Google DeepMind just dropped a paper arguing that the universal training task computer vision has been missing was hiding in plain sight: video generation. Their new model, GenCeption, repurposes Alibaba’s open-source Wan2.1 video generator to perform depth estimation, surface normal mapping, segmentation, and 3D pose estimation — all through a single architecture using nothing more than text prompts and a single forward pass. No iterative denoising. No task-specific heads for the core vision work.
The thing that makes you sit up is the data efficiency. While models like D4RT and VGGT Omega trained on millions of videos, GenCeption matched or beat them using a synthetic dataset of just 7,500 clips. That’s 7 to 500 times less data. The team rendered 800 digital humans performing 200 motion-captured sequences in Blender, varied the backgrounds and camera angles, and called it a day. Real video only entered the picture for language-guided segmentation. On benchmarks, GenCeption goes toe-to-toe with DepthAnything 3 for depth, beats NormalCrafter and Lotus-2 on surface normals, and outperforms Genmo and TRAM on 3D pose. For complex segmentation, it matches Meta’s SAM 3 paired with Gemini 3.5 Flash.
What’s genuinely surprising is the generalization. A model trained almost exclusively on synthetic single-person footage somehow produces crisp depth maps for real-world scenes with multiple people, cats, and humanoid robots. The paper notes that some outputs show finer detail than the Blender ground truth — whiskers and individual hair strands survive the process. The researchers attribute this to the generative pretraining itself, not just data volume. They say learning to predict the next frame forces a model to internalize spatial geometry, object permanence, and basic physics. It’s the vision equivalent of how next-token prediction gave LLMs grammar and world knowledge for free.
There’s a catch, though. Joint training across all tasks degraded 3D keypoint estimation, likely because the extra trainable modules needed for that task created interference. The team had to treat keypoint prediction differently — it doesn’t produce an RGB image, so it needed special components. Still, the core argument stands. If a single video generator can be repurposed across this many vision tasks with minimal data, the field’s obsession with specialized architectures may be a detour, not the destination.
