PROWBench checks whether generated videos faithfully realize program-specified entity trajectories, camera motion, and interactions, using replayable world states and timestamped events as the reference.
Takeaways
01Proxy conditioning improves entity control over camera conditioning alone, with strong video priors also improving event alignment.
02Proxy videos already encode camera motion, allowing camera control comparable to several models given explicit camera poses.
03General-purpose video models are competitive with specialized renderers on short clips across control and event fidelity.
04Fine-tuning offers mixed gains, while reliable domain-specific events still benefit from dedicated training.
05Better proxies need to bind identity, appearance, orientation, and articulation as well as position and motion.
06Long-horizon entity reappearance and identity consistency across independently generated views remain unresolved.
07Interaction Success and Logic-Render Alignment reveal failures visual and spatial scores miss, but judge calibration remains a limitation.
Leaderboard
Compare model results across five evaluation settings. Enable JavaScript to explore the table, or read the project paper.
Structural inputs
Coarse 3D, CWM proxy, and Colored-OBB are rendered from the same recorded world state and camera trajectory. The world record retains entity states and events even outside the camera’s field of view.
This gallery shows a subset of the benchmark results.
2 / 10 methods
Input
Input unavailable. Try another representation.
Press Play to load the selected videos.
About these comparisons
480p uses less data; 720p shows more detail. Lower-resolution originals keep their native size and are marked on the method card. For multi-view, the resolution refers to the complete four-view grid.
Selected gallery: 20 Verified-FF and 25 Unverified-FF examples on disjoint scenes, 6 multi-view examples × 4 cameras, and 6 long-horizon examples × 30 seconds. Input representations ↗
This gallery presents a selection of the supplied qualitative clips. The paper additionally evaluates 20 Gunfight cases and a fifth multi-view method, Seedance 2.5; those outputs are not available in this gallery. See the paper for quantitative results and evaluation protocols.
Scenes and interactions
40 Verified-FF appearance references
Crossing a streetCity drivingNavigating trafficDescending stepsPassing an objectPushing a street cartMoving a cratePushing a wooden boxSitting downNavigating a crowdCarrying upstairsCarrying a chairPushing a shopping cartClimbing stairsWalking on rocky terrainSliding on iceHelicopter flightRacing through a forestRacing on a wet trackExploring a caveExploring a fantasy forestExploring a desertDog on a snowy trailCat in a sunlit roomRabbit in an urban parkBirds over grasslandDuck on a rainy plazaDiving among fishRacing at duskPushing a crateSitting in a dark roomPushing a cart indoorsCats on a quiet streetCats in a city parkRabbits in the snowDucks on a park walkwayDucks in a neighborhoodDuck in the countrysideDucks on a snowy pathDiving over the seabed