Skip to main content

Google DeepMind's GenCeption: One Model to Rule Five Visual Tasks

Google DeepMind has introduced GenCeption, a new AI model that flips the script on traditional video generation. Instead of creating videos, it turns the process inside out to analyze visual data, handling five core tasks simultaneously: depth estimation, image segmentation, 3D pose estimation, camera motion tracking, and more. All this from a single model, in one forward pass.

Built on Alibaba's open-source Tongyi Wanxiang Wan2.1 framework, GenCeption takes a simple text prompt and outputs a rich set of information—depth maps, segmentation masks, camera trajectories, you name it. The key innovation? It's trained on just 7,500 synthetic videos of single people, rendered in Blender. Yet it generalizes remarkably well to real-world multi-person videos, animals, and even robots.

In tests, the smaller model processes 81 frames in about 6 seconds, while the 14-billion-parameter large model takes around 10 seconds. The detail restoration is so good it actually surpasses the original synthetic training data, clearly preserving edges like cat whiskers and individual strands of hair.

Image

Key Points

  • Single model, five tasks: Depth estimation, segmentation, 3D pose, camera motion, and more.
  • Trained on synthetic data: Only 7,500 single-person Blender videos, yet works on real-world multi-person scenes.
  • Fast processing: 81 frames in ~6 seconds (small model), ~10 seconds (large model).
  • Impressive detail: Preserves fine features like cat whiskers and hair strands.