Google Deepmind researchers have developed GenCeption, a model that repurposes a pre-trained video generator to perform classic computer vision tasks such as depth estimation and segmentation. The system, which builds on Alibaba's open-source Wan2.1 video model, delivers results in a single forward pass, guided by text prompts. It trained on a small set of synthetic videos and requires far less data than competing approaches. The model's ability to generalize to real-world footage and untrained categories like animals highlights its potential in computer vision.
GenCeption achieves state-of-the-art performance in depth estimation, segmentation, and 3D pose estimation while needing very little training data. The researchers argue that video generators can serve as the basis for universal world models in computer vision, bridging a gap that has long existed in the field. By using a single architecture for multiple tasks, the model represents outputs as standard three-channel RGB images, enabling it to handle diverse vision tasks efficiently. A text prompt instructs the model on which task to perform, much like an instruction given to a chatbot.
The study highlights that GenCeption matches or beats state-of-the-art results across many benchmarks despite using one architecture for every task. Its depth estimates match those from specialized models such as DepthAnything 3, and it outperforms models like NormalCrafter and Lotus-2 on surface normal estimation. The model also achieves results comparable to Meta's SAM 3 combined with Gemini 3.5 Flash on complex language-guided segmentation. These results suggest that video generation could provide useful representations of space and movement, potentially reducing the need for vast amounts of labeled data.
Source: thedecoder
