README.md

metadata

library_name: diffusers
pipeline_tag: image-to-video
license: other
tags:
  - video-generation
  - audio-driven-animation
  - avatar
  - wan
  - safetensors

HY-Video-Avatar2 checkpoints

This repository contains experimental WanModelAudio transformer checkpoints for HY-Video-Avatar2.

Checkpoints

Directory	Description inferred from run name
1229_fsdp_high_ct2_4800	High-noise CT2 checkpoint
0104_fsdp_low_ct2_600	Low-noise CT2 checkpoint
0115_fs3_emo_1_2600	Emotion-focused checkpoint
0120_fs10_motion_6_1200	Motion-focused checkpoint
0128_fs10_balanced_4_600	Balanced checkpoint
0129_fs3_balanced_1_1200	Balanced checkpoint
0202/ckpt/high_model	High-noise model checkpoint
0202/ckpt/low_model	Distilled low-noise model checkpoint

Each checkpoint directory contains:

• config.json: WanModelAudio architecture configuration.
• diffusion_pytorch_model.safetensors: transformer weights.

Loading

Use the custom WanModelAudio implementation from the HY-Video-Avatar2 codebase and point from_pretrained at one checkpoint directory, for example:

from infer.wan.modules.model_audio import WanModelAudio

model = WanModelAudio.from_pretrained(
    "bigfacing/HY-Video-Avatar2",
    subfolder="0129_fs3_balanced_1_1200",
)

These are transformer-only checkpoints; the remaining inference components and configuration must be supplied by the HY-Video-Avatar2 codebase.

Notes

• Checkpoint names are preserved from the original training runs for reproducibility.
• The descriptions above are inferred from those run names.
• Intended for research use; evaluate outputs, identity handling, consent, and applicable rights before deployment.