yrlyrl/physjepa-moplan-checkpoints

PhysJEPA / MoPlan Stage-2 checkpoints (public backup)

Wan2.2-I2V-A14B LoRA (r32) + 27-channel motion adapter, both experts (high, low). Load with scripts/infer_stage2_cond_v2.py --high-checkpoint /high --low-checkpoint /low (config lvr_s2_cond_v2_len41.yaml).

• raft41h_s1r2_light300: FINAL model = Stage-1 v2 (yrlyrl/physjepa-stage1-hb-real22-step6000) + GT-SFT v2 @4000 + one light RAFT round (300 steps, p=0.5, winner = highest group reward)
• raft41h_s1r2_still2: same RAFT samples, 900 steps, p=0.8, stillness-reselected winners (ablation)
• gtsft2_s4000: GT-SFT v2 @4000 (41 frames @7 fps), the no-RL ablation Inference rule: Stage-1 conditions past its 4 s horizon hold the last known latent (hold_known_beyond_horizon.py).