Generative Vision Atlas

Research line · emerging

Agentic and search-augmented generation

Give the generator tools, search, or a planning loop, so it can render subjects its training data never covered.

What defines membership

Some generation failures are knowledge failures rather than capability failures, and the fix is retrieval or agency rather than more parameters.

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

Re-Imagen · 2022-09landmark

Retrieves image-text pairs from an external knowledge base to render entities the model never learned, and benchmarks exactly that failure.

Idea2Img · 2023-10core

Puts a multimodal model in a loop that drafts, critiques and rewrites its own prompt.

Evidence

Agentic Visual Generation · 2026-09emerging

Surveys the space once it became a wave rather than a handful of papers.

Improvement

GenArtist · 2024-07core

An agent that plans, decomposes into tool calls, and verifies the result.

At scale

SearchGen · 2026-07emerging

Gives the generator live search so it can render subjects its training data never covered.

What it gets right

  • Attacks a failure mode that scaling does not fix: the model simply not knowing what a thing looks like
  • Composes with any generator, since it changes what reaches the model rather than the model itself

Where it is weak

  • Two papers old, with no independent replication
  • Adds latency and an external dependency at inference time
  • Unclear how much of the gain is retrieval versus the extra compute of a loop