Google DeepMind's GenCeption: One AI Model Does Multiple Vision Tasks

Google DeepMind has introduced GenCeption, a model that turns a video generator into a general-purpose vision tool. It's built on Alibaba's open-source Wan2.1 model and handles tasks like depth estimation, image segmentation, and 3D pose estimation with a single model.

Unlike traditional models that take many steps, GenCeption makes predictions in one go, making it faster. A small model processes 81 frames in about 6 seconds, while a larger one (14 billion parameters) takes about 10 seconds.

The model was trained on just 7,500 synthetic videos from 800 digital humans and 200 motion sequences. Even though it only saw single-person synthetic clips, it works on real multi-person videos and animals, capturing fine details like cat whiskers and single hair strands.