Sajjad Pakdamansavoji* · Amin Fadaeinejad*
Independent Research · *equal contribution
Top: source photographs. Bottom: the image AutoPolish keeps after ten rounds of critique by a simulated audience.
Text-guided image editing is a back-and-forth in which one person decides what to change, so the result reflects that person's taste rather than the audience's. AutoPolish replaces that person with a synthetic audience: many copies of one frozen vision-language model, each reacting to the image as a different viewer described by a short profile. Their combined complaints become edit instructions, and an independent scorer keeps an edit only if the image scores higher and still resembles the original. Nothing is trained.
- Groups are predictable, individuals are not. Only 19–47% of one person's rating is shared with other viewers, while the averages of two separate groups agree at up to 0.96.
- The audience tells viewer groups apart. It separates them up to 10× better than the same model without profiles (0.166 against 0.016 on paintings).
- It transfers to AI-generated images. On 898 pairs it agrees with the human majority 6 percentage points more often than always picking the more popular model.
- In editing it proposes more. AutoPolish matches generic critics in average gain while keeping images recognizable, and proposes 3× as many distinct changes per round.
- Simulate the audience. Each persona card describes a real rater from a public dataset.
A frozen
Qwen2-VL-7B-Instructrates the image as that person and writes the one change they would want. - Turn complaints into an edit. The panel's complaints are condensed into one instruction
of at most fifteen words, and
FLUX.1-Kontext-devapplies the accumulated instruction to the original image, so artifacts cannot pile up. - Keep only real improvements. A LAION aesthetic predictor (a different model family from the critic) scores each candidate, and DINOv2 checks that it still resembles the original. The loop keeps an edit only if both checks pass, for ten rounds.
git clone https://github.com/forall-labs/AutoPolish && cd AutoPolish
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt # analysis, figures and data download (CPU)
tools/setup_gpu.sh # adds torch and the model stack (GPU, optional)The editor, FLUX.1-Kontext-dev, is a gated model: accept its license on the
model page and run
huggingface-cli login before the first edit.
The playground notebook runs a ten-persona panel and
one full editing round on the example images in assets/examples/, and reprints the paper's
headline numbers. In code:
import sys; sys.path.insert(0, "src")
from PIL import Image
from persona import QwenVLBackend
from editor import SocietyCritic, distill_instruction
backend = QwenVLBackend("Qwen/Qwen2-VL-7B-Instruct") # frozen, shared by every persona
panel = [
"a female aged 30-34, education level: university; art experience: proficient; "
"photography experience: competent.",
"a male aged 18-21, education level: junior college; art experience: beginner; "
"photography experience: beginner.",
# ... one plain-text card per viewer
]
image = Image.open("assets/examples/basketball.jpg").convert("RGB")
critique = SocietyCritic(backend, panel).critique(image)
print(critique.panel_scores) # one score per viewer
print(critique.complaints) # the change each viewer wants
print(distill_instruction(backend, image, accumulated="", new_complaints=critique.complaints))The full experiment, ten rounds on 100 photographs with four critics, is one command:
OUTPUT_ROOT=outputs/my_run tools/run_autopolish.shEvery number and figure in the paper can be regenerated. The reproduction guide has three tiers, cheapest first:
| Tier | Needs | Time | Reproduces |
|---|---|---|---|
| 1 | CPU + experiment outputs (2.2 GB) | minutes | every figure and number |
| 2 | CPU + experiment outputs (2.2 GB) | ~1 hour | the analysis itself, from the raw run logs |
| 3 | GPUs + source datasets | days | the raw run logs, from scratch |
python tools/fetch_from_hf.py autopolish # experiment outputs from the Hugging Face Hub
cd analysis && python paper_figs.py # every paper figure -> results/paper_figs/| Source | Terms | |
|---|---|---|
| PARA | dataset page | access on request |
| EVA | kang-gnak/eva-dataset | CC0-1.0 |
| LAPIS | dataset page | terms of use form |
| Rapidata | Rapidata/700k_Human_Preference_Dataset_FLUX_SD3_MJ_DALLE3 | CDLA-Permissive-2.0 |
| Our outputs | savoji/AUTOPOLISH | analysis, run logs and edited images |
The source datasets are not redistributed here. DATASETS.md explains how to download them and where to put them.
src/
persona/ the synthetic audience: a frozen VLM role-plays one viewer at a time
editor/ the AutoPolish loop: critics, FLUX editor, aesthetic scorer, similarity check
experiments/ drivers that run the models and write run logs
{para,eva,lapis}_pipeline.py replay each dataset's real raters as personas
rapidata_pipeline.py pairwise preferences on AI-generated images
c4_refine.py the AutoPolish editing loop (the paper's main system)
claim4_pipeline.py the earlier self-refine loop the paper compares against
analysis/ CPU-only re-analysis of the run logs -> results/*.json and the paper figures
tools/ data download, GPU setup, the one-command editing run, website assets
notebooks/ autopolish_playground.ipynb, a hands-on tour
docs/ reproduction guide, datasets, GPU runbook, architecture
website/ the project page (GitHub Pages)
ARCHITECTURE.md explains how the pieces fit together and the design decisions behind them.
@article{pakdamansavoji2026autopolish,
title = {AutoPolish: Improving Visual Aesthetics with Synthetic Audience Critique},
author = {Pakdamansavoji, Sajjad and Fadaeinejad, Amin},
journal = {arXiv preprint arXiv:XXXX.XXXXX},
year = {2026}
}The code is released under the MIT License. The datasets and model weights it uses
keep their own licenses; in particular, FLUX.1-Kontext-dev is distributed under the
FLUX.1 [dev] Non-Commercial License.
We thank the authors of PARA, EVA, LAPIS and the Rapidata preference dataset for releasing their data, and the authors of Qwen2-VL, FLUX.1 Kontext, the LAION aesthetic predictor and DINOv2 for their models.
