NeurIPS 2026 · Evaluations & DatasetsDATA FOR WORLDS THAT KEEP GOING

GameVerse

A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling

Continuous gameplay, camera trajectories, and structured semantics for worlds that keep going.

Kang He · Wenshuo Peng · Chuanhao Li · Zihui Gao · Xiaojie Xu · Zhengyuan Lin · Yuzhe Ding · Donghong Ji · Kaipeng Zhang · Yongtao Ge

69,678video clips
1,655gameplay hours
349.6Mvideo frames
40,752two-minute clips
DATASET RECORDING

Step into the recorded worlds.

Browse real retained gameplay from GameVerse.

DATASET RECORDING
WORLD / 05

Elden Ring

Explore

Real retained footage · Short preview · Original creators credited below

A longer view of interactive worlds.

World models need more than glimpses. GameVerse preserves continuous gameplay up to two minutes across first- and third-person perspectives.

Minute-scale continuityEstimated camera geometryStructured scene semantics

One world.
Two uninterrupted minutes.

Follow one continuous Red Dead Redemption 2 recording through a sunlit forest. Select a frame to jump to that moment.

ONE CONTINUOUS RECORDING / 02:00

Red Dead Redemption 2 · Forest walk

Original source ↗
A SINGLE UNBROKEN SCENESELECT A MOMENT ↓

Five frames from the same retained 120-second clip · The final card shows its last available frame.

Clip duration distribution

Horizontal axis: clips · Vertical axis: video duration (seconds).

THE TWO-MINUTE WINDOW82.1%

of gameplay hours come from two-minute clips.

0s60s120s
How these statistics are measured

Watch the scene.
Read its layers.

Follow a Red Dead Redemption 2 city walk. Scrub through time to read the scene, player, world, and camera descriptions aligned to each interval.

Captions follow video intervals. Camera centers below follow the 30 fps pose frames sampled from this same recording.

CAMERA GEOMETRY

A path through the world.

Play or scrub the city-walk video above to follow its frame-linked camera path. Switch projection to view the same pose frames from another plane.

This case: original 60 fps video → every second frame sampled for 30 fps poses · Web preview: 15 fps

CAMERA PATH / FRAME-ALIGNED0:00 / 0:46
Estimated camera-center trajectoryCamera centers sampled at 30 fps from the original 60 fps recording. The position marker follows the corresponding pose frame during video playback.

From recorded worlds
to generated ones.

Revisit the scenarios shown in our paper through full-resolution screenshots from the original camera-conditioned generation videos.

Screenshots from the original camera-conditioned evaluation videos.

Text prompt

Explore the project.

Code and dataset access in one place. Public dataset access and the paper link will be added with the release.

Cite GameVerse.

Use this BibTeX entry when referencing the dataset. Proceedings details will be updated when available.

BIBTEX · NEURIPS 2026
@inproceedings{he2026gameverse,
  title = {GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling},
  author = {He, Kang and Peng, Wenshuo and Li, Chuanhao and Gao, Zihui and Xu, Xiaojie and Lin, Zhengyuan and Ding, Yuzhe and Ji, Donghong and Zhang, Kaipeng and Ge, Yongtao},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026}
}
DATASET RECORDING

Annotation excerpt

Original source ↗

Short excerpts of retained recordings. Original clip length and preview length are listed separately. Captions describe the stated source interval.