Google Deepmind researchers claim video generation models already contain the world models computer vision has been missing.
The team built GenCeption, a system that takes a pre-trained video generator and uses it for classic computer vision tasks. It performs depth estimation, segmentation, and 3D pose estimation while needing very little training data.
This approach contrasts with the field of computer vision, where specialists dominate. Models like “Segment Anything” handle one job, while “Depth Anything” handles another. Each uses its own architecture and requires specific training.
Deepmind argues that large text-to-video models can bridge this gap. Generating realistic videos requires understanding the spatial geometry of a scene, how objects move, and basic physics. These models also use text descriptions during training, which links language to visual content. They can train on vast datasets because labeling the data is relatively cheap.
GenCeption repurposes a video generator for vision tasks
GenCeption builds on Alibaba’s open source Wan2.1 video model. The main change is a simpler architecture. Diffusion models typically generate video from noise through many small steps. GenCeption instead produces a prediction in one forward pass, making it fast enough for practical computer vision tasks.
The researchers use a simple method to make one model handle many tasks. GenCeption represents every output as a standard three channel RGB image, whether the result is a depth map, surface normal map, or segmentation mask. It also converts camera movement into an image representation.
A text prompt tells the model which task to perform, much like an instruction given to a chatbot. The team adds trainable modules for tasks such as 3D keypoint prediction, which don’t produce images.
Training uses one loss function across all tasks. Rather than changing the architecture for each task, the researchers process the training data to account for task specific requirements.
Matching specialized models with a fraction of the data
Most of the training data came from a synthetic dataset containing just 7,500 videos. The team combined 800 digital human models with 200 motion sequences from a motion capture dataset. It then rendered the results in Blender with different backgrounds and camera angles. Real videos were used only for language guided segmentation.
According to the study, GenCeption matches or beats state-of-the-art results across many benchmarks despite using one architecture for every task. Its depth estimates match those from specialized models such as DepthAnything 3. GenCeption beats NormalCrafter and Lotus-2 on surface normal estimation. It also outperforms Genmo and TRAM on 3D pose recognition. On complex language guided segmentation, it matches Meta’s SAM 3 combined with Gemini 3.5 Flash.
Models such as D4RT and VGGT Omega trained on millions of videos. GenCeption reaches similar results using 7 to 500 times less data. When tested under the same conditions, pretraining on video generation also beats methods such as V-JEPA and VideoMAE V2. The authors attribute those results to the generation task rather than data volume alone. They argue that video generation helps models learn useful representations of space and movement.
Strong generalization beyond training data, but with clear limits
GenCeption trained almost entirely on synthetic videos showing one person at a time, but it still works on real videos with several people in the frame, as well as unrelated categories such as animals and humanoid robots. According to the study, some outputs contain more detail than the Blender renderings used for training. The results can preserve a cat’s whiskers and the edges of individual strands of hair.
Joint training across all tasks does hurt 3D keypoint estimation. The researchers suspect that the extra components needed for this task interfere with mechanisms learned by the base model during pretraining.
Their conclusion is that as few changes as possible should be made to the original model’s architecture. Processing speed also needs work. The smaller model takes about six seconds to process a video with 81 frames, the larger model has 14 billion parameters and takes about ten seconds.
Video generators may contain useful world models without understanding physics
The authors reject the idea that video generators are only entertainment tools. They argue that these models already contain a type of universal “world model” that could support a foundation model for computer vision. Such a system could fill the same role in image processing that large language models fill in text.
Whether that framing holds up is debated. An international research team recently proposed a uniform definition with OpenWorldLib and explicitly excluded text-to-video models because they lack feedback from the real world.
Former Meta chief AI scientist Yann LeCun has gone further, arguing that generative video models are a dead end. Meta’s V-JEPA 2 follows the alternative he has championed, predicting abstract concepts rather than pixels. A benchmark from Tsinghua University points to the limits of pixel prediction. Sora 2, Seedance 2.0, and Veo 3.1 repeatedly failed tests of basic physics and logic even when their output looked convincing.
GenCeption uses video generation for a narrower purpose. The researchers don’t ask Wan2.1 to predict how the world will behave. They use it to extract features for specific tasks such as depth estimation and segmentation, where those learned features can outperform specialized systems. Google Deepmind is also exploring a different approach with Genie 3, which builds interactive 3D environments for training AI agents.
