Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/12/2026, 6:50:26 AM
Summary
The paper introduces Evolving Creatives with Clicks (ECC), a framework for optimizing e-commerce advertising images by fine-grained editing rather than generation from scratch. ECC identifies 28 CTR-sensitive visual attributes using ANOVA and Random Forest regression on 550k images. It employs a two-stage learning process: instruction tuning to generate editing suggestions based on attribute contrasts, and preference optimization using real click feedback. A novel Consistent Preference Optimization (CPO) strategy is proposed to mitigate DPO overfitting by enforcing consistency across shuffled editing suggestions.
Entities (13)
Relation Signals (8)
ECC → optimizes → CTR
confidence 98% · ECC significantly improves the CTR of ad images
ECC → uses → CPO
confidence 95% · we introduce Consistent Preference Optimization (CPO) as ECC's preference optimization strategy
ECC → uses → Random Forest
confidence 92% · ECC first identifies 28 attributes significantly correlated with CTR via ... Random Forest regression.
ECC → uses → ANOVA
confidence 92% · ECC first identifies 28 attributes significantly correlated with CTR via ANOVA
CPO → mitigates → DPO overfitting
confidence 90% · CPO ... enforces consistency ... to mitigate DPO overfitting
Qwen3-VL-32B → usedfor → Attribute Extraction
confidence 88% · use Qwen3-VL-32B ... to extract 51 predefined visual attributes
Qwen3-VL-8B-Instruct → usedfor → Editing Suggestions
confidence 87% · We use Qwen3-VL-8B-Instruct as the base multimodal model for generating editing suggestions.
GPT-5.5 → usedfor → Suggestion Generation
confidence 85% · GPT-5.5 ... compares the two creatives and generates executable editing suggestions
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2610.13066v1
- Canonical: https://arxiv.org/abs/2610.13066v1
Trouble viewing inline? Open PDF directly →
Full Text
73,326 characters extracted from source content.
Expand or collapse full text
Evolving Creatives with Clicks: CTR-Driven Advertising image Editing via Attribute Disentanglement and Preference Alignment Xu Zhang 1,2∗† , Haohan Wang 2† , Zheng Zhang 2‡ , Yingjie Xia 3 , Wei Feng 2 , Honghe Zhu 2 Jingjing Lv 2 , Changping Peng 2 , Xuheng Qin 2 , Zehua Zhang 2 , Zheng Luo 2 , Qicheng Li 1‡ 1 TMCC, College of Computer Science, Nankai University, 2 JD.COM, 3 VCIP, Computer Science, Nankai University xu_zhang@mail.nankai.edu.cn Abstract The optimization objective for e-commerce advertising cre- atives has shifted from aesthetic quality to online business metrics like click-through rate (CTR). Recent methods gener- ate 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 gen- eration from scratch. ECC selectively refines attributes that hurt CTR while preserving those that contribute positively. Specifically, faced with a rich attribute space across mas- sive images, ECC first identifies 28 attributes significantly correlated with CTR via ANOVA and Random Forest regres- sion. Based on attribute contrasts between images with dis- parate 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 benefi- cial. Furthermore, we introduce Consistent Preference Opti- mization (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, out- performing 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 Advertising creatives, especially advertising images (ad im- ages), serve as a key lever to capture user clicks (Mishra et al. 2020; Wang et al. 2021; Chen et al. 2021). Early works primarily focused on enhancing the aesthetics or controlla- bility of generated ad images (Li et al. 2023; Du et al. 2024; Chen et al. 2024, 2025a). However, such improvements rarely translate into click-through rate (CTR) gains, as the criteria correlate loosely with online business performance. ∗ Work done while interning at JD.com. † These authors contributed equally. ‡ Corresponding author. CTR: 0.050 Product Size: Large TrainInference Has_model: Yes Layout: Right-Left CTR: 0.020 CTR: 0.030 Has_model: No Positive Sample CTR: 0.040 Product Size: Large Has_model: No Layout: Product-Only Layout: Product-Only Negative Sample CTR: 0.020 Product Size: Small Layout: Product-Only Layout: Left-Right Negative Sample CTR: 0.010 Product_size: Large Layout: Left-Right Has_model: Yes Negative Sample CTR: 0.020 Layout: Right-Left Has_model: No Positive Sample CTR: 0.050 Has_model: No Positive Sample CTR: 0.025 Product Size: Small Has_model: Yes Layout: Right-Left Positive Sample Negative Sample CTR-Positive Attributes CTR-Negative Attributes Has_model: No Product Size: Small Product Size: Small Product Size: Large Training Figure 1: Comparison of Existing Advertising Creative Op- timization Paradigms. Recent advances directly adopt CTR as the optimization objective (Yang et al. 2024; Chen et al. 2025b). The most prominent paradigm leverages CTR-derived rewards to fine- tune a prompt model via Direct Preference Optimization (DPO) (Rafailov et al. 2023), thereby generating prompts that yield better-performing images. Despite achieving cer- tain CTR improvements, this paradigm has three major lim- itations. First, as shown in Fig. 1(a), DPO treats prompts de- scribing the entire ad image as positive or negative samples, causing all visual attributes in an underperforming image to be penalized equally, even though some may contribute posi- tively to CTR. Such erroneous signals are difficult to recover because promising visual ideas may appear only sparsely in the vast design space. Second, ad images in practice originate from diverse production pipelines, while a unified framework that generates prompts and renders images from product im- ages can hardly preserve or reproduce existing high-quality creatives. Finally, DPO is prone to overfitting when pref- erence supervision is limited (Lin et al. 2024; Chowdhury, Kini, and Natarajan 2024). Although expanding the pref- erence dataset can alleviate this issue, collecting additional samples from real online feedback requires substantial image arXiv:2610.13066v1 [cs.CV] 9 Oct 2026 generation, traffic, and experimentation time, making over- fitting reduction and data acquisition cost difficult to balance. To address the first two limitations, we propose Evolving Creatives with Clicks (ECC), which reformulates creative optimization as fine-grained editing of existing ad images rather than generation from scratch. As shown in Fig. 1(b), ECC selectively refines attributes that hurt CTR while pre- serving those that contribute positively. We define 51 visual attribute dimensions based on large-scale e-commerce data and expert knowledge, and provide candidate values for most attributes to reduce ambiguity and avoid overly dispersed descriptions. We then leverage MLLMs to extract attribute values from online ad images and identify 28 CTR-sensitive attributes through statistical analysis. These attributes trans- form the vast editing space into a structured set of actionable dimensions, enabling more targeted CTR optimization. Based on the selected attributes, ECC adopts a two-stage learning strategy. In the first stage, we construct image pairs with disparate CTRs and derive attribute-level editing schemes. Since high-CTR creatives may still contain sub- optimal attributes, the resulting suggestions are not always beneficial. We therefore randomly sample a subset of oper- ations for each training instance to train the prompt model to generate diverse executable editing suggestions. In the second stage, the instruction-tuned prompt model generates multiple suggestions for existing ad images, which are exe- cuted to produce new creatives. Since accurately predicting CTR from images remains challenging, with prior CTR re- ward models achieving only about 58% accuracy (Chen et al. 2025b), we deploy these creatives online rather than relying on a learned reward model. Preference pairs are then con- structed from real exposure and click feedback to distinguish effective suggestions and further refine the prompt model toward genuine CTR gains. To mitigate DPO overfitting without further expanding the preference dataset, we propose Consistent Preference Optimization (CPO) as the preference optimization strategy of ECC. Inspired by invariance learning in self-supervised learning (Chen et al. 2020), we identify that shuffling in- dependent operations does not affect the edited image, so the original and shuffled chosen editing suggestions should receive consistent preferences. CPO therefore introduces a lightweight consistency objective into DPO, encouraging the prompt model to focus on editing semantics rather than super- ficial ordering patterns and improving generalization without additional online data. Overall, our contributions are as fol- lows: • CTR-Aware Attribute Mining and Dataset Construc- tion Pipeline. We develop a unified pipeline for CTR- aware attribute mining and editing-suggestion data con- struction. By extracting 51 visual attributes from over 550,000 online ad images and identifying 28 CTR- sensitive dimensions through MLLM-based extraction and statistical analysis, we construct the first large-scale CTR-driven editing-suggestion datasets for instruction tuning and reinforcement learning. • ECC: The First CTR-Driven Advertising Image Edit- ing Framework. We propose ECC, the first framework to formulate ad image editing from a CTR-oriented per- spective. Guided by CTR-sensitive attributes, it learns executable editing suggestions through instruction fine- tuning and aligns them with online click preferences through reinforcement learning. • Consistent Preference Optimization Strategy. We pro- pose Consistent Preference Optimization (CPO), which augments DPO with consistency regularization between chosen responses and their semantics-preserving shuffled variants, addressing the overfitting problem in preference learning. Large-scale