Google DeepMind tests video generation as a shortcut for computer vision

Google DeepMind researchers have introduced GenCeption, a computer vision model that turns a pretrained text-to-video generator into a general-purpose perception system for tasks including depth estimation, surface normals, segmentation and 3D pose. The work is relevant for AI teams because it suggests video generators may be useful beyond content creation, though the researchers did not announce a product release or disclose deployment plans.

GenCeption is built on Alibaba’s open-source Wan2.1 video model. Rather than running the usual diffusion process across many denoising steps, the system produces task outputs in a single forward pass. A text prompt specifies the job, such as generating a depth map or segmentation mask, and the model returns the result from the same underlying architecture.

The design keeps most outputs in a common format: three-channel RGB images. Depth, surface normals, segmentation masks and even camera motion are represented as image-like outputs. For tasks that do not naturally fit that format, including 3D keypoints, the researchers added trainable components. Training uses a shared loss across tasks, with preprocessing handling task-specific requirements rather than separate model designs.

Small training set, broad task coverage

The training setup is the main claim. According to the paper, most of GenCeption’s training data came from 7,500 synthetic videos. The team created them from 800 digital human models and 200 motion-capture sequences, rendered in Blender with varied backgrounds and camera positions. Real video was used only for language-guided segmentation.

Google DeepMind says GenCeption matched or exceeded specialist systems on several benchmarks while using one model across tasks. The paper reports depth results comparable to DepthAnything 3, stronger surface-normal estimation than NormalCrafter and Lotus-2, better 3D pose results than Genmo and TRAM, and language-guided segmentation performance matching Meta’s SAM 3 combined with Gemini 3.5 Flash.

The researchers also compare GenCeption’s data requirements with models such as D4RT and VGGT Omega, which they say were trained on millions of videos. Their claim is that GenCeption reached similar performance with 7 to 500 times less data. In same-condition tests, they report that pretraining on video generation outperformed representation-learning approaches including V-JEPA and VideoMAE V2.

Generalization claims and limits

The paper says the model generalized beyond its narrow synthetic training distribution. Although it was trained mostly on single-person synthetic videos, GenCeption worked on real footage with multiple people and on categories it had not seen during training, including animals and humanoid robots. The researchers report that some outputs retained fine visual details, such as whiskers and hair boundaries, despite the synthetic training data being less detailed.

There are limits. Joint training across tasks reduced performance on 3D keypoint estimation. The researchers suggest the added modules required for that task may interfere with representations learned during video pretraining. Runtime is also not yet trivial: the smaller version takes about six seconds to process an 81-frame video, while the larger 14-billion-parameter model takes about ten seconds.

The broader argument is contested. The DeepMind authors say video generators may contain useful “world model” representations because generating plausible video requires spatial and motion information. Other researchers have pushed back on classifying text-to-video systems as world models because they lack real-world feedback. Yann LeCun, formerly Meta’s chief AI scientist, has argued against generative video as a route to robust world modeling, while Meta’s V-JEPA 2 follows a different approach based on predicting abstract representations rather than pixels.

GenCeption does not prove that video generators understand physics. It uses a video model as a feature extractor for bounded perception tasks, and the reported results suggest that may be a practical reuse case. For computer vision teams, the signal is narrower but concrete: pretrained video generators could become competitive backbones for perception workloads that have been served by separate specialist models.

This story draws on original reporting from The Decoder.