Generative Vision Atlas

Research line · dominant

Encoder plus projector

Bolt a frozen vision encoder onto a language model with a small trained projection, and treat the projected patches as prompt tokens.

What defines membership

The language model's embedding space is a good enough interface that a linear map into it is sufficient; nothing about the language model needs to change.

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

LLaVA · 2023-04landmark

A single linear projection from frozen CLIP into a language model, trained on generated instruction data.

At scale

LLaVA-OneVision · 2024-08core

Extends the same interface to multi-image and video with resolution-adaptive tiling, without architectural change.

Combination

MetaQuery · 2025-04core

Runs the pattern in reverse for generation: learnable queries read a frozen multimodal model and drive a diffusion decoder.

What it gets right

  • Cheapest possible bridge between vision and language, and easy to train
  • The frozen language model keeps its abilities intact
  • The same interface can be reversed to condition a generator, which is how frozen-model generation works

Where it is weak

  • A linear map is a narrow channel for everything an image contains
  • Visual tokens consume context length the language model needs elsewhere
  • Fine spatial grounding is limited by what the frozen encoder chose to keep