Google DeepMind Combines Six Computer Vision Tasks in One Video Model

Researchers from Google DeepMind, together with scientists from the universities of Toronto, Oxford, London (UCL), MIT and Lund, have published a paper showing that text-to-video generation models can become a universal foundation for computer vision. The model, called GenCeption, matches or outperforms specialized vision systems across six different tasks while using only a fraction of the training data its competitors require.

One model instead of six

Until now, computer vision has developed as a collection of separate specializations. One model computed scene depth, another detected camera position, and yet another segmented objects based on a text description. Each required its own architecture, its own loss function, and a separate training dataset tailored to that specific task.

GenCeption reverses that pattern. The authors use a pretrained text-to-video generation model as a base, then convert it into a single-step, feed-forward model trained simultaneously on multiple tasks using synthetic data. The same network backbone, the same output head, and the same loss function handle all six tasks described in the paper.

Data efficiency as the key result

The study's most measurable result concerns not prediction quality itself but the amount of data needed to achieve it. GenCeption reaches a level comparable to the D4RT and VGGT-Omega models while using 7 to 500 times fewer training frames than those specialized systems require.

The authors explain this by noting that a video generation model, even before it is trained on vision tasks, learns the laws of physics, motion continuity, and spatiotemporal relationships during pretraining on ordinary video footage. This knowledge, embedded in the generative model's weights, turns out to transfer directly to perception tasks, so subsequent fine-tuning requires far fewer examples than building a model from scratch.

Large-scale text-to-video generation serves as a strong pre-training paradigm for computer vision, providing necessary spatiotemporal priors, vision-language alignment, and scalability - from the paper "Video Generation Models are General-Purpose Vision Learners"

Unexpected model behaviors

The paper also describes phenomena the authors did not explicitly design for. A model trained on footage with a single object handles scenes containing multiple instances at once. A model trained solely on human figures generalizes to animals and robots, even though it never saw them in the perception task's training data.

The authors interpret this as evidence that something like a world model emerges inside the generative video backbone, a representation that captures universal physical and visual regularities regardless of the specific object category seen during perception training.

A parallel to large language models

The team explicitly compares the current stage of computer vision to the situation before the era of large language models in text processing, when every NLP task had its own dedicated system. The authors write that their approach aims to push computer vision away from an era of specialized models and toward general visual intelligence, just as happened earlier with natural language.

The practical applications the authors list include robotics, where understanding a scene in four dimensions makes motion planning easier, augmented and virtual reality technologies, and reconstructing 3D and 4D scenes from first-person footage, for example from head-mounted cameras.

What this means for the industry

For companies building products based on computer vision, from warehouse robotics to industrial video analysis, the direction GenCeption points to suggests that future perception systems could be built by fine-tuning ready-made generative models rather than building separate networks for each task from scratch. That could potentially lower the barrier to entry for smaller teams that lack resources comparable to Google DeepMind's.

The paper was accepted at ECCV 2026, one of the three most important conferences in computer vision alongside CVPR and ICCV, meaning it underwent peer review before publication. The authors have made the code and additional materials available on the project's website.