🤖 Google DeepMind has introduced GenCeption — a single feed-forward model for multi-task computer vision.
The model utilizes a frozen diffusion foundation (text-to-video) and is trained on synthetic videos, allowing it to perform tasks such as depth estimation, segmentation, and camera pose estimation via text instructions.
🌍 The transition to a unified vision backbone reduces computational complexity and simplifies the deployment of multi-task systems in robotics and autonomous vehicles.
👤 Instead of installing multiple disparate neural networks, a single compact model can be used that understands everything — from spatial geometry to objects — based on a text description.
Source 1: https://genception.github.io/