Google DeepMind's new paper "Video Generation Models are General-Purpose Vision Learners" argues: video generation models can be turned around to do general-purpose visual understanding. The team (Kaiming He, Joao Carreira, Andrew Zisserman as co-authors) releases GenCeption — using a pretrained text-to-video diffusion model as a perception backbone, switching tasks via text instructions. On depth, surface normal, camera pose, segmentation, 3D keypoint tasks, it matches or exceeds specialized models like DepthAnything3, SAM3, D4RT, VGGT-Omega, and significantly leads V-JEPA and Video MAE. Data efficiency is stunning: to reach the same performance as D4RT and VGGT-Omega, training data only needs 1/7 to 1/500. Training only on synthetic human videos, it generalizes to real scenes and OOD objects — typical emergent behavior. It directly answers the old question of the visual-domain equivalent of next-token prediction: the answer is to use large-scale video generation itself as a pretraining paradigm. When video generation models are not just creation tools but the foundation of general visual intelligence, the Sora / Kling / Veo track will be repriced: generation is the entry point, understanding is the end-game. ECCV 2026 acceptance is just the beginning.