Generative Vision Atlas

Research line · ascendant

Medical: domain-trained latent

Train an autoencoder on medical data itself, then run diffusion in that latent.

What defines membership

A latent is necessary for 3D volumes, but it must be trained on the target anatomy rather than borrowed from natural images.

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

Pinaya brain-MRI LDM · 2022-09core

Latent diffusion for 3D brain MRI conditioned on age, sex and structure volumes, with 100,000 synthetic brains released.

Evidence

MONAI Generative Models · 2023-07core

Turns the recipe into shared infrastructure most later medical generative work builds on.

At scale

MAISI · 2024-09core

3D CT at 512-cubed via a domain-trained VAE-GAN compressing to 128-cubed, with the memory constraint stated outright as the reason.

What it gets right

  • The only practical route to full 3D volumes at clinical resolution
  • Latent statistics match the anatomy rather than natural photographs
  • Explicitly reasoned: the memory argument is stated, not assumed

Where it is weak

  • Needs enough in-domain data to train an autoencoder, which many modalities lack
  • Demonstrated memorization risk in the 3D latent setting
  • Little downstream clinical evidence compared with the pixel-space line

Open problems it has not solved