T2V β GRPO checkpoints (Wan2.1 + SANA)
Unified folder for our T2V GRPO ablation checkpoints.
Layout
wan21/
wan-base/ # Wan2.1-T2V-1.3B base (diffusers layout)
kl-0.004-step-200/ # transformer.pt + training_state.json
kl-0.007-step-200/
kl-0.01-step-200/
sana/
kl-0.004-step-200/ # SANA-Video GRPO, motion_score=30
kl-0.007-step-200/
kl-0.01-step-200/ # job acloud-65026330dfde
kl-0.01-step-200-747dd8535868/ # second KL=0.01 run (acloud-747dd8535868)
Wan KL runs were previously published at ttthhz1/wan21-t2v-grpo-kl-ablation and are mirrored here under wan21/.