Google DeepMind researchers have made a counterintuitive argument: the video generation models built for creating realistic footage already contain the foundational "world models" that computer vision has been searching for. A new model called GenCeption demonstrates this by repurposing a pre-trained video generator to handle classic computer vision tasks like depth estimation, surface normal mapping, and 3D pose recognition, achieving results that match or exceed specialized systems designed for each individual task.

Why Does This Matter for Computer Vision?

Computer vision has long relied on specialized models, each built from scratch for a specific job. "Segment Anything" handles image segmentation, "Depth Anything" estimates how far objects are from the camera, and other models tackle pose estimation or surface analysis. This fragmented approach contrasts sharply with how large language models (LLMs) work. LLMs are general-purpose systems trained on a single task, predicting the next word, yet they develop broad capabilities across writing, reasoning, and problem-solving as a side effect.

Computer vision lacks an equivalent foundation. GenCeption suggests video generation could fill that gap. Creating realistic videos requires understanding spatial geometry, how objects move through space, and basic physics principles. These models also learn from text descriptions during training, linking language to visual content. Because labeling video data is relatively inexpensive compared to other approaches, video generators can train on massive datasets.

How Does GenCeption Actually Work?

GenCeption builds on Alibaba's open-source Wan2.1 video model but simplifies its architecture significantly. Traditional diffusion models generate video by starting with noise and refining it through many small steps. GenCeption instead produces predictions in a single forward pass, making it fast enough for practical computer vision applications.

The model uses an elegant unified approach: every output, whether a depth map, surface normal map, or segmentation mask, is represented as a standard three-channel RGB image. Camera movement is also converted into an image representation. A text prompt tells the model which task to perform, similar to how you might instruct a chatbot. For tasks like 3D keypoint prediction that don't naturally produce images, the researchers add trainable modules.

The Surprising Data Efficiency

Perhaps the most striking finding is how little training data GenCeption needs. Most of the training came from just 7,500 synthetic videos created by combining 800 digital human models with 200 motion capture sequences, then rendering them in Blender with different backgrounds and camera angles. Real videos were used only for language-guided segmentation tasks.

Despite this minimal dataset, GenCeption matches or beats state-of-the-art results across multiple benchmarks. Its depth estimates rival specialized models like DepthAnything 3. It outperforms NormalCrafter and Lotus-2 on surface normal estimation and beats Genmo and TRAM on 3D pose recognition. On complex language-guided segmentation, it matches Meta's SAM 3 combined with Gemini 3.5 Flash.

To put this in perspective, competing models like D4RT and VGGT Omega trained on millions of videos. GenCeption reaches similar results using 7 to 500 times less data. When tested under identical conditions, pretraining on video generation also outperformed methods like V-JEPA and VideoMAE V2. The researchers attribute these results to the generation task itself rather than data volume alone, arguing that video generation helps models learn useful representations of space and movement.

Steps to Understanding GenCeption's Unified Architecture

• Single Model, Multiple Tasks: Rather than maintaining separate architectures for depth estimation, segmentation, and pose estimation, GenCeption uses one base model for all tasks, reducing complexity and training overhead.
• Text-Guided Task Selection: A text prompt specifies which computer vision task to perform, allowing the same model to switch between different outputs without architectural changes.
• Unified Output Representation: All outputs are converted to standard RGB image format, enabling the model to process diverse tasks through a consistent pipeline regardless of their natural output format.
• Minimal Synthetic Training Data: The model trains primarily on 7,500 synthetic videos rather than millions of real videos, demonstrating that video generation pretraining provides sufficient foundational knowledge for downstream vision tasks.

Can It Generalize Beyond Its Training?

GenCeption trained almost entirely on synthetic videos showing one person at a time, yet it still works on real videos with multiple people in the frame and even unrelated categories like animals and humanoid robots. Some outputs contain more detail than the Blender renderings used for training, preserving fine details like a cat's whiskers and individual strands of hair.

However, the approach has clear limitations. Joint training across all tasks does hurt 3D keypoint estimation, with researchers suspecting that the extra components needed for this task interfere with mechanisms learned by the base model during pretraining. Processing speed also needs improvement; the smaller model takes about six seconds to process a video with 81 frames, while the larger 14-billion-parameter model takes about ten seconds.

Is This Really a Universal World Model?

The Google DeepMind team argues that video generators already contain a type of universal "world model" that could support a foundation model for computer vision, filling the same role in image processing that large language models fill in text. However, this framing remains contested in the research community.

An international research team recently proposed a uniform definition of world models 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 an alternative approach, predicting abstract concepts rather than pixels. A benchmark from Tsinghua University points to the limits of pixel prediction, showing that 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, more pragmatic purpose. Rather than asking the video model to predict how the world will behave, researchers use it to extract features for specific tasks like 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.

The research suggests that video generators may be more useful for computer vision than previously thought, even if they don't fully solve the world model problem. By demonstrating that a single repurposed video model can match specialized systems while using a fraction of the training data, GenCeption opens a new perspective on how foundation models might develop across different domains of AI.