Generative Vision Atlas

Research line · emerging

Editing in a representation latent

Perform the edit inside a foundation-model feature space rather than a reconstruction latent.

What defines membership

If generation is better in a semantic space, editing should be too, because the axes of that space correspond to meaningful attributes.

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

PS-VAE · 2025-12core

Builds instruction editing from DINOv2 and SigLIP2 features, and states that a frozen representation autoencoder's weak reconstruction leaves details inconsistent with the input — so it unfreezes the encoder.

Evidence

DiDAE · 2026-01core

Gets genuinely frozen-latent editing to work by finding directions through Procrustes alignment and SVD in a frozen CLIP space, at the price of restricting to face attributes and synthetic shapes.

Improvement

UniSpace · 2026-08emerging

A unified model on a reparameterized frozen vision-transformer space; flagged as needing independent verification.

Limitation

RPiAE · 2026-03core

Independently reaches the same diagnosis: frozen-encoder tokenizers have limited reconstruction fidelity, which degrades editing, so it fine-tunes the encoder too.

What it gets right

  • Edit directions in a semantic space are meaningful by construction
  • Directly extends the representation-latent generation line into editing
  • The one genuinely open frontier in an otherwise crowded area

Where it is weak

  • Reconstruction fidelity is the named bottleneck, by two independent groups
  • The only strictly-frozen success is confined to narrow domains
  • Both general-purpose attempts abandoned strict freezing to make it work
  • No inversion, adapter or control analogue exists yet for this latent family

Competing answers

Open problems it has not solved