D33pStateTech/asian-ten-wan21-lora

Asian Ten — Wan 2.1 T2V LoRA

Custom LoRA for Wan 2.1 Text-to-Video 14B, trained with musubi-tuner on Modal A100-80GB.

Details

• Base model: Wan 2.1 T2V 14B (fp8)
• Training framework: musubi-tuner (kohya-ss)
• LoRA dim: 32
• Epochs: 20
• Resolution: 480×832
• Trigger: "Asian ten" or describe the subject in prompts

Usage

Load with any Wan 2.1 compatible inference tool (ComfyUI, musubi-tuner, Replicate, etc.) as a LoRA weight file.

# Example with musubi-tuner CLI
python wan_generate_video.py \
  --task t2v-14B \
  --dit wan2.1_t2v_14B_fp8_e4m3fn.safetensors \
  --vae wan_2.1_vae.safetensors \
  --t5 models_t5_umt5-xxl-enc-bf16.pth \
  --lora_weight asian_ten_wan21.safetensors \
  --prompt "Asian ten woman, portrait, studio lighting" \
  --video_length 49 --video_size 480 832

Checkpoints

File	Epoch	Notes
asian_ten_wan21.safetensors	20 (final)	Best quality