PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolving Creatives with Clicks: CTR-Driven Advertising image Editing via Attribute Disentanglement and Preference Alignment

Xu Zhang, Haohan Wang, Zheng Zhang, Yingjie Xia, Wei Feng, Honghe Zhu, Jingjing Lv, Changping Peng, Xuheng Qin, Zehua Zhang, Zheng Luo, Qicheng Li

arXiv:2610.13066v1cs.CVcs.AI

TL;DR

Existing ad-image optimization often emphasizes aesthetics, while CTR-oriented prompt methods can penalize all attributes of an underperforming image, including beneficial ones. ECC edits existing creatives using CTR-sensitive attributes and real click preferences, with CPO promoting consistency across shuffled equivalent suggestions. Online experiments report the highest aggregate CTR gain across diverse product categories and advertising scenarios.

  • Problem

    Aesthetic improvements in ad images rarely translate reliably into CTR gains, and whole-image preference labels can penalize click-beneficial attributes along with harmful ones.

    Source: Introduction · p. 1

  • Method

    ECC identifies CTR-sensitive attributes, generates executable editing suggestions for existing creatives, refines them with real click preferences, and uses CPO to align shuffled equivalent suggestions.

    Source: Introduction · p. 2

  • Results

    ECC achieves a 6.86% CTR gain, the highest aggregate gain across diverse product categories and advertising scenarios, outperforming strong MLLM-based baselines.

    Source: Main Online Results · p. 5, Conclusion · pp. 7–8

  • Takeaways & Limitations

    The online experiments demonstrate the effectiveness and practical value of fine-grained attribute editing for creative optimization.

    Source: Conclusion · p. 8

Abstract

from arXiv · show

The optimization objective for e-commerce advertising creatives has shifted from aesthetic quality to online business metrics like click-through rate (CTR). Recent methods generate appealing ad images from product images by constructing CTR-based reward signals and refining prompt generation via Reinforcement Learning. Such paradigms equally penalize all attributes within an underperforming image, even though some are actually beneficial to clicks. In this work, we propose Evolving Creatives with Clicks (ECC), which reformulates creative optimization as fine-grained editing rather than generation from scratch. ECC selectively refines attributes that hurt CTR while preserving those that contribute positively. Specifically, faced with a rich attribute space across massive images, ECC first identifies 28 attributes significantly correlated with CTR via ANOVA and Random Forest regression. Based on attribute contrasts between images with disparate CTRs, we instruction-tune a prompt model to generate editing suggestions, then refine it with real click preferences since reference-derived suggestions are not uniformly beneficial. Furthermore, we introduce Consistent Preference Optimization (CPO) as ECC's preference optimization strategy, which enforces consistency across semantically equivalent shuffled suggestions to mitigate DPO overfitting and improve preference generalization. Extensive experiments demonstrate that ECC significantly improves the CTR of ad images, outperforming state-of-the-art models by a substantial margin. The complete data pipeline, including attribute extraction and analysis, the resulting dataset, and source code will be released upon acceptance.

Introduction

The introduction argues that CTR-oriented ad optimization should edit existing creatives selectively rather than penalize every feature of an underperforming image. ECC uses CTR-sensitive attributes, editing suggestions, and real click preferences to guide refinement, with CPO addressing preference overfitting.

  • ECC reframes creative optimization as fine-grained editing of existing ad images, refining attributes that hurt CTR while preserving those that help.Source: Introduction · p. 2
  • ECC identifies 28 CTR-sensitive attributes from 51 dimensions using MLLM extraction and statistical analysis, then uses them to structure actionable editing supervision.Source: Introduction · p. 2
  • Instruction tuning teaches executable suggestions from attribute-level contrasts, and real exposure and click feedback further refines them because reference-derived suggestions are not always beneficial.Source: Introduction · p. 2
  • CPO adds consistency between chosen suggestions and semantics-preserving shuffled variants to mitigate DPO overfitting without expanding the preference dataset.Source: Introduction · p. 2
  • Large-scale online experiments report that ECC achieves the highest CTR gain among strong MLLM-based baselines.Source: Introduction · p. 2

Related Work

Prior ad-image optimization often targets visual quality or selection rather than explicit CTR-aware editing. ECC combines CTR-sensitive visual-factor analysis with editing suggestions guided by online click feedback.

