🎥 Video Generation – Maitreya

Predicting frames instead of pixels - what a model has to get right for a generated sequence to be a video and not a slideshow.

A Good Frame Is Not a Good Video

generative-models

video

world-models

An image model learns whether a frame looks plausible. A video model has to learn which frames could belong to the same trajectory. This post builds two generators with provably identical per-frame distributions — one produces motion, the other produces a flicker — and shows why metrics that depend only on per-frame marginals are blind to the difference.

Aug 12, 2026

Deep dive