Google Deepmind turns video generators into universal world models
Google Deepmind's GenCeption repurposes video models for depth estimation and segmentation, matching specialized models with 90% less data and suggesting video generators are the world models vision needs.
google deepmind researchers released GenCeption, a system that repurposes text-to-video generators to solve dense prediction tasks in computer vision without architectural changes for each specific job. by converting outputs like depth maps or segmentation masks into standard RGB images and using text prompts to switch tasks, the model processes inputs in a single fast forward pass. benchmark results show GenCeption matches established specialized models like Segment Anything despite training on a dataset of only 7,500 synthetic videos. the findings suggest that the physical and spatial understanding required for realistic video generation provides the foundation for general-purpose world models in vision.