E-commerce ad image Optimization

  • Earlier ad-image work improves visual appeal, composition, product presentation, or editing flexibility, while CTR-driven ranking selects existing creatives and generation can limit foreground diversity.ECC instead edits existing creatives using real CTR feedback.Source: E-commerce ad image Optimization · p. 2

CTR Attribution and Visual Factor Analysis

  • CTR prediction models can incorporate visual information but often provide limited insight into which visual factors drive CTR differences; ANOVA and Random Forest support attribute significance and importance analysis.Source: CTR Attribution and Visual Factor… · p. 2

Multimodal Instruction-Guided Image Editing

Methods

ECC mines CTR-sensitive visual attributes, trains a prompt model to propose executable edits, and aligns those suggestions with online click preferences using DPO and CPO.

Methods

  • ECC’s workflow turns online ad attributes into instruction-tuning data, deploys generated edits to collect click feedback, and uses that feedback to construct preference data.Source: Methods · p. 3

CTR-Aware Visual Attribute Mining

  • 28 CTR-sensitive attributes are retained from 37 candidates using CTR associations, Random Forest contributions, reliability, redundancy, and editability.The process begins with 51 attributes and excludes 14 before statistical analysis; ANOVA uses within-product normalized CTR, while Random Forest estimates contributions under attribute interactions.Source: CTR-Aware Visual Attribute Mining · pp. 3–4

Instruction-Guided Editing Learning

  • Instruction tuning teaches the model to produce executable editing suggestions, while random sampling of 1–3 suggested operations diversifies training targets for later online preference refinement.The model is given a source ad image and instruction, and learns with the standard autoregressive loss.Source: Instruction-Guided Editing Learning · p. 4

Preference Optimization

  • Online exposure and click feedback form preference pairs for DPO, which aligns editing suggestions according to their post-editing CTR.Source: Preference Optimization · pp. 4–5
  • CPO adds a consistency regularizer so chosen suggestions and semantics-preserving shuffled variants receive consistent preferences, addressing DPO overfitting to superficial ordering.The method retains the DPO objective and explicitly encourages attention to editing semantics rather than operation order.Source: Preference Optimization · p. 5

Datasets

ECC builds separate datasets for learning executable edits and aligning them with real click preferences.

  • The instruction-tuning dataset teaches actionable modifications, while the online preference dataset aligns suggestions with real user click preferences.Source: Datasets · p. 5
  • 101,699 instruction-tuning samples pair lower-CTR creatives with editing suggestions generated from contrasts against higher-CTR references using 28 selected attributes.References and extracted attribute values are used only during data construction, not as model input.Source: Datasets · p. 5
  • 68,704 preference pairs compare suggestions for the same source image and instruction, with higher- and lower-post-editing-CTR suggestions labeled chosen and rejected.Pairs are collected from deployed edits using real exposure and click feedback.Source: Datasets · p. 5

Experiments

ECC’s experiments assess online CTR, cross-category performance, aesthetic quality, preference order robustness, attribute relevance, and representative edits. Its main online result is a 6.86% CTR gain, with additional evaluations examining where and how the approach performs.

Experimental Setup

Online Results

  • ECC achieves 6.86% CTR gain, versus 0.20% for CAIG, and exceeds GPT and Gemini by 3.61 and 1.34 percentage points, respectively.Under the current traffic scale, ECC is expected to generate approximately RMB 117 million in additional annual revenue.Source: Main Online Results · p. 5, Table 1 · p. 6
  • ECC improves CTR across all evaluated product categories and placements, with particularly strong gains in Home Appliances, Computers, Beauty, the Order Page, and the Product Detail Page.Source: Analysis Across Product Categories and… · p. 6, Table 2 · p. 6

Aesthetic Evaluation

  • ECC achieves the best overall ranking across four aesthetic evaluators, but evaluator rankings are inconsistent and do not fully align with online CTR gains.Lower overall ranking scores indicate better performance.Source: Aesthetic Evaluation · p. 6, Table 3 · p. 6

Order Robustness Analysis

Validation of CTR-Sensitive Visual Attributes

  • 59.56% accuracy is achieved with prompts describing the selected 28 attributes, the highest among the tested input settings.The pairwise task predicts which of two same-product images has higher CTR; the selected attributes retain CTR-relevant information while excluding redundant or unstable factors.Source: Validation of CTR-Sensitive Visual… · pp. 6–7, Table 5 · p. 7

Ablation Study

Qualitative Analysis

  • Representative edits show cleaner layouts, more suitable backgrounds, clearer product presentation, and more focused copy in optimized images.Figure 4 places original images on the left and optimized images on the right; the edited images have higher online CTR.Source: Qualitative Analysis · p. 7, Figure 4 · p. 7

Conclusion

ECC refines existing advertising images using CTR-sensitive attributes and real click preferences. Online experiments report the highest aggregate CTR gain across diverse product categories and advertising scenarios, supporting fine-grained attribute editing for creative optimization.

