Skip to main content
GameDev.net gamedev.net

World Models News

The latest World Models coverage curated for game developers.

Building Pretraining Data for World Models: An Unreal Engine-Based Pipeline for Action-Conditioned Video Generation

Researchers have built an Unreal Engine pipeline that generates action-conditioned video for world-model pretraining, pairing control inputs with rendered scene changes. The system produced 2,691 hours of 1080p footage and …

Research arXiv cs.GR · 1 week, 5 days ago
WorldRover: A Scalable Synthetic Video Data Engine for World Exploration with Rich Annotations

WorldRover turns long-horizon world exploration into a synthetic data problem, using Unreal Engine to render minute-scale traversals with full trajectories and scene geometry. The resulting 10M-sequence dataset pairs RGB with …

Research arXiv cs.GR · 4 weeks, 1 day ago
CG-World: A Large-Scale World-State Dataset and Protocol for World Models

A new CG-World dataset packages 850,000 temporally aligned 1-5 second segments from industrial graphics pipelines into a world-state format. It records latent state, motion, camera, lighting, physics, and events, giving …

Research arXiv cs.GR · 1 month, 2 weeks ago
Wonder: Video World Model Done Better

Wonder is a new video world model that turns a single image or conditioned clip into a playable scene you can explore by moving the camera in real time. The …

Research arXiv cs.GR · 1 month, 2 weeks ago
Echo-Memory: A Controlled Study of Memory in Action World Models

A controlled study of action-conditioned world models finds that memory design matters more than replay quality suggests. With a shared video diffusion backbone and fixed action interface, the authors show …

Research arXiv cs.GR · 3 months, 1 week ago
The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show

A study of video diffusion models found they can linearly expose physical plausibility in their internal states, even though that signal is missing from the VAE input. On IntPhys and …

Research arXiv cs.GR · 3 months, 1 week ago
Fewer, Better Frames: A Compute-Normalized Proof of Concept for Coherence-First World-Model Rendering with Model-Guided FSR4 Frame Generation

A proof-of-concept rendering pipeline that spends less compute on raw frame count and more on temporal coherence looks promising for graphics teams. Using a 15 FPS anchor cadence and FSR4-based …

Research arXiv cs.GR · 3 months, 1 week ago
Improved Baselines with Representation Autoencoders

A new paper claims RAEv2 cuts diffusion training time by more than 10x while improving image quality, which is relevant for teams building generative tools and world models. The key …

Research arXiv cs.GR · 3 months, 4 weeks ago