Wuji Hand Gesture VAM TI2V-5B 30L

Final checkpoint from W&B run: https://wandb.ai/wuji-tech/wuji_hand_gesture/runs/w468mpbb

Training summary:

• Backbone: Wan2.2-TI2V-5B
• Action DiT: 30 layers, dim 1024, 24 heads
• Dataset: knightnemo/wuji-hand-gestures-cropped
• Resolution: 256x256
• Raw frames: 49
• Target video frames: 13 (action_video_freq_ratio=4)
• Action horizon: 48
• Action/proprio dim: 20
• Full reference length: 65 frames
• Final step: 10000

Files:

• step-10000.safetensors: final training checkpoint
• log_node0.txt: rank-0 training log

Base model