Generative Vision Atlas

Updates

A pipeline sweeps arXiv and Hugging Face daily papers every three days, filters for genuine relevance, scores what survives, and proposes it for review. Nothing enters the atlas automatically, and no explanatory claim is written without a fetched source.

The pipeline

discoverfilter for relevancescore qualitycheck for duplicatesproposeverifymerge

Last run 2026-09-10, which processed 321 new preprints:

321
discovered
257
not image generation
56
out of scope
0
already here
5
proposed for review

The aggressive filtering is deliberate. World-action models, depth estimation and language-model post-training papers share almost all of generative vision's vocabulary, and without a hard topic gate they dominate the queue.

Waiting for review (5)

Proposed by the last sweep, not yet part of the atlas. Each shows why it was surfaced.

15.0 Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models

core topic appears in the title · affiliated with a major lab or university · code or weights released

15.0 Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

core topic appears in the title · affiliated with a major lab or university · title suggests a substantive contribution

13.0 FlowCPO: A Unified Divergence View of Preference Alignment for Flow Models

affiliated with a major lab or university · title suggests a substantive contribution

11.0 Efficient Fairness Auditing Across Guidance Scales in Text-to-Image Diffusion Models via Causal

core topic appears in the title · affiliated with a major lab or university · title suggests a substantive contribution

6.0 Muon-C: Operator-Aligned Muon for Convolutional Kernels

Coverage, stated honestly

What is actually written versus what is still only metadata. The validator tracks this, so these numbers cannot quietly drift.

203/204
have the authors' abstract
fetched verbatim from arXiv, never paraphrased
136/138
key papers explained from full text
problem, idea, method, evidence, limitations, and why it matters
0
key papers still to write up
these show the abstract and structured facts, and say so

Standards

Every paper is checked

203 of 204 entries have had their arXiv record fetched and their title matched. The remainder is a journal paper with a DOI and no arXiv id. Six wrong titles have been caught this way so far.

Automation cannot promote

The pipeline never proposes a paper above the lowest tier, and never writes to the atlas directly. A human or a review session decides.

Gaps are stated

Where a line has thin coverage, no independent replication, or an unverified claim, its page says so rather than omitting it.