README.md

metadata

license: cc-by-4.0
pretty_name: FactorAtlas
size_categories:
  - 10K<n<100K
tags:
  - image
  - synthetic-data
  - vision-language
  - benchmark

FactorAtlas

FactorAtlas is the dataset accompanying Seeing Is Not Addressing: Auditing Linguistic Access to Frozen Visual Geometry.

It contains 23,040 controlled synthetic images spanning 8 shapes, 12 hues, 10 surface patterns, 12 rendering contexts, and 2 renderer seeds.

Use with the reference code

Clone the reference implementation, install its dependencies, then run from the repository root:

pip install -r requirements.txt
bash scripts/run_factoratlas.sh

The runner downloads FactorAtlas into data/factoratlas, caches frozen model features, and writes the complete audit report locally.

Files

• metadata.json — one record per image with factor labels and rendering keys
• data/factoratlas-*.tar — 24 image shards containing 224×224 PNGs
• data/SHA256SUMS — SHA-256 checksums for the shards