A world model can be capable of generating convincing gameplay and still lose coherence after about a minute. The bottleneck may be less about model size than about what the model has been shown: most gameplay datasets break action into short clips and coarse captions. GameVerse targets that data gap with sustained sequences and annotations designed to track both viewpoint and events over time.
Key Findings
- The dataset contains 69,678 clips, totaling 1,655 hours and 349.6 million frames; 40,752 clips reach two minutes. That continuity gives models training examples of how scenes and actions evolve, rather than only disconnected snapshots.
- Camera trajectories are estimated across clips, while captions distinguish stable scene context from changing events and agent actions. This supports conditioning on what the world is like as well as what happens next.
- Fine-tuning the same HY-World 1.5 backbone on GameVerse improves all six reported VBench metrics, including subject and background consistency. This suggests that long-horizon data can help without changing the model architecture.
Who It’s For—and the Trade-Off
A strong fit for researchers studying interactive world models, long-form video generation, or temporal consistency who need gameplay data with camera and event structure. Look elsewhere if the target is real-world video: the dataset is limited to gameplay, clips top out at two minutes, and its camera poses and generated captions can contain errors. The release provides clip metadata and annotations rather than redistributing the MP4 files.