Image Generators are Generalist Vision Learners - Deep Dive

https://arxiv.org/abs/2604.20329v1 What does it actually mean for a generative visual model to "understand" what it sees? That's the question driving this episode, and it turns out to be harder to answer than it sounds. We start with "Image Generators are Generalist Vision Learners," which introduces VisionBanana, a model built by instruction-tuning NanoBanana Pro on a mix of its original data and a small amount of task-specific vision data. The trick is reframing perception itself as image generation, treating outputs like segmentation masks and depth maps as RGB images. The result is a single generalist that holds its own against dedicated specialists like SAM3 and Depth Anything, suggesting image generation plays a role for vision similar to what next-token prediction plays for language. From there we widen the lens with two companion papers. The first asks whether video models can genuinely reason by generating frames, using maze-solving as a test. The second probes video world models from the inside, looking at where physical variables like velocity and mass are actually encoded. Together, the three papers sketch a more honest picture of what generative pretraining does, and doesn't, buy us. Related papers discussed: - Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks: https://arxiv.org/abs/2511.15065v1 - Interpreting Physics in Video World Models: https://arxiv.org/abs/2602.07050v1 This podcast is from Colin Davis (colin-davis.com) using Claude & Elevenlabs.