GameVerse supplies two-minute gameplay clips for world models
GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling · arXiv:2610.37257
In plain words
GameVerse is a collection of gameplay videos made for training systems that generate interactive scenes. Many gameplay clips are only a few seconds long, so GameVerse keeps clips of up to two minutes and adds estimated camera paths and captions that distinguish the setting, events and actions.
In one test, the same video-generation model scored higher on every reported measure after training on GameVerse. That result covers 60-second videos from 103 inputs, not every model or kind of scene.
The research
Interactive scenes that remain coherent for about a minute need training footage that lasts long enough to show how scenes evolve.
GameVerse adds continuous gameplay clips and descriptions of both surroundings and changing actions, shifting the question from whether such data exists to whether training on it changes measured generation quality in a 60-second test.
GameVerse is a gameplay video dataset for interactive world models, pairing long scene-consistent clips with camera paths and captions that separate setting, events and actions.
The construction pipeline segments long gameplay recordings at scene changes, filters visual and motion problems, estimates camera paths, and pairs each clip with captions separating setting, events and actions. The finished dataset includes more than 40.6 thousand clips at its 120-second limit.
For 60-second videos from the same 103 inputs, HY-World-FT, trained further on GameVerse, scored higher than HY-World 1.5 on all six VBench measures and all seven WorldScore components. This supports improvement in that comparison, while leaving the contribution of each dataset element unresolved.
The generation experiments cover gameplay videos up to 60 seconds, while GameVerse clips extend to two minutes.
The study does not test transfer to real-world scenes or separate the effects of clip length, camera supervision and structured captions; the exact starting checkpoint for the HY-World comparison is unspecified.
What the evidence shows
- A dataset for longer scenes
GameVerse is a gameplay video dataset for training interactive world models, which generate changing scenes in response to actions and camera movement. The paper frames a mismatch: systems discussed in its introduction tend to lose scene coherence after about a minute, while their training clips are mostly shorter than ten seconds.
GameVerse supplies clips as long as two minutes, giving training data more continuous gameplay at the scale of the stated problem.
- Filtering footage and camera paths
The dataset is built from long gameplay recordings, segmented at scene changes into clips of up to 120 seconds. A sequence of quality checks filters visual defects, unstable motion and duplicates.
The pipeline also estimates camera positions and orientations for each frame, then removes trajectories with jumps, jitter or drift. This turns raw recordings into clips with both continuous footage and camera-motion information; the camera paths are estimated from video rather than taken from game-engine parameters.
- Captions separate scene details
By separating stable setting from events and actions, the captions make different kinds of scene information explicit for training. More than 40.6 thousand clips reach the dataset’s 120-second segmentation limit, showing that two-minute footage is a substantial part of the collection.
- The same model scores higher
The same-backbone comparison asks whether further training on GameVerse changes scores under matched conditions.
On the same 103 inputs, the further-trained model scored higher across both video-quality and control and scene-quality measures in 60-second generation, although its exact starting checkpoint is unspecified.
- Model differences bound the claim
Because the original models differ in backbone and frame rate, their higher aesthetic and imaging scores cannot isolate which difference caused the gap. This short-video diagnostic therefore cautions against attributing every cross-model score gap in the 60-second test to GameVerse training.
Deep
- Original paper authors
- Kang He, Wenshuo Peng, Chuanhao Li, Zihui Gao, Xiaojie Xu
- Submitted to arXiv
- License
- arXiv non-exclusive