AutoPolish experiment outputs

Every output behind the numbers and figures of AutoPolish: Improving Visual Aesthetics with Synthetic Audience Critique: the analysis results, the raw logs of every model run, and the images produced by every editing run.

Code · Project page · Paper (arXiv, coming soon)

Download

The files are stored verbatim with their directory layout intact, because the analysis scripts read them by path. The code repository restores them in place:

git clone https://github.com/forall-labs/AutoPolish && cd AutoPolish
pip install -r requirements.txt
python tools/fetch_from_hf.py autopolish

Repo prefix	Restores to	Files	Size
raw/analysis/	results/	47	2 MB
raw/runs/	data/results/	2,321	2.2 GB

layout_manifest.json records that mapping. The download is cached by the Hub client before it is copied into place, so budget about twice its size in free disk space.

Contents

raw/analysis/ holds the cached analysis results (*.json) and figures that the reproduction guide maps to each result in the paper.

raw/runs/ holds one folder per run:

Run	What it is
para_full, eva_full, lapis_full	the synthetic audience replaying each dataset's real raters as personas (greedy decoding)
para_full_t07, eva_full_t07, lapis_full_t07	the same, sampled at temperature 0.7
para_blind, eva_blind, lapis_blind	the persona-blind control
rapidata_persona	pairwise preferences on AI-generated images, replaying Rapidata voters
c4_run2	the editing run reported in the paper: four critics, ten rounds, 100 photographs, with every candidate and committed image
c4_run1	an earlier editing run, kept for comparison
c4_selfrefine	the earlier self-refine loop the paper compares against

Judge runs are split into a summary file per shard (configuration, the persona description of every rater, and precomputed metrics) and chunk files holding one record per (image, rater). Editing runs hold logs/ (one record per image and round) and edits/ (the images).

Terms

These outputs are derived from four public datasets and inherit their terms:

• the persona descriptions in the run logs are built from rater attributes released with PARA, EVA and LAPIS;
• the edited images are derivatives of PARA and EVA photographs;
• the pairwise runs use the Rapidata preference dataset (CDLA-Permissive-2.0).

Use them for research and in line with the terms of those datasets. The source images themselves are not included; download them from the links above. The code that produced and analyzes these files is MIT-licensed.

Citation

@article{fadaeinejad2026autopolish,
  title   = {AutoPolish: Improving Visual Aesthetics with Synthetic Audience Critique},
  author  = {Fadaeinejad, Amin and Pakdamansavoji, Sajjad},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}