Forall Labs

ProjectsImage editingSynthetic audiences

AutoPolish

Improving visual aesthetics with synthetic audience critique

Many copies of one vision-language model each react to a photograph as a different viewer. Their combined complaints become edit instructions, and an independent scorer keeps an edit only if the image scores higher and still looks like the original.

Authors
Amin Fadaeinejad*
Sajjad Pakdamansavoji*
* equal contribution
Affiliation
Independent Research
Released
October 2026
arXiv soon Code Hugging Face soon Thread soon BibTeX
A concrete wall with a bright red bench after ten rounds of AutoPolish Round 10
edit kept edit rejected, best image unchanged
Held-out score
5.54+0.50
Similarity
0.91floor 0.78
Edits kept
4of 10

Round 10 kept. The edit raised the score, so it became the new best image.

Instruction tried

Vividly color the red bench to stand out more against the muted background.

The panel's reactions that led to itscore of 100
      Figure 1.Ten viewers, ten rounds, one bench: a real run from the paper on an EVA photograph. Choose a round to see what each simulated viewer scored, the complaints they shared, the instruction those complaints became, and whether the edit was kept. Hold the button on the image to compare with the image before that round.

      Text-guided image editing is an iterative, human-in-the-loop process that does not scale and restricts the final image to one operator's taste, which often misaligns with the target audience. We propose AutoPolish, which replaces that operator with a synthetic audience: many copies of a vision-language model, each reacting as a different viewer described by a short profile. An auto-research style loop turns their combined feedback into edits, keeping each edit only when an independent scorer rates it higher.

      Across three datasets of photographs and paintings, only 19 to 47% of one person's rating is shared with other viewers, while the averages of two separate groups correlate at up to 0.96, so one rater poorly predicts the population average. Our audience captures how viewer groups differ up to 10× better than the same model without profiles, and agrees with the human majority on pairs of AI-generated images 6 percentage points more often than a simple baseline. In editing, AutoPolish matches generic critics in average gain while keeping images recognizable; it improves fewer images than a fixed instruction but makes larger edits and proposes 3× as many distinct changes.

      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.

      Reliability, three datasets
      10×

      better separation of viewer groups than the same model without profiles, 0.166 against 0.016 on paintings.

      Group separation, LAPIS
      +6 pts

      agreement with the human majority on AI-generated pairs, 65.7% with twenty personas against 60.0% for always picking the more popular model.

      Rapidata pairs
      3×

      as many distinct changes proposed per round as a generic critic, 3.12 against 1.00, at the same average gain.

      Editing, 100 photographs

      How AutoPolish works

      Nothing is trained. One frozen model plays every viewer, and the model that judges an edit comes from a different family than the one that proposes it, so the loop cannot raise its own score without improving the image. An audience proposes, an independent judge decides.

      Pipeline: persona cards feed a role-playing vision-language model whose reactions are aggregated; the complaints steer an instruction, an image editor applies it, and a held-out judge keeps the edit only if it scores higher.
      Figure 2.Top. A frozen Qwen2-VL-7B is given N persona cards, each describing one viewer in plain text, and their reactions are combined into a group reaction. Bottom. The complaints become one instruction for FLUX.1 Kontext, and a held-out judge keeps the edit only if it scores higher while staying similar to the original.
      1. Step 01

        Simulate the audience

        Each persona card describes a real rater from the dataset: age, gender, education, art and photography experience, personality or nationality. The model rates the image as that person and writes the one change they would want.

      2. Step 02

        Turn complaints into an edit

        The complaints of the whole panel are condensed into one instruction of at most fifteen words. The editor always applies the accumulated instruction to the original image, so artifacts cannot pile up, and proposes three candidates.

      3. Step 03

        Keep only real improvements

        A LAION aesthetic predictor scores each candidate, and DINOv2 checks that it still resembles the original (similarity at least 0.78). The best candidate replaces the current image only if both checks pass. The loop runs for ten rounds.

      Groups are predictable where individuals are not

      We test the audience on 4,000 photographs from PARA, 4,070 from EVA and 4,000 paintings from LAPIS, replaying each dataset's real raters as personas, and on 898 pairs of AI-generated images from Rapidata.

      The persona changes what the model writes more than what it scores

      Three bar charts: group error before and after calibration; attribute AUC from the comment text with and without a persona; distinct comments per rater.
      Figure 3.Left. Error of the audience's average rating before and after calibration, against the prior. Middle. How well a classifier recovers a rater's attribute from the comment alone. Right. Distinct comments per rater of an image, with and without a persona. Calibration beats the prior on every dataset, and the persona makes the comments more diverse and more telling of the viewer.

      The persona moves the scores the way real groups differ

      Three scatter plots of real against simulated group departures for PARA, EVA and LAPIS, with correlations of +0.37, −0.24 and +0.39.
      Figure 4.How far each group of viewers departs from the crowd in its real ratings (horizontal) and in the model's ratings under that group's persona (vertical). Each point is one group, sized by how many ratings it has. The model follows the real groups on PARA and LAPIS but not on EVA, which records the least about its raters.

      A larger panel agrees more often with the crowd

      Left: agreement with the crowd majority rising with panel size above the popularity prior. Right: panel preference share against crowd preference share per pair.
      Figure 5.Left. Agreement with the crowd's majority on AI-generated pairs as the panel grows, against the prior of always picking the more popular model. Right. The panel's preference share against the crowd's, per pair (r = 0.47).

      In editing, AutoPolish matches the other critics on average and makes larger changes

      Left: best-so-far score over ten rounds for four critics. Right: final gain against similarity to the original, all above the 0.78 floor.
      Figure 6.Left. Best score so far at each round, averaged over 100 photographs, with a 95% bootstrap band. Right. Final gain of each image against its similarity to the original, with the 0.78 floor marked. Almost every image stays far above the floor.
      Table 1.Ten rounds on 100 photographs (50 PARA, 50 EVA). Gain is the rise in the held-out aesthetic score. The differences in average gain between critics are within their confidence intervals; reward-only optimizes the scorer directly and is a reference, not a comparable critic.
      CriticGain95% CIImages improvedGain per improvedEdit sizeComplaints per round
      static string0.164[0.135, 0.196]84%0.1950.032n/a
      blind VLM0.155[0.126, 0.185]72%0.2150.0361.00
      AutoPolish0.170[0.136, 0.207]69%0.2470.0493.12
      reward-only0.187[0.157, 0.221]86%0.2180.040n/a

      AutoPolish often changes what the photo contains, while the other critics mostly adjust light and color. Choose a photograph below to compare each critic's best image after ten rounds.

      Figure 7.Each critic's best image after ten rounds, with its held-out score and the change from the source.

      Citation

      Please cite this work as

      @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}
}

      Reproducing

      The code, the analysis behind every figure and table, and the steps to rerun them are in the GitHub repository; start with the reproduction guide.

      Data

      The experiments use PARA, EVA, LAPIS and Rapidata under their authors' terms. The photographs on this page come from EVA, released under CC0.