The Image Looks Good. But Whose Taste Is It?

TL;DR for operators

A retailer can generate an attractive product image that has little connection to the customer viewing it. A creator can receive a polished illustration that looks unlike anything they would normally publish. In both cases, conventional image-quality scores may miss the personalization failure.

Research on personalized image generation from user histories introduces two benchmarks and a system called Pearl that reasons over historical activity, generates an image, and inspects the result before revising its prompt.¹ Pearl improves several offline personalization measures, but does not consistently produce the most aesthetically appealing or visually similar images.

For content operations, three implications follow:

• Evaluate personal relevance independently of image quality. A visually attractive image need not reflect the intended user’s history.
• Inspect the rendered output, not just the generation prompt. Intended personalization details can disappear during rendering.
• Treat feedback optimization as conditional improvement. Some metrics improve while others regress, and the study does not establish higher engagement or conversion.

The operational opportunity is a more testable personalization workflow, not a demonstrated increase in marketing effectiveness.

A better personalization score does not guarantee a prettier image

Consider two customers viewing the same product. Their purchase histories and reviews may suggest different activities, surroundings, and visual preferences. A personalized image-generation system should preserve the product while presenting it in a context that fits each customer’s history.

The difficulty is determining whether it succeeded.

In the paper’s social-media benchmark, Pearl achieves an inter-category user-retrieval Recall@1 of 0.876, compared with 0.748 for Pigeon, a competing personalization method. Yet PMG, another baseline, receives an aesthetic score of 5.84, higher than Pearl’s 5.45 (Section 4.3, Table 2).

The retrieval test asks whether an evaluator can associate a generated image with the correct user’s history among competing histories. It measures recognizable user-specific information, not whether the picture is attractive.

Source: Section 4.3, Table 2. Higher values are better for these measures.

The 0.128 absolute improvement in inter-category Recall@1 is meaningful within the benchmark’s retrieval setting. It does not mean that 87.6% of target users would prefer Pearl’s images.

This difference changes how personalization systems should be selected. A marketing team choosing the most attractive output may overlook user alignment. A team optimizing retrieval scores alone may overlook visual quality or preservation of the advertised product.

PMH-IG tests what a model learns from ordinary user histories

The paper’s first contribution is PMH-IG, a benchmark that makes historical user activity the basis of personalization.

Traditional image personalization often starts with curated reference pictures of a subject or visual concept. PMH-IG instead asks what can be inferred from distributed evidence in reviews, photographs, captions, and related activity.

It contains two complementary tasks.

Personalized Scene Generation uses Amazon review histories to infer lifestyle preferences and generate scenes around a specified product. The dataset includes 5,000 users and 50,000 reviews across six product categories, with 518 test users.

Personalized Creative Generation uses Instagram posting histories to generate new imagery consistent with a creator’s established aesthetic, given a requested topic. Its dataset includes 4,413 users and 50,656 posts, with 207 test users and 314 test user-target pairs.

Both tasks use user-disjoint training and evaluation splits (Appendix B, Table 3).

The two tasks separate different requirements. In e-commerce, product preservation is central: the model should personalize the surrounding scene without misrepresenting what is being presented. In creative generation, consistency with historical visual preferences becomes more prominent.

The benchmark also separates general aesthetics, image similarity, multimodal assessments of content and style, and retrieval-based personalization. This measurement design is itself a contribution because these criteria need not rank systems identically.

There is an important asymmetry: scene generation lacks naturally occurring ground-truth personalized scene images. Its personalization assessment therefore depends on indirect retrieval and model-judge measurements, rather than direct comparison with an image the customer actually wanted.

Pearl checks whether the generated image expresses the intended preferences

Pearl addresses a practical failure that occurs between planning an image and rendering it.

A model can correctly infer preferences from a user’s history and describe an appropriate scene. The image generator may still omit relevant objects, emphasize inconsistent stylistic elements, or render the required content incorrectly.

Pearl separates the workflow into three functional stages:

1. Plan: A Qwen2.5-VL-7B multimodal planner analyzes user history and the generation condition, producing an explicit preference-oriented reasoning trace and a self-contained prompt.
2. Render: A frozen SDXL-1.0 image generator produces an initial image.
3. Reflect and rerender: A second Qwen2.5-VL-7B policy examines the rendered image against the history and generation requirement, revises the prompt, and sends it through the same generator.

