Generative Vision Atlas

Research line · dominant

Editing in the VAE latent

Train an editor that reads an instruction and rewrites the image inside a standard reconstruction-trained latent.

What defines membership

Editing is a supervised generation problem: given enough instruction-image-result triplets, a model learns what to change and what to leave.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

InstructPix2Pix · 2022-11landmark

Manufactures the training data the task needed, with a language model writing instructions and a prompt-to-prompt pipeline rendering before and after images.

Evidence

MagicBrush · 2023-06core

Human annotators required to leave unrelated regions untouched; concludes even tuned models fall short of practical use.

Improvement

Emu Edit · 2023-11core

One editor across region, free-form and recognition tasks, steered by learned task embeddings.

Step1X-Edit · 2025-04core

A multimodal language model parses the instruction and a diffusion decoder executes it; source of GEdit-Bench.

At scale

FLUX.1 Kontext · 2025-06landmark

Concatenates reference and target tokens into one sequence in a rectified-flow model, and concedes multi-turn character drift as its named limitation.

What it gets right

  • By far the best-populated line, with mature tooling and open weights
  • Handles open-vocabulary instructions rather than a fixed operation set
  • Inherits every improvement to the underlying text-to-image model

Where it is weak

  • Preservation of untouched regions is learned statistically, never guaranteed
  • Drift compounds across multi-turn edits
  • Depends on synthetic training triplets whose quality bounds the result

Competing answers

Open problems it has not solved