Generative Vision Atlas

Research line · ascendant

Agglomerative multi-teacher distillation

Distil several specialist foundation models into one backbone rather than choosing between their objectives.

What defines membership

No single pretraining objective is best at everything, so the right move is to inherit several at once through distillation.

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

AM-RADIO · 2023-12core

Distils CLIP, DINOv2 and SAM into one student backbone.

Evidence

Theia · 2024-07strong-followup

The same bet pursued independently, validated in robotics.

UNIC · 2024-08strong-followup

The same bet again from another group, with teacher-dropping regularisation. Together these show agglomeration is a general idea, not one company's.

Improvement

RADIOv2.5 · 2024-12strong-followup

Fixes resolution and mode-switching failures in the first recipe.

At scale

C-RADIOv4 · 2026-01emerging

Now distils DINOv3, SigLIP 2 and SAM 3, the same three encoders the generative-latent work independently converged on.

What it gets right

  • Sidesteps the self-distillation versus contrastive argument by taking both
  • One backbone to deploy instead of several specialists
  • Its choice of teachers is a useful read on what the field currently rates

Where it is weak

  • A student is bounded by its teachers and inherits their blind spots
  • Not yet evaluated as a generative latent, which is the use this atlas cares about most
  • Distillation adds a training stage and a set of hyperparameters of its own