Notes

A representation is what an encoder turns pixels into before any task runs, and every one is learned by answering a single question: what should the model be forced to predict? Match a caption, fill in a masked patch, or predict a hidden feature. That one choice decides what the features encode and what they stay blind to, and this atlas reads it across images, video, and 4D geometry.

Every current vision-language model is the same three-part machine: a vision encoder turns pixels into tokens, a connector reshapes them, and a language decoder reasons over vision and text together. This atlas holds that skeleton fixed and maps the choices that fill it, from the LLaVA baseline through to models that early-fuse everything and act.

Almost every modern text-to-video model runs one loop: a text encoder sets the target, a denoiser cleans noise step by step inside a compressed latent space, and a decoder returns pixels. The atlas walks the four slots that loop leaves open, and the quadratic wall that explains why clips stay short.

A world action model predicts a future and keeps that forecast in the path that produces the action. That contract, not any particular backbone, is what separates it from a vision-language-action policy or a plain world model. A full walkthrough of the survey that named the family, and of its thesis that the strongest models generate less of the future while keeping what control needs.