I Am AdMan: A Pipeline for Automatic Generation of Personalized Advertising Imagery

1 Introduction

• Existing personalization largely matches products to customers, while advertisement presentation remains generic and detailed-profile-to-image generation at scale is underexplored.Prior image-generation work often focuses on inpainting or uses only a few customer-data dimensions.Source: §1 · pp. 1–2
• The study implements an autonomous pipeline and evaluates OpenAI GPT and Google Gemini configurations using 1,745 images from 100 real customers and six celebrities.It combines expert qualitative evaluation with quantitative analysis and focuses on technical feasibility rather than direct customer-engagement effects.Source: §1 · p. 2
• The pipeline produces personalized scenes, but quality varies by product complexity and model configuration, with product text, complex interactions, and spatial arrangements remaining challenging.Text rendering is a persistent artifact source; judge behavior also differs between the Google Gemini and OpenAI GPT configurations.Source: §1 · p. 2

2 Background & related work

• The paper positions its work at the intersection of marketing personalization, generative AI for personalized content, and multi-agent systems for scalable creative workflows.Source: §2 · p. 2

2.1 Generative artificial intelligence

• Generative AI creates novel content from learned patterns, while LLMs generate text and diffusion models synthesize images from prompts.These capabilities provide a technological basis for pipelines that turn structured creative instructions into visuals.Source: §2.1 · p. 2

2.2 Multi-agent systems

• Role-specialized LLM agents collaborate, critique, and refine outputs, but prior work has not applied such systems to detailed-persona-driven ad imagery at scale.The passage describes this as a gap addressed by the proposed pipeline.Source: §2.2 · p. 3

2.3 Personalization in marketing

• Marketing personalization adjusts the marketing mix to customer profiles, while personalized advertising specifically adjusts advertisement content to customer tastes or recognizability.Source: §2.3 · p. 3

2.4 AI in personalized marketing

• Prior personalized image-generation approaches often use inpainting or limited customer preferences, rather than systematically using detailed profiles such as demographics, values, and lifestyle [29].Pigeon generates entire advertisement imagery but is limited to historic interactions and movie posters.Source: §2.4 · p. 4

3 ADMAN: A pipeline for automatic generation of personalized advertising imagery

• ADMAN organizes fully automated personalized-ad generation into data generation, image generation, image evaluation, and user interaction.Source: §3 · p. 4
• The data-generation stage turns customer attributes into a textual persona, which, together with a product reference image, feeds the next stage.The described attributes include demographics, interests, and behavioral signals.Source: §3 · p. 4
• A prompt-generation agent creates persona-aligned scene and style instructions, and an image-generator agent uses them with the product reference image.The generated ad is intended to reflect the persona while presenting the product consistently with its reference.Source: §3 · p. 5
• A judge agent checks for minimum quality and obvious artifacts, with the prototype using fixed-limit regeneration and best-of selection if no version passes.Source: §3 · p. 5
• The pipeline serves an advertisement to the customer, selecting the best available generated candidate when none passes the judge.Source: §3 · p. 5

4 Experimental setup

• ADMAN was evaluated qualitatively using the OpenAI GPT configuration with celebrity customer data, then quantitatively using both OpenAI GPT and Google Gemini configurations on 100 real individuals, assessing image artifacts and pass rate.Source: §4 · p. 5

4.1 Pipeline implementation

• ADMAN maps tabular customer attributes to personas and prompts, generates images with product references, and uses a judge for bounded regeneration; four products span FCB Grid quadrants.Source: §4.1 · p. 6

4.2 Data collection

• Evaluation profiles comprised six celebrities, whose identities ADMAN did not receive, and 100 Prolific participants surveyed across customer-data variables.Source: §4.2 · p. 7

4.3 Evaluation

• The evaluation combined expert focus-group feedback with quantitative artifact-rate assessment, using identical generation settings across products and personas.Source: §4.3 · p. 7

4.3.1 Qualitative Evaluation

• Five non-author experts reviewed 24 judge-selected final advertisements generated with OpenAI GPT, alongside the personas, to assess feasibility and recognizable personalization.The focus group included eight AI researchers total; three author-team members moderated and provided methodological context.Source: §4.3.1 · p. 7

4.3.2 Quantitative Evaluation

• Pass, artifact, and relaxed artifact rates assess judge approvals and human-labeled defects across first, judge-passed, and delivered image sets.Three reviewers label each image independently, with majority vote determining artifacts; Wilson intervals and paired McNemar tests support selected comparisons.Source: §4.3.2 · pp. 7–9

