GenCeption

GenCeption is a research project from Google DeepMind, described in the paper "Video Generation Models are General-Purpose Vision Learners" (ECCV 2026). It repurposes a pre-trained text-to-video diffusion model as a unified vision model. The model takes an input video and a text prompt that specifies the desired output, and it produces that output in a single forward pass instead of a multi-step diffusion process.

The model handles both dense and sparse perception tasks. Dense tasks, such as depth and normals, are unified in RGB space with supervision applied in latent space. Sparse tasks, such as keypoints, are handled by adding learnable tokens to the diffusion transformer. The project page lists grounded 4D reconstruction with camera pose, 4D human keypoints, and text-referred segmentation. Creator coverage also mentions DensePose. Post-training uses mostly synthetic data.

The authors report results competitive with or better than task-specific models such as DepthAnything3, D4RT, VGGT-Ω, SAM3 and Sapiens. They also report that the video-generative backbone beats V-JEPA and VideoMAE V2 under matched fine-tuning data, and that it reaches D4RT/VGGT-Ω-level performance with 7× to 500× less training data. The model transfers from synthetic to real footage and to unseen object categories. The provided material shows a project page and citation only. It does not show released code, weights, a demo or a paper link.