Generative Vision Atlas

Research line · ascendant

Unified understanding and generation

Build one model that both interprets images and produces them, instead of separate understanding and generation systems.

What defines membership

Understanding and generation should share parameters, and doing so makes both better.

How the line developed

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

Before

Emu2 · 2023-12core

Autoregressive prediction of visual embeddings alongside text at multimodal LLM scale.

The idea

Janus-Pro · 2025-01core

Decouples the encoder used for understanding from the one used for generation while sharing the language backbone.

Evidence

BLIP3-o · 2025-05core

Systematic ablation of CLIP-versus-VAE features and flow-matching-versus-MSE objectives, fully open.

Improvement

MetaQuery · 2025-04core

Connects a fully frozen multimodal LLM to a diffusion decoder with learnable queries, so understanding ability cannot be damaged.

Show-o2 · 2025-06core

Autoregressive text and diffusion image generation in one set of weights, now on 3D-VAE latents.

OmniGen2 · 2025-06core

Two decoding pathways with unshared parameters rather than one forced shared path.

At scale

HunyuanImage 3.0 · 2025-09core

An 80B mixture-of-experts unified model with native chain-of-thought before image output.

What it gets right

  • One model to train, serve, and improve
  • Understanding can inform generation, enabling instruction following and reasoning about the requested image
  • The direction every frontier lab is visibly pursuing

Where it is weak

  • Generation training frequently degrades understanding ability
  • What is actually shared varies enormously between systems, which makes the label 'unified' nearly meaningless without qualification
  • Evaluation is immature: few benchmarks test both capabilities together

Also in this line

BAGEL, Chameleon, Emu3, Emu3.5, Janus, STARFlow2, Transfusion