DeepMind Says Video Models Are Already the Missing World Models for Computer Vision

A team at Google DeepMind argues that generative video models like Veo 3 already function as the world models that computer vision has lacked for years. The claim is a direct response to critics in the research community who, in April 2026, formally excluded text-to-video generators from the definition of a true world model.

What Veo 3 Demonstrates

According to DeepMind researchers, Veo 3 handles tasks from entirely different areas of computer vision without any fine-tuning at all. It detects edges, performs image segmentation, and upscales footage, despite never having been trained specifically for those tasks.

A second area is physical understanding. The model simulates buoyancy, air resistance, and light reflections, and in one demonstrated example it removes blocks from a Jenga tower in a physically plausible way while keeping the overall structure coherent.

Reasoning Frame by Frame

The most surprising part of the demonstration concerns visual reasoning. Veo 3 solves mazes, recognizes symmetries, sorts numbers, and completes sudoku puzzles. Researchers call this phenomenon 'chain-of-frames': each successive video frame becomes a step in the reasoning process, much like consecutive sentences in the chain-of-thought process used by large language models.

The model can also manipulate images: removing backgrounds, changing the color of objects, and generating new shots of the same scene, in some cases preserving more detail than specialized photo-editing tools.

We believe video models will become universal, general-purpose foundation models for machine vision, just as large language models became foundation models for natural language processing - Google DeepMind researchers

The Dispute Over Defining a World Model

DeepMind's claim lands right in the middle of a dispute that has been running for months. When OpenAI called Sora a 'world simulator,' and DeepMind chief Demis Hassabis voiced similar views about Veo, part of the research community saw it as an abuse of terminology. In April 2026, the OpenWorldLib team published a paper explicitly excluding text-to-video generators from the definition of a world model.

Their argument is simple: a model that only generates video from text does not perceive its surroundings or interact with them. A true world model must close a feedback loop with its environment, something passive content generation does not provide. Yann LeCun, Meta's head of AI, has joined this critique, consistently calling the generative approach to world models a dead end and advocating for predictive architectures as the alternative.

Why It Matters

The dispute is not just an academic argument over definitions. If DeepMind's thesis holds up, a single universal video model could replace dozens of specialized computer vision systems currently used in robotics, surveillance, medical imaging, and industry, much as large language models displaced narrowly specialized text-processing tools.

For companies building products around computer vision, this points to a potential shift in strategy: instead of training a separate model for every task, they could try using a general video model and simply prompt it appropriately. DeepMind is developing the Genie family in parallel, including Genie 3, which generates interactive three-dimensional worlds that stay coherent for several minutes at 720p and 24 frames per second, a different approach from passive text-to-video generation and closer to the definition of a world model pushed by the critics.

The dispute is unlikely to be settled quickly. Both sides agree on one thing: video models are rapidly gaining capabilities that nobody expected of them even two years ago. Whether we end up calling them world models or simply very capable content generators will largely determine how companies and researchers plan their coming years of investment in computer vision.