PermaVid · full-GT revisit benchmark

Does a video model remember the scene when the camera returns?

A comprehensive evaluation of camera-controlled / world-model video generation. Every model gets an observed context + the real camera trajectory and must generate the rest; because the scene is fully rendered, every generated frame is scored against a real GT frame (whole-trajectory + multi_visit + moving_back).

🏆
Comprehensive comparison
Video-first: every example plays all methods vs real GT in sync, then a same-camera-pose frame comparison at the revisit. Plus ranked table (multi_visit + moving_back) + degradation curve. Start here.
📉
Multi-visit degradation table
One viewpoint, revisited up to 8× — every model's render at visit 1, 2, 3, … as a frame table, GT on top, with a Δ(first→last) column. Watch who forgets the scene as the camera keeps coming back.
🔬
Ranked revisit matrix
All 440 revisit frames pulled out — every method's reconstruction at the exact returning frames, in a grid sorted best→worst by revisit PSNR, GT on top. Read down a column to see who holds the viewpoint.
🎚️
Span 0.33 revisit comparison
The slow / small-motion setting. Each scene is video-first (all methods vs real GT in sync), then a same-camera-pose frame comparison at the revisit — every model at the exact returning pose, ranked by PSNR.
🎬
The dataset
78 UE scenes × ~20 diverse camera trajectories, with stats and per-scene detail.
🌐
WorldScore metrics
Separate ICCV'25 benchmark on its own camera tasks — overall WorldScore + 7 dimensions (camera control, 3D/photometric consistency, subjective quality…) for 9 methods.
🔁
Per-scene results
GT-vs-render montages for each held-out scene + metrics table.
📉
Revisit degradation
Per-visit fidelity curves + playable clips showing who drifts on the return.