Google DeepMind Introduces GenCeption, Repurposing Video Generators for Computer Vision

Google DeepMind has introduced GenCeption, a model based on Alibaba's open-source Wan2.1 video generator that can perform multiple classical computer vision tasks within a single framework. Unlike traditional task-specific models, GenCeption repurposes a video generation model to handle depth estimation, segmentation, and pose estimation. This approach challenges the traditional paradigm of using separate, task-specific models for different computer vision tasks by demonstrating that video generation models already contain rich representations of the physical world. It could pave the way for more unified, general-purpose vision models that are faster and more versatile. GenCeption processes tasks in a single forward pass instead of multi-step denoising, allowing a 14-billion-parameter version to process an 81-frame video in about 10 seconds. Although trained primarily on synthetic data of single humans, the model generalizes remarkably well to real-world footage, multiple people, and animals.