DeepMind unveils GenCeption vision model

Video generation models are proving to be the long-sought foundation for general computer vision, moving beyond static images to leverage internal world models of physics and time. Repurposing video diffusion backbones rather than starting from scratch represents a massive leap in data and compute efficiency.

• –**Spatio-Temporal Prior Extraction:** Repurposing video diffusion backbones unlocks native physical world understanding (4D causality) without generative sampling overhead.
• –**Instructable Multi-Tasking:** Handles highly diverse tasks—from depth and pose estimation to segmentation—through a single, text-guided feed-forward model.
• –**Massive Data Efficiency:** Achieves state-of-the-art results while requiring up to 500x less downstream training data than specialized models.
• –**Paradigmatic Shift:** Validates next-token/frame prediction on video as a universal computer vision pre-training objective, mimicking the NLP foundation model trajectory.