Tech
Video Generation AI Transformed into a General-Purpose Vision Model: The Boundaries of "World Understanding" Revealed by GenCeption
GenCeption, developed by a team including Google DeepMind, repurposes the internal representations of video generation AI to jointly estimate multiple visual quantities such as depth and surface normals. It quantitatively demonstrates that pretraining for video generation is effective for understanding the structure of the world, achieving accuracy comparable to specialized models with minimal additional training.
