Generative Vision Atlas

Research line · ascendant

Editing inside a unified model

Treat editing as one more thing a single multimodal model does, with no editing-specific pipeline.

What defines membership

A model that understands images and generates them needs no separate editing machinery; the source image is simply more context.

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

BAGEL · 2025-05core

Mixture-of-transformer-experts routing understanding and generation over shared sequences, trained on interleaved data.

Improvement

OmniGen2 · 2025-06core

Unified at the sequence level but explicit that it uses two decoding pathways with unshared parameters.

Emu3.5 · 2025-10core

The closest to literally one network: a single autoregressive transformer over interleaved tokens with one objective.

HiDream-O1-Image · 2026-05core

Removes the VAE and the separate text encoder entirely, embedding pixels, text and condition tokens in one shared space processed by the same blocks, so editing and generation are the same forward pass.

What it gets right

  • One model to train and serve for understanding, generation and editing
  • Instruction understanding comes from the language side rather than a bolt-on parser
  • The direction every frontier lab is visibly pursuing

Where it is weak

  • 'Unified' hides real architectural differences: most still route generation through a separate head or expert
  • Generation training can degrade understanding
  • The most unified example so far is limited to 512 pixel training