byeongjun-park/FlashRender

This model was presented in FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow.

Please refer to the Github README for usage.

Checkpoints

Stage	File	Sampling
1	epoch_20.safetensors	50 steps, inference_mode=mul
2	meanflow_epoch_20.safetensors	4 steps, inference_mode=any
3	onpolicy_epoch_5.safetensors	4 steps, inference_mode=any

onpolicy_epoch_5.safetensors is the final model and the default in configs/base.yaml.

Each file stores only the trainable parameter subset, not a stand-alone diffusers/transformers checkpoint. Weights are bfloat16, except the relative-pose tokens and the per-block RoPE phase MLPs (rel_pose_{src,tgt}_token, self_attn.rope_phase_{qk,vo}), which are stored in float32. They are loaded with strict=False onto the base Wan DiT after it is patched by flashrender_utils.model_utils.adjust_to_FlashRender, so use the code repository:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="byeongjun-park/FlashRender",
    local_dir="models/checkpoints",
    allow_patterns=["*.safetensors", "config.json"],
)