README.md

metadata

language:
  - en
pipeline_tag: image-to-video
tags:
  - image-to-video
  - audio-conditioned
  - diffusion
  - talking-avatar
  - pytorch

AvatarForcing

One-Step Streaming Talking Avatars via Local-Future Sliding-Window Denoising

AvatarForcing is a one-step streaming diffusion framework for talking avatars. It generates video from one reference image + speech audio + (optional) text prompt, using local-future sliding-window denoising with heterogeneous noise levels and dual-anchor temporal forcing for long-form stability. For method details, see: https://arxiv.org/abs/2603.14331

This Hugging Face repo (lycui/AvatarForcing) provides two training-stage checkpoints:

• ode_audio_init.pt: stage-1 ODE initialization weights
• model.pt: stage-2 DMD weights

Model Download

Models	Download Link	Notes
Wan2.1-T2V-1.3B	🤗 Huggingface	Base model (student)
AvatarForcing	🤗 Huggingface	ode_audio_init.pt (ODE) + model.pt (DMD)
Wav2Vec	🤗 Huggingface	Audio encoder

Download models using huggingface-cli:

pip install "huggingface_hub[cli]"
mkdir -p pretrained_models

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir ./pretrained_models/Wan2.1-T2V-1.3B
huggingface-cli download facebook/wav2vec2-base-960h --local-dir ./pretrained_models/wav2vec2-base-960h
huggingface-cli download lycui/AvatarForcing --local-dir ./pretrained_models/AvatarForcing

Citation

@misc{cui2026avatarforcingonestepstreamingtalking,
      title={AvatarForcing: One-Step Streaming Talking Avatars via Local-Future Sliding-Window Denoising},
      author={Liyuan Cui and Wentao Hu and Wenyuan Zhang and Zesong Yang and Fan Shi and Xiaoqiang Liu},
      year={2026},
      eprint={2603.14331},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.14331},
}