I Am AdMan: A Pipeline for Automatic Generation of Personalized Advertising Imagery
Victor Kolominsky-Rabas, Leopold Müller, Claudius Budcke, Niklas Kühl
cs.AI · cs.HC
Quick verdict
Proposes multi-agent pipeline turning customer data into personalized ad images; credible feasibility demo, yet no baselines, single-run and judge-limited.
Summary
The paper proposes ADMAN, a multi-agent pipeline that converts tabular customer attributes into a textual persona, uses an LLM prompt agent plus a text-to-image model conditioned on a product reference photo to generate personalized advertisement images, and applies an LLM judge with bounded (three-attempt) regeneration and best-of fallback. The authors instantiate it with an OpenAI (gpt-5/gpt-image-1) and a Google Gemini configuration and evaluate on four products spanning the FCB grid (electric vehicle, luxury watch, detergent, soft drink), combining a qualitative expert focus group (8 researchers, 3 of them authors; six fictional celebrity personas) with 100 real Prolific profiles — 1745 images total, 800 delivered. Human-labeled artifact rates favor Gemini (final-set AR 0.20 vs 0.42; rAR 0.04 vs 0.21), while Gemini's judge passes far fewer images (PR 0.40 vs 0.79); all McNemar comparisons are significant (p<0.001). Text rendering dominates artifacts, and both judges detect mainly text defects, missing structural ones. Limitations: a single generation run per configuration, no non-personalized baseline, personalization assessed only qualitatively on 24 images, low annotation agreement in two headline cells (κ 0.171/0.318), and no cost or downstream-effect measurement. Code is promised upon publication.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A fully automated multi-agent pipeline can turn individual customer profiles into photorealistic personalized advertising images, with persona attributes visibly propagating into scenes (RQ1).
- Generation quality varies substantially by product complexity and model configuration, with text rendering a persistent artifact source across both configurations.
- The LLM judge improves delivered quality but functions mainly as a text-defect detector, missing structural and physical artifacts (RQ2).
- Bounded regeneration does not reduce artifact prevalence; when all versions fail, best-of selection still delivers flawed images.
Strengths
Exemplary statistical hygiene for a prototype paper
Wilson 95% CIs on all proportions, exact McNemar tests on 400 paired customer-product comparisons (all p<0.001), Fisher exact with Holm-Bonferroni for product contrasts, generalized Fleiss' κ (0.548/0.647), and customer-level clustered sensitivity checks (Table 3, Sec. 5.2).
Clean decomposition of judge versus generator
The Ifirst discrimination analysis (AR 0.32 vs 0.69 OpenAI; 0.03 vs 0.29 Gemini; rAR separation weak or absent at p=0.007 and p=0.589) shows both LLM judges act mainly as text-defect detectors — an insight that generalizes beyond this pipeline (Sec. 5.2, 6.1).
Useful negative results
Theory-grounded, transparent evaluation design
Honest scoping of claims
Weaknesses
No baseline of any kind
Core claim tested only qualitatively, with author overlap
Single-run results
Label reliability collapses in headline cells
Reproducibility currently aspirational
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ADMAN (Sec. 3) is a design-science prototype: a three-agent pipeline — prompt generator (A1: gpt-5 or gemini-3.1-pro-preview), image generator (A2: gpt-image-1 or gemini-3-pro-image-preview), and judge (A3, same LLM as A1) — that converts tabular customer attributes into a deterministic, template-based textual persona, generates an advertisement image conditioned on a product reference photo, and applies bounded regeneration (version limit three) with judge-selected best-of fallback. Claims (Sec. 1, 7): (1) fully automated, individual-level visual personalization is technically feasible (RQ1); (2) LLM-based judge supervision improves quality control, with quantified limits (RQ2); (3) performance depends strongly on product complexity and model configuration. The authors explicitly scope the study to technical feasibility, not downstream customer effects.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the AdMan pipeline?
AdMan is a research prototype that automatically creates personalized advertisement images: it converts a customer's profile data (demographics, hobbies, lifestyle, appearance) into a textual persona, uses one AI agent to write an image prompt, a text-to-image model to render the ad using a product reference photo, and a third LLM-based judge agent to reject flawed images and regenerate up to three times, falling back to the best version if none passes.
Which AI models were compared and which performed better?
The paper compares an all-OpenAI configuration (gpt-5 + gpt-image-1) with an all-Google Gemini configuration. Gemini's image generator produced fewer human-labeled artifact defects (final artifact rate 0.20 vs 0.42), but Gemini's judge passed far fewer images (40% vs 79%), meaning 60% of Gemini deliveries were best-of fallbacks. Overall Gemini delivered cleaner images; neither judge detected non-text structural defects reliably.
What were the main problems with the AI-generated ads in this study?
Text on product packaging (e.g., Coca-Cola and Tide labels) was the dominant failure mode in both configurations, and laundry detergent scenes were hardest (final artifact rate 0.69 in both). The LLM judges missed physical/structural artifacts such as reversed car trunks and missing washing machine doors, and regenerating up to three times never reduced artifact rates.
Is the AdMan code or dataset publicly available?
Not yet. The paper states that code and dataset will be made publicly available upon publication; in the current preprint they are not accessible, and the prompts are described only generically.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure content (the actual advertisement images, the red-highlighted artifacts in Fig. 2, and the judge-reasoning panels in Figs. 4-5) is not assessable from the text extraction; visual-quality statements here rely on the paper's reported labels, κ values, and statistics.
- The abstract in the supplied metadata ends mid-word, but the complete abstract appears intact in the extracted paper text, so no content was lost there.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.