Google DeepMind and five university partners built a single feed-forward network that performs depth estimation, camera pose recovery, segmentation, and human keypoint tracking without a separate model for each task. The system, called GenCeption, is described in a paper titled “Video Generation Models are General-Purpose Vision Learners,” accepted at the European Conference on Computer Vision. It targets an assumption that has structured computer vision for over a decade: that every perception task requires its own specialized architecture and training run.
The team, which includes researchers from the University of Toronto, University College London, Oxford, MIT, and Lund University, starts from a pretrained text-to-video diffusion model rather than building a vision network from scratch. That base model already encodes spatial and temporal patterns from generating video, plus alignment between text and pixels from following text prompts. GenCeption adapts those learned representations into a single architecture fine-tuned across multiple perception tasks at once, collapsing what used to be a multi-step generative process into one forward pass.
The instruction interface is text. A user specifies which task they want, depth, surface normals, camera pose, foreground segmentation, or 3D keypoints, and the same network produces it. The paper also demonstrates referring segmentation, where GenCeption isolates an object described in natural language by reasoning about its color, position, and motion, including objects like a rocket that never appeared in its training data.
On accuracy, the researchers report GenCeption is competitive with or beats specialized state-of-the-art models built for individual tasks, naming DepthAnything3, D4RT, VGGT-Omega, SAM3, Sapiens, DAVID, Genmo, and Lotus-2 as the comparison points. The paper does not disclose whether any of those baselines were re-tuned specifically for this comparison, so the margins should be read as reported rather than independently reproduced.
The more consequential claim is about data efficiency. Under matched fine-tuning data, the video-pretrained backbone outperforms alternative pretraining approaches including V-JEPA and VideoMAE V2, and the paper states it reaches accuracy comparable to D4RT and VGGT-Omega using seven to five hundred times less training data. If that efficiency holds up under outside replication, it changes the cost calculus for teams that currently maintain separate labeled datasets and separate models for depth, pose, and segmentation pipelines.
The model was fine-tuned predominantly on synthetic video, largely of humans, which makes the reported generalization results the paper’s most interesting evidence. GenCeption trained only on synthetic footage transfers to real-world video without additional training. Trained on scenes with a single subject, it generalizes to real video containing multiple people. Trained only on humans, it extends to categories it never saw during fine-tuning, including animals and robots. The authors frame this as evidence that a video generation backbone builds something closer to a general world model than a narrow perception tool.
This is a research paper, not a shipped product or an API. There is no benchmark suite, deployment cost, or inference-latency figure published alongside the accuracy claims, and no indication yet of a public model release. The comparison is nonetheless a meaningful data point for an industry currently deciding whether to keep building bespoke perception stacks for robotics and AR/VR or to bet on generalist backbones the way natural language processing did when task-specific models gave way to single instruction-following systems.
Teams building perception pipelines for robotics or AR/VR should treat the reported 7x to 500x data-efficiency gap as the number worth trying to reproduce before committing to a next-generation vision stack, since that efficiency claim, more than the raw accuracy parity, is what would make a generalist backbone cheaper to operate than today’s task-specific models.
Per the GenCeption project page published by Google DeepMind researchers on genception.github.io, July 14, 2026.
