Google DeepMind: Video Generators Already Contain Computer Vision’s Missing World Model

GenCeption, built on Alibaba’s Wan2.1 video model, performs depth estimation, surface normal prediction, segmentation, and 3D pose estimation in a single forward pass via simplified architecture. It represents all outputs as standard RGB images and uses text prompts to specify tasks, adding trainable modules only for non-image outputs (e.g., 3D keypoints). Training uses 7,500 synthetic videos (800 digital human models × 200 action sequences rendered in Blender) plus a small amount of real video for language-guided segmentation. In benchmarks, GenCeption’s depth estimation matches DepthAnything 3, surface normal estimation surpasses NormalCrafter and Lotus-2, 3D pose recognition outperforms Genmo and TRAM, and language-guided segmentation matches Meta’s SAM 3 + Gemini 3.5 Flash. Training data volume is only 1/7 to 1/500 of models like D4RT. Notably, it generalizes to multi-person real videos and unseen animal categories despite training almost exclusively on single-person synthetic data. For developers, this suggests a future where one video generation model replaces many specialized models, drastically reducing data annotation and training costs. Speed remains a bottleneck: small models take ~6 seconds for 81 frames, 14B-parameter models ~10 seconds, and joint training degrades 3D keypoint accuracy. Researchers emphasize keeping the original model architecture intact.