online experiments show that ECC achieves the highest CTR gain among strong MLLM- based baselines. Related Work E-commerce ad image Optimization Prior work on ad image optimization primarily focuses on visual quality and controllability. Methods such as Au- toCO (Chen et al. 2021), PAID (Chen et al. 2025a), retrieval- assisted advertising image generation (Ku et al. 2023), and instruction-guided editing models (Brooks, Holynski, and Efros 2023; Zhang et al. 2023; Huang et al. 2024; Hui et al. 2024) improve visual appeal, composition, product presen- tation, or editing flexibility, but do not explicitly model user behavior. CTR-driven methods leverage online interaction signals for creative refinement and ranking (Mishra et al. 2020; Wang et al. 2021; Lin et al. 2022) and generation (Wei et al. 2022; Yang et al. 2024; Chen et al. 2025b). Rank- ing methods only select existing creatives, while generation- based methods often reuse fixed product foregrounds, limit- ing foreground diversity and creative variation. ECC instead edits existing creatives using real CTR feedback, enabling more flexible and diverse optimization. CTR Attribution and Visual Factor Analysis CTR prediction methods such as DeepFM (Guo et al. 2017) and xDeepFM (Lian et al. 2018) model feature interactions, while later studies incorporate visual information to improve prediction accuracy (Ge et al. 2018). However, these models often lack interpretability and provide limited insight into which visual factors drive CTR differences. Visual factor analysis commonly relies on statistical signif- icance and feature importance. ANOVA (Fisher 1970) evalu- ates whether different attribute values lead to significant CTR differences, while Random Forest regression (Breiman 2001) captures nonlinear relationships and estimates attribute im- portance. Building on these methods, we identify CTR- sensitive visual attributes and use them to guide editing- oriented data construction. Multimodal Instruction-Guided Image Editing MLLMs have shown strong visual understanding and rea- soning capabilities. Recent MLLMs, including LLaVA (Liu et al. 2023), Qwen2-VL (Wang et al. 2024), Qwen2.5-VL (Bai et al. 2025b), GPT-5.5 (OpenAI 2026), and Gemini 3.1 Pro Preview (Google 2026), have substantially improved visual grounding and fine-grained image understanding. (a) Attribute Extraction with MLLM 550k Ads images background symmetry has_text has_model ... 51 Attributions (b) CTR-aware Statistical Analysis ANOVA Analysis for Each Attribute Random Forest Attribute Importance Analysis Root (All data) Background ... ... Layout ... ... ... . . . . . . . . . Choose attribute with max MSE decrease High Importance Low Importance MSE Decrease = MSE parent -MSE children (b.1) Comprehensive Statistical Analysis CTR: 0.015 Low CTR Creative CTR: 0.049 High CTR Creative GPT-5.5 Role: - Ad Editing Expert 1111 Input: - Constraint Prompts - Low-CTR Image - High-CTR Image - 28 Key Attributes 111111 Output: - Editing Suggestions Replace background: ... 1111 Adjust layout: ... 1111 Add text: ... 1111 Replace product: ... 1111 Remove model/limbs: ... 1111 Add decoration: ... 1111 Add props: ... (c) LLM-based Pairwise Editing Knowledge Construction CTR-Driven Suggestion Dataset "messages": [ "role": "user", "content": [<image>, <instruction>], "role": "assistant", "content": <Suggestion>] (d) Dataset Building "messages": [ "role": "user", "content": [<image>, <instruction>], "role": "assistant", "content": <chosen_Suggestion>], "rejected_response": <rejected_Suggestion> (b.2) Analysis Results Figure 2: CTR-aware visual attribute mining and dataset construction pipeline. (a) shows the extraction of 51 predefined visual attributes from over 550K ad images using Qwen3-VL-32B. (b) shows the joint ANOVA and Random Forest analysis for selecting 28 CTR-sensitive attributes. (c) shows executable editing suggestions generated by comparing same-product creatives with different CTRs under attribute-guided prompts. (d) shows the structures of the instruction-tuning and preference datasets. Instruction-guided editing methods (Brooks, Holynski, and Efros 2023; Zhang et al. 2023; Huang et al. 2024; Cao et al. 2024; Hui et al. 2024; Sheynin et al. 2024; Yu et al. 2025; Xu et al. 2025; Qian et al. 2026) enable flexible image modification, but mainly target generic visual edits without modeling CTR. ECC instead learns CTR-oriented editing suggestions from real creatives and online click feedback. Methods Our workflow begins by extracting visual attributes from large-scale online ad images and identifying CTR-sensitive dimensions through statistical analysis, which are then used to construct instruction-tuning data. The instruction-tuned model generates executable editing suggestions that are ap- plied to source images for online deployment, and the re- sulting exposure and click feedback are used to construct preference data. Finally, CPO introduces