Wan2.2-5B-NonAR-DMD-4Step-LoRA-r64-iter1600-SLT-AR-30s-Local32-Sink8-run2-iter2200
Weight-only LongLive-LoRA streaming-long-tuning checkpoint at iteration 2200.
- Base architecture: Wan2.2-TI2V-5B
- Student: causal/autoregressive
- Teacher and critic attention: non-causal
- Student attention: local window 32 with sink 8
- Video length: 184 latent frames (about 30 seconds at 24 FPS after decoding)
- EMA: disabled
- Contents: generator LoRA and critic LoRA weights
Optimizer state is intentionally excluded from this Hugging Face export. The local
training checkpoint remains the source of truth for resuming training.
