byungjun-kim/DWM-Wan2.1-Fun-14b-LoRA

DWM WAN 2.1 14B Hand-Concat LoRA

This folder contains the DWM WAN hand-concat LoRA checkpoint packaged for release from:

• outputs/260330/wan2.1_14b_fun_inp_hand_concat_lora_batch_20260330_225432/checkpoint-5000

Files:

• lora_diffusion_pytorch_model.safetensors
• non_lora_weights.safetensors
• dwm_wan_14b_lora.yaml

This checkpoint is loaded with the DWM WAN inference path in this repository:

python training/wan/inference.py \
  --checkpoint_path hf_release/wan2.1_14b_fun_inp_hand_concat_lora_checkpoint-5000 \
  --experiment_config /path/to/this_repo/dwm_wan_14b_lora.yaml \
  --data_root data \
  --video_path /abs/path/to/data/realworld/videos/00001.mp4 \
  --prompt_subdir prompts_rewrite \
  --output_dir outputs_infer/wan_release_demo

The expected processed sample layout is:

<sample_root>/
├── videos/
│   └── <stem>.mp4
├── videos_static/
│   └── <stem>.mp4
├── videos_hands/
│   └── <stem>.mp4
└── prompts_rewrite/
    └── <stem>.txt

This release was validated in the DWM repository by loading the packaged folder directly and running single-video inference on realworld/videos/00001.mp4.