Generative Vision Atlas

▶ Open the atlas  ·  Source on GitHub

A living literature review of modern generative vision, built around a single organizing question: where should generation happen? In a compression code, inside a foundation model’s features, in discrete tokens, or in raw pixels?

Image generation has largely settled its arguments about objectives and architectures. Where generation should happen is still contested, so the atlas is organized around that disagreement rather than around a taxonomy of methods.

How it is organized

204 papers sit inside 35 research lines, each defined by one explicit bet. A line states what its group is claiming, what that claim costs them, and which lines disagree. Six design questions structure the whole atlas — what space you generate in, what happens at sampling time, how understanding and generation are arranged, and so on — and lines answering the same question compete directly.

Medical imaging is kept deliberately separate: 27 papers across 5 lines. Clinical imaging answers the same design question differently, and answers to a different evidence bar — not a better distribution distance, but whether a diagnostic model improves and whether a radiologist is fooled.

What it does that a paper list does not

It refuses to publish a leaderboard. One paper reports the same model at FID 1.65, 1.49, 1.14 and 1.06 under four guidance methods — a swing larger than most architectural claims. Every one of the 46 benchmark numbers here carries the guidance method, training budget and model size that produced it, so the reader can see when a comparison is not actually a comparison.

It tracks what is unsolved. 11 open problems, each linked to the papers attacking it and to the research lines that carry it as a known weakness. Where a line is weak, its own page says so.

It keeps itself current. A pipeline sweeps arXiv and Hugging Face every three days, filters for genuine relevance, and proposes candidates for review. Nothing is added automatically, and nothing is claimed without a fetched source — 203 of the 204 entries are verified against arXiv.

Reproducible evidence

Four notebooks with real outputs from an A100: what each latent discards, RAE measured against the VAE it replaces, how editing methods actually differ, and why guidance breaks benchmark comparisons.