UAR · A field guide to visual generation
Represent · Reconstruct · Generate · Understand
From reconstructing an image to modeling a visual world.
You know how to program, but visual generation may be new to you. Each article connects prior work, the new idea, architecture, equations, and implementation. Manipulate a mechanism, then inspect the evidence behind its claims.
Start reading · Learning map · Compare methods
Recover one image
Start with the bottleneck. Then compare a one-pass decoder with iterative reconstruction.
Generate new images
Separate the representation from the distribution that generates it.
Read and write images
Ask which features preserve pixels, labels, and text alignment. Test each claim separately.
These are question-driven reading paths, not a historical citation graph. The learning map includes all 9 foundation lessons and 21 papers.