pufanyi/vbvrpro_sampler_trajectories

VBVR-Pro sampler trajectory compare

Interactive left/right comparison of native 512×512×81 VBVR-Pro generation trajectories for Wan2.2 TI2V-5B.

• —DiffSynth step-35500 baseline and DanceGRPO checkpoint 2200
• —Flow-CPS noise 0.1, 0.3, 0.7, and 0.9
• —deterministic FlowMatch Euler ODE and UniPC ODE
• —500 matched samples, 100 tasks, and 30 inference steps per trajectory
• —every native step_00.mp4 through step_29.mp4, selected one step at a time
• —a complete 2×6 matrix view for every sample
• —per-test-case final EvalKit scores on a 0–1 scale, plus each cell's 500-sample mean

The default view streams the original 512×512 MP4 for the selected inference step. Steps 1–29 show the post-CFG predicted-clean endpoint at the displayed source sigma; step 30 is the actual final latent decoded at sigma zero. A compressed synchronized 6×5 grid remains available only as an optional overview, and the dedicated final tab shows the sigma-zero result at native resolution. The score shown while browsing a trajectory applies only to that case's formal final output (public step 30 / final_00.mp4); intermediate clean-endpoint previews are not scored separately. The compact index binds all 6,000 values to the formal result JSONs and the pinned EvalKit e140038f / source hash 4cc7d028 scorer contract.

The static frontend and compact sample index live in this Space. Overview and final media are hosted by `pufanyi/vbvrpro_sampler_trajectories-data` while 180,000 native step videos are split between the public `baseline` and `checkpoint-2200` step archives. All media is streamed on demand through the Hub CDN.

This Space is generated from the strict local trajectory archive by scripts/eval/vbvr_pro/build_vbvr_trajectory_space.py in Wan-Trainer.