GenCeption turns video generation into a general-purpose vision AI

Just as next-token prediction turned NLP into a generalist foundation, video generation may do the same for computer vision. In a new ECCV 2026 paper, a team led by Letian Wang shows that large-scale text-to-video generative models can be repurposed as a powerful pre-training paradigm for vision tasks. Their system, GenCeption, takes a frozen video diffusion backbone and uses it as a feed-forward perception model guided by text instructions. This setup provides rich spatiotemporal priors and vision-language alignment without requiring task-specific architecture changes.

GenCeption achieves state-of-the-art results across a diverse suite of tasks: depth estimation, surface normal prediction, camera pose estimation, expression-referring segmentation, and 3D keypoint detection—often matching or surpassing highly specialized models (e.g., DepthAnything3, SAM3, D4RT, VGGT-Omega). Critically, it does so with drastically less data—7x to 500x fewer training examples—while showing early signs of scaling with model size and data volume. The model also exhibits emergent generalization: trained solely on synthetic human videos, it works on real-world footage and out-of-distribution categories like animals and robots. These findings suggest text-to-video generation is not just a synthesis tool but a foundational step toward generalist vision AI for the physical world.