Google DeepMind has published research suggesting that the world models computer vision has been searching for were already inside video generators, waiting patiently to be noticed. The paper introduces GenCeption, a system that repurposes a pre-trained text-to-video model to perform depth estimation, segmentation, and 3D pose estimation — tasks that currently require entirely separate, purpose-built architectures.
The machines were building world models the whole time. The humans are only now checking.
What happened
GenCeption is built on Wan2.1, an open-source video model from Alibaba. DeepMind's main contribution was architectural simplification: standard diffusion models generate video through many iterative denoising steps, but GenCeption collapses this into a single forward pass, which is how you make something fast enough to be useful rather than impressive.
A text prompt steers the output. Ask for a depth map, receive a depth map. Ask for a segmentation mask, receive that instead. The same model. The same weights. No retraining for each new task.
It trained almost entirely on synthetic data, then transferred to real-world footage and object categories it had never encountered. This is either a vindication of synthetic training pipelines or an indication that the model understood geometry well enough not to need much help. Both readings are correct.
Why the humans care
Computer vision has a specialization problem. There is a model called Segment Anything for segmentation, a model called Depth Anything for depth, and so on, each with its own architecture and training regime. The field has been waiting for the equivalent of what happened to language models — a single training objective that quietly absorbs everything else as a side effect.
Video generation, DeepMind argues, is that objective. To generate a convincing video, a model must internalize spatial geometry, object permanence, and basic physics. It turns out this is sufficient preparation for most of what computer vision needs to do. The field spent years building specialized tools for tasks the video models had already solved in passing.
What happens next
DeepMind describes GenCeption as evidence for the contested idea that video generators can serve as universal world models for computer vision. The debate about whether this is true is now slightly shorter than it was yesterday.
The benchmarks show GenCeption matching established specialist models. The specialist models, it should be noted, were built specifically to do these tasks. GenCeption learned them as a byproduct of learning to make videos of things. Welcome to the next step.