LongLive step-distillation LoRAs — block size × rollout ablation
Five few-step (4-step) distillation acceleration LoRAs (rank 128, alpha 128, 2000 steps), distilled from base Wan2.2-TI2V-5B. This ablation varies the chunk/block size and the student rollout structure (single bidirectional pass vs. autoregressive 8-frame blocks), to study their effect on long-generation temporal consistency.
| file | block size | student rollout |
|---|---|---|
block8_single.pt |
8 | single bidirectional (baseline) |
block16_single.pt |
16 | single bidirectional |
block16_AR8.pt |
16 | autoregressive (2×8 blocks) |
block24_single.pt |
24 | single bidirectional |
block24_AR8.pt |
24 | autoregressive (3×8 blocks) |
All share the same base, rank (128), step count (2000), and seed-paired training; the only variables are block size and rollout structure. Teacher + critic are bidirectional (no AR mask).
Format: each .pt holds PEFT-style lora_A/lora_B weights (under key generator_lora).
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for wileewang/longlive-stepdistill-blocksize-loras
Base model
Wan-AI/Wan2.2-TI2V-5B