GAMESLASHPlayGen2 opens a playable AI world model that remembers 3D maps and off-screen entities
By GameSlash
Research released November 5 separates persistent 3D world state from neural rendering. A local Freedoom2 prototype tests new maps, entity interactions and programmable rules.
A paper posted on November 5, 2026 introduces PlayGen2, a research prototype that separates the 3D state of a playable world from the AI model that renders its images. Unlike frame-only world models, the system remembers geometry and tracks the player and other entities even when they are off-screen. Its authors provide a technical paper and public code and pretrained weights.
What changed compared with AI-generated gameplay video?
The system stores static geometry in a volumetric occupancy grid, alongside identity-indexed states for monsters, projectiles and items. An action-conditioned model updates that state and fills in newly explored regions. A separate neural renderer uses depth, semantic projections and visual history to draw a first-person view. Generated images do not feed back into world-state prediction, reducing dependence on visual memory alone.
The researchers trained on approximately 500 hours of gameplay collected from Freedoom2 MAP01. They then demonstrated navigation and entity motion on previously unseen maps. A model-controlled monster can keep moving when the player looks away: an example of why persistent game state matters beyond impressive-looking video.
Playable tests, speed and limits
The paper describes local execution on a MacBook Pro with M4 Pro. A detailed profiling setup with eight denoising steps reports about 4.98 generated frames per second when run serially, or 7.16 FPS with overlapped processing. Those are experimental backend measurements, not a claim of smooth 60 FPS or readiness to replace Unity or Godot.
The study also composes two single-player models to demonstrate synchronized participant states and adds explicit rules for health and death without retraining the models. This is a research example, not a multiplayer release for general players. The available demo is local and runs on a single shared session; its README says it is not an authenticated public hosting service.
Why game creators should care
PlayGen2 illustrates a useful design lesson: keep world geometry, entity state and game rules as inspectable structures, then treat AI rendering as a separate layer. For teams experimenting with world models, it provides a reference for avoiding disappearing corridors, off-screen character resets and visual-only simulations. The code repository includes setup instructions and a local browser interface, but game developers would need substantial engineering before using this approach in a shipped title.
Sources: PlayGen2 paper (November 5, 2026); official source and README. For a different approach to browser world exploration, see the existing GameSlash Frameboy report.