INT4-Convrot

Wan Dancer 14B INT8/INT4

...

💜 Wan-Dancer ｜ 🖥️ GitHub | 🤗 Hugging Face | 🤖 ModelScope | 📑 Paper

Wan-Dancer: A Hierarchical Framework for Minute-scale Coherent Music-to-Dance Generation

🔥 Latest News!!

• July 13, 2026: 💃 We introduce Wan-Dancer, a method can generate long-duration, high-quality, rhythmic dance videos from music with global structure and temporal continuity. We released the model weights and inference code. And now you can try it on ModelScope Studio or HuggingFace Space!

---

📦 Quantized Weights (this repo)

This repo hosts INT8 and INT4 ConvRot-quantized versions of Wan-Dancer-14B's diffusion transformer and both text encoders, converted from the original BF16 release for lower VRAM footprint in ComfyUI. All quantization was done with ConvRot (rotation-based weight quantization, tensor-wise scaling) — weights are stored as INT8/INT4-range tensors alongside an FP32 per-tensor scale factor, decoded back to full precision at inference time.

A small number of layers whose input dimension isn't evenly divisible by the ConvRot group size (e.g. music_projection.weight) are kept at original BF16 precision automatically — quantization is skipped only for those layers, everything else is quantized.

Model files

File	Component	Quantization	Precision	Size	Original (BF16)
global_model-int8_convrot.safetensors	Diffusion transformer	INT8 ConvRot	INT8 weights + FP32 scale	16.30 GB	TBD
global_model-int4_convrot.safetensors	Diffusion transformer	INT4 ConvRot	INT4-range weights + FP32 scale	8.38 GB	TBD
models_t5_umt5-xxl-enc-int8_convrot.safetensors	Text encoder (T5-XXL)	INT8 ConvRot	INT8 weights + FP32 scale	6.27 GB	10.58 GB
models_t5_umt5-xxl-enc-int4_convrot.safetensors	Text encoder (T5-XXL)	INT4 ConvRot	INT4-range weights + FP32 scale	4.12 GB	10.58 GB

Original BF16 diffusion model size marked TBD — fill in once you have it, to make the compression ratio explicit for readers.

Which one should I use?

• INT8 — better fidelity, larger file. Good default if you have the VRAM to spare.
• INT4 — smallest footprint, some precision loss. Best for constrained VRAM setups; recommend testing output quality against INT8 or the original BF16 before committing to it for production work.

Usage

1. Download the diffusion model file into ComfyUI/models/diffusion_models/
2. Download the text encoder file into ComfyUI/models/text_encoders/
3. Load as you would the original BF16 Wan-Dancer-14B checkpoint — no workflow changes required, these are drop-in replacements at the weight level.

Notes

• Quantized with convert_to_quant (ctq) using --int8/--int4 --scaling_mode tensor --convrot.
• These are unofficial community-quantized weights, not produced or endorsed by the original Wan-Dancer-14B authors. Please refer to the original repo for the base model, license, and citation details.

---

Citation

If you use Wan-Dancer, please cite the original paper:

@article{wandancer2026,
  title={Wan-Dancer: A Hierarchical Framework for Minute-scale Coherent Music-to-Dance Generation},
  journal={arXiv preprint arXiv:2607.09581},
  year={2026}
}