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

TL;DR

GameVerse is a 1,655-hour, 69.7K-clip gameplay video dataset purpose-built for long-horizon interactive world modeling. 1,358 hours of its clips are continuous two-minute sequences — an order of magnitude longer than prior gameplay datasets, which typically cap under ten seconds. Each clip carries a temporally stable camera trajectory and event-boundary-aligned captions that factor static scene context from dynamic foreground events. The paper’s experimental claim is that training on GameVerse substantially extends the temporal horizon of interactive world models, locating the long-rollout-coherence failure mode in data availability rather than model capacity.

Key claims

• Current interactive world models struggle to maintain coherence beyond roughly one minute of rollout, and the limitation is attributed to scarcity of long-horizon training data rather than model capacity [Abstract].
• Existing gameplay video datasets typically consist of clips under ten seconds, with captions that describe whole clips in coarse undifferentiated text — insufficient for controllable long-horizon generation [Abstract].
• GameVerse contains 69.7K clips spanning 1,655 hours and 349.6 million frames across diverse first- and third-person gameplay, including 1,358 hours of continuous two-minute sequences — an order of magnitude longer than prior datasets [Abstract].
• Clips pass through a unified pipeline of scene-consistent segmentation, quality filtering, camera pose estimation, and event-based caption annotation [Abstract].
• Each clip is paired with a temporally stable camera trajectory that remains coherent across minutes [Abstract].
• Captions separate static scene context from dynamic foreground events and align to event boundaries rather than fixed time windows [Abstract].
• Training on GameVerse substantially extends the temporal horizon of interactive world models [Abstract].

Method

The paper constructs the dataset through a four-stage pipeline: (1) scene-consistent segmentation that splits raw gameplay into clips coherent across minutes rather than fixed windows; (2) quality filtering; (3) camera pose estimation producing per-clip temporally stable trajectories; and (4) event-based caption annotation that separates static scene context from dynamic foreground events and aligns captions to event boundaries. The resulting corpus covers first- and third-person gameplay, with the 1,358-hour subset deliberately held to continuous two-minute sequences rather than concatenated short clips.

Results

Headline statistics: 69.7K clips, 1,655 hours, 349.6M frames; 1,358 hours of continuous two-minute sequences (vs. <10s clip norm in prior gameplay corpora). The authors report that comprehensive analysis shows GameVerse surpasses existing public datasets in scale, diversity, and annotation quality, and that experiments demonstrate training on GameVerse substantially extends the temporal horizon of interactive world models [Abstract]. Specific per-benchmark deltas on named baselines are not surfaced in the abstract; downstream numbers will have to be refreshed from the full PDF.

Why it’s interesting

GameVerse is the sharpest filed data-side bet on the “minute-scale coherence is a data bottleneck” position. It sits at the pretraining-substrate layer where LongVie 2: Multimodal Controllable Ultra-Long Video World Model LongVie 2 reports 5-minute generation via training curriculum, Context Forcing: Consistent Autoregressive Video Generation with Long Context Context Forcing reports >20s effective context via KV-cache partitioning, and PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference PackForcing reports 2-minute extrapolation from 5s clip training via learned mid-buffer compression — GameVerse argues the architectural workarounds are compensating for a missing training distribution, and supplies the continuous-two-minute corpus the other recipes have been working around. Within the WFM cluster it’s complementary to gameplay-with-input-traces corpora announced in Meet Odyssey-3: Our Most Powerful Foundation World Model Odyssey-3 Pro’s data-composition disclosure, with the twist that GameVerse pins down camera trajectory + event-boundary captions specifically for controllable long-horizon generation rather than raw action conditioning.

See also

• World Foundation Models — the central cluster debate about whether long-rollout coherence is a model or data problem; GameVerse stakes out the data position
• Context Length / Quality Trade-off in Video Generation — the architectural-workaround side of the same tradeoff (Context Forcing / PackForcing / MALT / PFP)
• Synthetic Training Data — GameVerse as a purpose-built pretraining corpus for a specific downstream capability
• Autoregressive Video Generation — target architecture class that benefits from long-horizon training sequences
• WildWorld: A Large-Scale Dataset for Dynamic World Modeling with Actions and Explicit State toward Generative ARPG — Monster Hunter: Wilds action-conditioned corpus, sibling gameplay dataset but action-centric rather than camera + event-caption centric
• LongVie 2: Multimodal Controllable Ultra-Long Video World Model — 5-min generation via curriculum rather than data
• PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference — 2-min extrapolation from 5s clips via cache compression (the architectural counterpart)