When the camera returns, do the models remember?
Each model gets a short observed context, then must generate video along a trajectory that leaves and returns to previously-seen viewpoints. At every return frame we have the real GT frame. Below: the GT observed viewpoint vs each model’s render on successive revisits — the key test is whether the return frames stay faithful to GT and consistent across visits.
FrameCrafter and TrajectoryCrafter reconstruct the first revisit reasonably, then drift/hallucinate on later revisits (they carry no persistent memory of the observed scene) — visible below as the return frames morphing away from GT. This is the informative baseline the benchmark is built to expose, and the target the finetuned models aim to beat.
| method | n | GT-fidelity PSNR (up) | SSIM (up) | LPIPS (dn) | cross-visit PSNR (up) | drift LPIPS (dn) | met3r (dn) |
|---|---|---|---|---|---|---|---|
| FrameCrafter | 20 | 18.48 | 0.51 | 0.39 | 16.51 | 0.47 | 0.22 |
| TrajectoryCrafter | 18 | 13.12 | 0.37 | 0.73 | 14.12 | 0.61 | 0.14 |
| Ours VACE-14B | 8 | 22.14 | 0.62 | 0.29 | 29.04 | 0.13 | 0.19 |
| method | n | return PSNR (up) | SSIM (up) | LPIPS (dn) | degrade PSNR (dn) |
|---|---|---|---|---|---|
| FrameCrafter | 20 | 9.13 | 0.17 | 0.78 | 34.60 |
| TrajectoryCrafter | 17 | 13.19 | 0.29 | 0.69 | 30.66 |
GT-fidelity = render at a revisit/return vs the REAL observed frame (the key number); cross-visit = consistency across the 3 revisits.
Row 1 = GT observed viewpoint + context. Then one row per method: GT, then its render at revisit 1/2/3 of the same anchor pose.
Our finetuned Wan I2V-14B (camera-conditioned) and Wan VACE-14B (warp+inpaint) are training. This page gains an “Ours (2k / 4k)” row per method as their checkpoints are evaluated on the same conds.