Generative Vision Atlas

Topics

Six subject areas, plus medical imaging as a separate domain. Every paper belongs to at least one.

Topics or research lines: which do you want?

A topic is a subject area, like editing or unified models. Browse by topic when you want to know what exists in a field.

A research line is a family of work sharing one core design bet, like generating in a frozen foundation model's features. Lines cut across topics, and lines answering the same question compete directly. Browse by line when you want to know why groups disagree.

Image generation

Making an image from a prompt: the objectives, backbones and latent spaces that do it.

99 papers · 18 research lines

Representation-space generation

What the generative latent should be. The question this atlas is built around.

46 papers · 6 research lines

Image editing

Changing an existing image, which adds a constraint generation never faces.

38 papers · 9 research lines

Unified understanding and generation

One model that both reads images and makes them, for varying values of 'one'.

20 papers · 3 research lines

Vision foundation models

The pretrained encoders that generative models increasingly depend on.

18 papers · 3 research lines

Vision-language models

Covered where they bear on generation: as text encoders and instruction parsers.

15 papers · 4 research lines

Separate domain

Medical imaging

Kept apart from the six topics above, because clinical imaging answers to a different evidence bar and has reached different conclusions about where generation should happen.

27 papers · 5 research lines