StableAvatar: Infinite-Length Audio-Driven Avatar Video Generation,you can try it in ComfyUI

• 提示词可使用中文，务必详细描述场景既首帧人物的详细信息。
• 修复TCD和LCM因为全零pad引发的伪影，LCM因为融合机制的问题，批次过渡会有一定的闪烁，但好过之前，cfg不要设置为1。
• 设置GPU_memory_mode 为None 时，关闭mmgp的fp8量化/ when set GPU_memory_mode to None， will disabled mmgp ‘s FP8 quantize;
• when step=4 and lightX2V lora will run in LCM mode/ 步数为4时（搭配lightX2V lora）自动开启lcm；
• 可以使用lightX2V lora进行10步推理，节省2.5倍的时间/ use lightX2V lora infer in 10 steps.
• 同步官方最新平滑及对数滑动窗口机制代码；
• Infinite-Length Audio-Driven / 特点，无限长（音频多长就推理多长）
• 如果也使用了echomimic V3，会自动调用v3的共用模型

In the ./ComfyUI/custom_nodes directory, run the following:

git clone https://github.com/smthemex/ComfyUI_StableAvatar.git

audio-separator仅在推理歌曲时有用。

pip install audio-separator --no-deps # optional if need vocal 
pip install -r requirements.txt

3.1 fromFrancisRing/StableAvatar downlaod "Wan2.1_VAE.pth" ,"diffusion_pytorch_model.safetensors" and "config.json ","Kim_Vocal_2.onnx" ,"transformer3d-rec-vec.pt" or "transformer3d-square.pt " 底模有2个可选
3.2 use comfyui ,clipvison-h and umt5_xxl_fp8_e4m3fn_scaled.safetensors
3.3 wav2vec2-base-960h
3.4 if use echomimic v3，just only download"transformer3d-rec-vec.pt" or "transformer3d-square.pt " and "Kim_Vocal_2.onnx" / 如果也用echomimic v3，仅需下载底模和Kim_Vocal_2.onnx，会自动调用echomimic的模型
3.5 可选/optional lora kijai

├── ComfyUI/models/StableAvatar/transformer 
|         ├── diffusion_pytorch_model.safetensors  # Wan2.1-Fun-V1.1-1.3B-InP transformer #3.13G 务必注意模型同名。
|         ├── config.json
├── ComfyUI/models/StableAvatar/wav2vec2-base-960h
|         ├── all config json files 
|         ├──  model.safetensors
├── ComfyUI/models/clip
|         ├── umt5_xxl_fp8_e4m3fn_scaled.safetensors #comfy
├── ComfyUI/models/clip_vision
|         ├──clipvison-h # 1.26G comfy
├── ComfyUI/models/diffusion_models/
|         ├──transformer3d-rec-vec.pt  # FrancisRing/StableAvatar 二选一
|         ├──transformer3d-square.pt   # FrancisRing/StableAvatar
├── ComfyUI/models/vae
|         ├── Wan2.1_VAE.pth
├── ComfyUI/models/StableAvatar/  # 音频分离用
|         ├──Kim_Vocal_2.onnx
├── ComfyUI/models/loras/    
|         ├──lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors  #KJ

• 480x832 or 832x480 or 512x512 resolution /模型在这三种分辨率训练，推荐使用
• overlap_window_length 越大越好越慢/ high will get best quality but more times
• step 25~50 如果使用lightX2V 的lora step=10;
• 二种gpu卸载方式,推荐用mmgp
• duration>0 时裁切，裁切数值为秒/ if duration>0 will cut the input audio
• 推理歌曲时，须开启audio_separator分离人声 / when infer a song need turn on the audio_separator to get vocal.
• 暂时不要开启teacache，会花

@article{tu2025stableavatar,
  title={StableAvatar: Infinite-Length Audio-Driven Avatar Video Generation},
  author={Tu, Shuyuan and Pan, Yueming and Huang, Yinming and Han, Xintong and Xing, Zhen and Dai, Qi and Luo, Chong and Wu, Zuxuan and Jiang Yu-Gang},
  journal={arXiv preprint arXiv:2508.08248},
  year={2025}
}