  • ECC improves existing ad images through fine-grained editing rather than generating images from scratch.Source: Conclusion · p. 7
  • ECC identifies 28 CTR-sensitive attributes from 51 predefined dimensions and uses them to construct actionable editing supervision.Source: Conclusion · p. 7
  • ECC learns executable editing suggestions through instruction tuning and refines them with preference optimization based on real click feedback.Source: Conclusion · p. 7
  • CPO adds a consistency regularizer to DPO to mitigate overfitting to spurious ordering patterns and improve preference generalization.Source: Conclusion · p. 7
  • Online experiments show ECC outperforms strong MLLM-based baselines and achieves the highest aggregate CTR gain across diverse product categories and advertising scenarios.The paper identifies fine-grained attribute editing as effective and practically valuable for creative optimization.Source: Conclusion · pp. 7–8

Overview

The supplementary material outlines CTR-aware attribute mining, the visual schema, and preliminary screening. It describes a 51-attribute schema and the criteria used to narrow attributes for analysis.

Overview

  • The supplement presents the mining process, representative instruction-tuning examples, and additional qualitative editing results across products and scenarios.The examples illustrate executable suggestions constructed from same-product creative pairs.Source: Overview · p. 10

A. CTR-Aware Visual Attribute Mining

  • The pipeline extracts 51 visual attributes and identifies 28 CTR-sensitive attributes after statistical analysis and consideration of reliability, redundancy, and editability.The attributes are extracted by Qwen3-VL-32B.Source: A. CTR-Aware Visual Attribute Mining · p. 10

A.1 Visual Attribute Schema

  • The hierarchical schema covers six ad-image areas and assigns candidate labels to reduce semantic fragmentation during large-scale extraction.Most attributes use predefined Strict labels; Extensible attributes permit additional values when predefined candidates are insufficient.Source: §A.1 · p. 10

A.2 Preliminary Attribute Screening

  • Before statistical analysis, 14 attributes are excluded for unreliable attribution, including unstable extraction, subjective labels, imbalance, redundancy, or insufficient samples.Table 8 summarizes the pre-excluded attributes and rationales.Source: §A.2 · p. 12, Table 8 · p. 12

A.3 Statistical Analysis and Final Selection

The analysis combines one-way ANOVA and Random Forest contribution to select CTR-sensitive attributes, with further adjustments for reliability, redundancy, and editability.

A.3 Statistical Analysis and Final Selection

  • ANOVA tests each attribute’s association with within-product normalized CTR, while Random Forest estimates contribution in a multivariate setting that captures nonlinear patterns and interactions.Attributes are initially selected when ANOVA is significant (p < 0.05) or normalized Random Forest contribution exceeds 5.0%.Source: §A.3 · p. 12, Analysis of Variance · p. 12, Random Forest Regression · p. 13
  • The mining procedure narrows 51 initial attributes to 28 CTR-sensitive attributes, with 37 candidates remaining after preliminary exclusion.The statistical analysis table covers candidate attributes after preliminary screening.Source: Analysis of Variance · p. 13, Table 9 · p. 14
  • Final selection also accounts for distribution reliability, semantic redundancy, and practical editability, including manual adjustments to the joint statistical criterion.Product Quantity is retained for high Random Forest contribution and actionable edits; other attributes are removed for redundancy, unstable distributions, weak contribution, or overlapping semantics.Source: Random Forest Regression · p. 14

B. Instruction-Tuning Dataset Examples

The instruction-tuning examples pair low-CTR creatives with executable editing suggestions, using same-product high-CTR references only in constructing and validating the suggestions.

B. Instruction-Tuning Dataset Examples

  • Figure 5 shows low-CTR inputs, editing-suggestion targets, and same-product high-CTR references used only during suggestion construction and validation.Colored phrases mark the CTR-sensitive attributes involved in each editing operation; the reference is not provided to the model during training.Source: Figure 5 · p. 15, B. Instruction-Tuning Dataset Examples · p. 15

C. Additional Qualitative Results

Additional qualitative examples show CTR-driven edits across several creative elements while retaining advertised-product identity and core content.

C. Additional Qualitative Results

  • Figure 6 examples cover background, layout, product presentation, visual style, and copywriting, with clearer organization and improved online CTR reported.The resulting creatives preserve the advertised products’ identity and core content.Source: C. Additional Qualitative Results · p. 16, Figure 6 · p. 16
Loading 2610.13066v1…