DiffusionART · HunyuanVideo

HunyuanVideo-13B fine-tuned with DiffusionART for text-to-video generation.

This repository contains the full transformer weights, not a LoRA or a complete pipeline. The inference entry below downloads the base model's VAE and text encoders automatically.

Inference

Use the included infer.py. Expand the setup once, then run the example from the same directory.

Environment setup

Linux, Python 3.12, a CUDA 12.1-compatible driver and toolkit (nvcc), and approximately 100 GiB of available GPU memory and disk space for the verified default path. The installer uses a pinned public DanceGRPO inference runtime.

export RUN_DIR="${RUN_DIR:-$PWD/diffusionart-t2v-demo}"
export HF_HOME="$RUN_DIR/cache/hf"
export PIP_CACHE_DIR="$RUN_DIR/pip-cache"
export TMPDIR="$RUN_DIR/tmp"
export XDG_CACHE_HOME="$RUN_DIR/xdg-cache"
mkdir -p "$RUN_DIR" "$HF_HOME" "$PIP_CACHE_DIR" "$TMPDIR" "$XDG_CACHE_HOME"
cd "$RUN_DIR"
python3.12 -m venv bootstrap
bootstrap/bin/python -m pip install 'huggingface-hub==0.26.1'
bootstrap/bin/huggingface-cli login

bootstrap/bin/python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download(
    "vanzl/DiffusionART-HunyuanVideo",
    revision="b9f83d8fc677be078cf20e75e2149a5f69059220",
    allow_patterns=["infer.py", "install.sh", "requirements-inference.txt"],
    local_dir="release",
)
PY
bash release/install.sh "$RUN_DIR/venv" "$RUN_DIR/runtime"

The fixed revision preserves the released weights and inference entry. Hugging Face login is used by that entry even though this model is public.

venv/bin/python - <<'PY'
import runpy
import sys
import torch

# Use the attention backend from the successful inference checks.
torch.backends.cuda.enable_cudnn_sdp(False)
sys.argv = [
    "infer.py",
    "--revision", "b9f83d8fc677be078cf20e75e2149a5f69059220",
    "--prompt", "A white horse gallops through shallow water along a sunlit beach.",
    "--output", "diffusionart.mp4",
    "--cache-dir", "cache",
]
runpy.run_path("release/infer.py", run_name="__main__")
PY

Defaults: 640×640 · 53 frames · 8 FPS · 30 sampling steps. The script uses the verified shift, guidance, precision and VAE settings. Change --prompt and --output for another video; an existing output is never overwritten.

Examples

Selected outputs from this model. See the Demo for the full gallery and pretrained comparisons.

A white horse gallops through shallow water along a sunlit beach.	The camera approaches a glass pavilion surrounded by autumn trees.
The camera follows a winding river through a valley of autumn trees.	Sunlight sparkles through a rotating crystal prism on a white surface.

License

Subject to the Tencent Hunyuan Community License and Notice. Training provenance and weight hashes are recorded in release.json.

Citation

@misc{wan2026diffusionartefficientstablescalable,
  title={DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models},
  author={Zhenglin Wan and Jiayu Chen and Zhengfeiyang Zhang and Pengfei Zhou and Wangbo Zhao and Fangyuan Kong and Pengfei Wan and Kun Gai and Yang You},
  year={2026},
  eprint={2610.12572},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.12572},
}