5.1 Qualitative results

• Experts praised photographic realism, coherent persona-aligned scenes, and contextual accuracy, while finding compelling examples across all four product categories.Source: §5.1 · p. 9, Figure 3 · p. 10
• Experts identified structural defects, awkward anatomy, implausible scenes, persona mismatches, compositional anomalies, and occasional gender- and age-stereotyped outputs.Source: §5.1 · pp. 9–10

5.2 Quantitative results

• AR = 0.69 for final detergent images under both configurations, while Gemini’s final AR for the electric vehicle, watch, and soft drink was 0.05, 0.04, and 0.03, respectively.Source: §5.2 · p. 11
• AR = 0.56 versus rAR = 0.01 for OpenAI GPT soft-drink first attempts, illustrating how product-text defects can dominate artifact counts.Source: §5.2 · p. 12
• Final PR was 0.79 for GPT and 0.40 for Gemini, so many deliveries were best-of fallbacks rather than judge-approved images.Source: §5.2 · p. 12
• Judges’ pass decisions aligned with human artifact labels mainly through text defects, with weak evidence of detecting structural artifacts.Source: §5.2 · pp. 12–13
• ∆AR = −0.22 (95% CI [−0.28, −0.17]) and ∆PR = −0.39 ([−0.44, −0.34]) for Gemini minus GPT, indicating fewer artifacts but fewer passes.Source: §5.2 · p. 12, Table 3 · p. 13
• AR rose across repeated versions, from 0.50 to 0.56 to 0.73 for GPT and from 0.23 to 0.28 to 0.29 for Gemini; later samples were increasingly selected toward harder pairs.Source: §5.2 · p. 13

6 Discussion

• The fully automated pipeline generates personalized advertising images at scale, but text rendering, product fidelity, and automated quality control remain bottlenecks to commercial readiness.These bottlenecks persist across the two model configurations, although artifact prevalence, judge behavior, and product-specific failure modes vary.Source: §6 · p. 13

6.1 Interpretation of findings

• Persona attributes visibly shape outputs, while increasing scene complexity degrades spatial coherence, physical realism, object permanence, and style control.The discussion locates the main constraint in image generation rather than persona-to-prompt translation.Source: §6.1 · pp. 13, 15
• Product and persona complexity interact, making consistent quality from a universal, product-agnostic pipeline unlikely.Products with complex packaging, text, or spatial arrangements impose constraints that can compound with detailed personas.Source: §6.1 · p. 15
• Repeated generation does not reliably resolve text errors, which persist across both configurations and may require separating text rendering from image generation.The discussion describes text rendering as a structural limitation of current image models.Source: §6.1 · p. 15
• Both judges largely detect text defects rather than structural implausibility; GPT accepts artifact-bearing images, while Gemini rejects many images reviewers deem clean.The discussion reports that neither judge reliably separates images with and without non-text artifacts.Source: §6.1 · p. 15
• Generated products are often recognizable but can depart from reference details, and recognizable celebrity faces raise rights-sensitive concerns in advertising.Exact product appearance and brand consistency may matter even when deviations do not make an image unusable.Source: §6.1 · pp. 15–16

6.2 Implications for research and practice

• Research should pursue product- and model-aware generation and judge calibration, while practitioners should assess generation quality separately and retain human review for complex, text-heavy products.The discussion also proposes targeted, feedback-driven refinement or inpainting as alternatives to restarting the full generation cycle.Source: §6.2 · p. 16

6.3 Limitations & future work

• Conclusions are bounded by two pipeline instantiations, a predominantly U.S.-English persona sample, four products, and no customer-outcome or deployment-cost evaluation.The study also does not measure product fidelity as a separate quantitative construct or assess judge calibration and repeated-evaluation consistency.Source: §6.3 · pp. 16–17

• The pipeline produces advertisements whose scenes reflect persona attributes, demonstrating the technical feasibility of individualized visual presentation.Source: §7 · p. 17
• Image-generation performance differs substantially between the OpenAI GPT and Google Gemini configurations.Source: §7 · p. 17
• The LLM judge improves output quality by filtering defective images, but its effectiveness varies by product complexity, artifact type, and model configuration.Source: §7 · p. 17
• When no candidate passes, best-of selection remains limited by the quality of the generated candidate pool, and complex packaging, prominent text, or strict brand-fidelity needs still warrant human oversight.Source: §7 · p. 17
• The findings motivate product-specific, model-aware generation, feedback-driven refinement, calibrated judges, and hybrid evaluation approaches.Source: §7 · p. 17