The reflector does not edit pixels directly. It changes the instructions used to produce the next image.

This architecture leaves the image generator fixed. The adaptable components are the policies responsible for preference interpretation and prompt correction.

For a production system, that separation could make failures easier to diagnose. An inappropriate image might originate from an incorrect preference inference, a prompt that inadequately expresses the preference, or a renderer that fails to follow the prompt. These are different failure modes and suggest different corrective actions.

Pearl provides a framework for examining them, although the paper does not establish that each failure can always be identified correctly.

Training against rendered images helps, but not on every measure

Pearl’s training procedure introduces another distinction: learning to imitate a desirable generation plan is different from learning which plans actually produce better images.

The authors first use supervised fine-tuning on teacher-generated trajectories. Gemini 2.5 Flash supplies these training examples with privileged access to held-out identity signals. Those privileged signals are part of the training arrangement, not information assumed available to Pearl at inference.

The second stage applies render-in-the-loop Direct Preference Optimization (DPO). It samples four candidate trajectories, renders their outcomes, and forms preference comparisons using task-specific rewards. Scene-generation rewards include recommendation ranking, identity-element coverage, and multimodal judgments; creative-generation rewards include CLIP and DINO image similarities.

Planner and reflector updates alternate while the other policy and the image generator remain fixed. The study evaluates one planner update followed by one reflector update, not repeated autonomous improvement.

The ablation results are particularly informative:

Source: Appendix E.1, Tables 5–6.

Full Pearl improves several measurements, especially creative user retrieval and scene style alignment. But scene Hit@5 declines, as does the creative content-judge score. Other measurements remain unchanged or show additional regressions.

Crucially, the ablated variant is labeled Pearl-Reflection, but the described intervention replaces preference optimization with supervised-only training. It does not cleanly remove the reflector from the architecture.

The experiment therefore measures the incremental effect of the preference-optimization procedure within the tested pipeline. It cannot establish how much of the overall advantage comes independently from reflection.

That distinction matters when assessing the operational cost of an additional rendering and inspection stage. The benchmark supports the complete approach’s comparative performance, but does not provide a component-isolated estimate of reflection’s benefit or a systematic compute-performance analysis.

Offline personalization is not yet commercial validation

Across the principal baseline comparisons, Pearl leads all six scene-generation measures and seven of ten creative-generation measures (Section 4.3, Tables 1–2). The authors report approximately 15% average improvement across personalization metrics, although individual gains vary.

A preliminary human evaluation also favors Pearl’s history fit in 19 of 25 scene comparisons against PMG and 16 of 25 creative comparisons against Pigeon.

These results strengthen the case for history-conditioned personalization, but their interpretation has limits.

Only one annotator assessed the 50 human-evaluation cases. The creative comparisons also used different history subsets for the two methods. Neither task measured preference judgments from the actual target users, and the main benchmark results lack reported uncertainty intervals. The source additionally identifies one scene account with possible target-review leakage in a separate rerendering test, leaving its original-output status unresolved.

Most importantly, no online experiment measures click-through rates, engagement, purchasing behavior, or customer satisfaction.

For a retailer, an appropriate next evaluation would therefore combine product-fidelity checks, history-alignment measurements, visual-quality thresholds, and actual customer-response experiments. A creator tool would similarly need to test whether creators themselves accept the generated aesthetic, rather than relying exclusively on retrieval probes or external judges.

These are Cognaptus operational inferences from the study, not deployment outcomes demonstrated by the authors.

Personalization needs its own acceptance criteria

The central contribution of this work is a more explicit definition of what successful personalized image generation should demonstrate.

PMH-IG establishes that user alignment can be evaluated separately from attractiveness and visual similarity. Pearl shows how preference inference, rendering, and post-render inspection can be organized into a single generation pipeline. Its training experiments show that optimizing against rendered outcomes can improve selected measures, without ensuring universal gains.

For teams building personalized advertising, e-commerce imagery, or creator tools, this suggests a concrete design principle: judge an image against the user’s history, the requested content, and visual-quality requirements independently.

An additional reflection stage may help enforce those requirements, but its incremental value, resource cost, and effect on actual users still require direct testing.

A generated image can look excellent and still belong to nobody in particular.

Cognaptus: Automate the Present, Incubate the Future.