Generative Vision Atlas

Research line · dominant

Contrastive language-image pretraining

Learn visual features by aligning images to their captions across web-scale data.

What defines membership

Language supervision is what makes a representation semantically meaningful and directly promptable, and that is worth more than what it costs in spatial precision.

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

CLIP · 2021-02landmark

Contrastive image-text pretraining at web scale, producing the zero-shot transfer that made vision-language models practical.

Evidence

Perception Encoder · 2025-04core

Finds the most useful embeddings sit in intermediate layers rather than at the output, which complicates the idea of reading a frozen encoder at a single point.

Improvement

SigLIP · 2023-03core

Swaps the softmax contrastive loss for a pairwise sigmoid, removing the dependence on very large batches.

SigLIP 2 · 2025-02core

Adds captioning and self-distillation objectives to the contrastive recipe, specifically to fix weak dense and localization features.

AIMv2 · 2024-11strong-followup

A third route: autoregressive prediction of patches and text together, rather than contrastive alignment.

What it gets right

  • Directly promptable and language-aligned, which self-distilled features are not
  • The default front end for every vision-language model
  • SigLIP 2 is the frozen encoder Scale-RAE chose for text-to-image

Where it is weak

  • Consistently beaten by DINOv2 as an alignment target in head-to-head ablations
  • Contrastive training optimizes a global match, which historically cost dense and localization quality
  • Caption quality bounds what the representation can learn

Competing answers

Open problems it has not solved