Computer vision has remained at the task-specific model stage; GenCeption wants to move past it.

GenCeption, developed by Google DeepMind and top universities, repurposes a video generation model into a generalist computer vision model for ECCV 2026.

Researchers from Google DeepMind, along with academics from Toronto, Oxford, MIT, and Lund, present GenCeption, a work that takes a pre-trained video generation model and recycles it into a generalist perception model. The paper will be presented at ECCV 2026.

The thesis rests on an analogy. Language processing has transitioned from models tailored for a single task to generalist foundation models, driven by generative pre-training. Computer vision, however, has remained at the stage of the specialized model: one for depth, one for segmentation, one for pose. For the authors, the visual equivalent of generative pre-training already exists, and it is text-to-video generation: to produce a credible video, a model must internalize 3D geometry, object permanence, and physical laws, while remaining natively aligned with language.

GenCeption leverages this foundation. The same architecture, the same set of weights, and the same loss function serve all tasks, with the textual prompt alone deciding the output: depth, surface normals, camera pose, segmentation, referring expression segmentation, and 2D and 3D keypoints. The diffusion model, which is usually iterative, is converted into a single-pass model, significantly accelerating inference.

Two results stand out. First, data efficiency: trained on essentially synthetic videos, the model achieves, according to the authors, performance comparable to state-of-the-art specialized models with seven to five hundred times less training data. Second, emergent behaviors: trained solely on synthetic videos of human characters, it transfers without retraining to real-world sequences, to scenes containing multiple subjects, and to unseen categories such as animals or robots. The authors see this as evidence of a universal world model embedded within video generators, and an argument for viewing video generation not merely as a synthesis tool, but as a path toward generalist vision.