metadata
license: apache-2.0
tags:
- text-to-video
- lora
- diffusion
- adversarial-distillation
AFD OpenVid Student LR Sweep
Self-Forcing 1.3B student LoRA checkpoints trained for 300 optimizer steps with a frozen OpenVid-warmup VideoReward discriminator.
- Student learning rates:
5e-6,1e-5,5e-5 - Checkpoints: 60, 120, 180, 240, 300
- Generator LoRA: r=256, alpha=256
- Discriminator learning rate: 0 (frozen)
- Teacher data: 200 deterministic OpenVid-1M samples, seed 42, at least 81 frames and 3–8 seconds
Only the active default student adapter is included. The training-only
old adapter, discriminator, optimizer, and rank states are excluded.