Generative Vision Atlas

Research line · dominant

Medical: borrowed natural-image latent

Fine-tune Stable Diffusion on medical data while keeping its natural-image VAE frozen.

What defines membership

The pretrained natural-image autoencoder is good enough for medical images, and the transfer-learning saving outweighs any domain mismatch.

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

RoentGen · 2022-11landmark

Adapts Stable Diffusion to chest X-rays with the VAE frozen; downstream classifiers improve by up to five points.

Improvement

RadEdit · 2023-12core

Masked editing to manufacture the distribution shifts a deployed model will meet, as a stress test.

BiomedJourney · 2023-10core

Disease-progression editing learned from pairs of longitudinal radiology reports.

At scale

PixCell · 2025-06core

A pathology foundation generator on SD3's VAE, using foundation-model embeddings as the condition because pathology has no captions.

What it gets right

  • Cheapest path to a working medical generator, reusing a large pretrained model
  • Best-populated line, with real editing systems and at least one reader study
  • Inherits the entire Stable Diffusion tooling ecosystem

Where it is weak

  • The autoencoder was trained on natural photographs and is never validated for clinically relevant detail retention
  • No paper in this line ablates its own central assumption
  • Evidence that ImageNet-family features miss clinical structure argues against it

Competing answers

Open problems it has not solved