consistency reg- ularization over semantics-preserving shuffles to explicitly learn order-invariant preferences and mitigate overfitting. CTR-Aware Visual Attribute Mining As shown in Fig. 2(a), we collect over 550K online ad im- ages and use Qwen3-VL-32B (Bai et al. 2025a) to extract 51 predefined visual attributes, including composition and copywriting, among others. Most attributes use constrained candidate values for consistent extraction, while Scene Type and Visual Style permit open-ended descriptions. We then analyze the extracted attribute values and their distributions, pre-excluding 14 attributes with unstable extraction, subjec- tive or redundant semantics, or weak attribution signals, and retain 37 candidates for statistical analysis. To identify visual attributes associated with click-through rate (CTR), we combine ANOVA significance with Ran- dom Forest feature importance, as illustrated in Fig. 2(b.1). To reduce product-level differences in popularity and user demand, CTR is first normalized within each product. For ANOVA (Fisher 1970), samples are grouped by their ex- tracted attribute values, and the between-group and within- group variations in normalized CTR are compared. The re- sulting F-statistic is converted into a p-value under the cor- responding F distribution, where a smaller value indicates a stronger association with CTR. Since ANOVA evaluates each attribute independently, we further train a Random For- est regressor (Breiman 2001) to predict normalized CTR and estimate attribute contributions under multi-attribute in- teractions. Categorical attributes are converted into one-hot features, and the contribution of each split is measured by the weighted reduction in mean squared error: ∆MSE = MSE(D)− |D L | |D| MSE(D L )− |D R | |D| MSE(D R ), (1) whereD L andD R denote the subsets produced by the split. Importance scores of one-hot branches for each attribute are summed and normalized as attribute-level contributions. The final selection considers ANOVA significance, Ran- dom Forest contribution, distribution reliability, semantic re- dundancy, and practical editability. Most attributes follow the two quantitative signals, with a few boundary cases adjusted accordingly. For example, Product Quantity is retained for its relatively high Random Forest contribution and direct Policy Model Reference Model Instruction Please analyze...and provide editing suggestions to improve its CTR. 111 Rules for copy/text edits: - When a suggestion... 111 Approved Copy List: • 强劲有力... • 静音技术无刷电机... • ... (c) CPO LOSS (b) Preference Optimization CTR: 0.009 Low CTR Creative (a) Instruction Tuning Shuffled Chosen Reject (c.1) DPO LOSS (c.2) Reg LOSS Origin Chosen Response Shuffled Chosen Response Policy Model Policy Model -1.2-0.8-0.5 ... -0.7-1.3-0.2 ... Adjust layout: ... 1111 Remove text: ... 1111 Replace product: ... L CPO_LOSS =L DPO_LOSS +λL Reg_ LOSS Add props: ... 111 Add model/limbs: ... 1111 CTR-Positive Suggections CTR-Negative Suggections Replace product: ... 1111 Adjust layout: ... 1111 Remove text: ... Adjust layout: ... 1111 Remove text: ... 1111 Replace product: ... Add props: ... 111 Add model/limbs: ... 1111 Prompt Model Adjust layout: ... 1111 Remove text: .... Add props: ..limbs: ... 1111 Adjust layout: ... 1111 Remove text: ... 1111 Replace product: ... 11 Add props: ... 111 Add model/limbs: ... 1111 ... ... After Instruction Tuning Preference Signal Training Figure 3: Training framework of ECC. (a) shows instruction tuning, where the model learns to generate CTR-driven editing suggestions. (b) shows preference optimization using original chosen, shuffled chosen, and rejected responses with policy and reference models. (c) presents the proposed CPO strategy, which combines the standard DPO loss with a consistency regularizer to enforce consistent preferences between chosen responses and their semantics-preserving shuffled variants. editability, while Copy Hierarchy Consistency and Product Shadow are excluded due to redundancy or unstable distribu- tions. This yields 28 CTR-sensitive attributes from 37 candi- dates. Their relative importance is visualized in Fig. 2(b.2), where larger words indicate higher importance; complete re- sults are provided in the supplementary material. Instruction-Guided Editing Learning As shown in Fig. 3(a), we perform instruction tuning us- ing the CTR-driven editing-suggestion dataset constructed based on the 28 selected attributes. Given a source ad image I r and an instruction q, the prompt model learns to gener- ate executable editing suggestions for visual factors such as composition, copywriting, and background. Since the attribute-guided supervision provides only an initial editing signal and not every operation necessarily improves CTR, we introduce Stochastic Editing Operation Sampling. For each training instance, 1–3 operations are randomly sampled from the complete suggestion and con- catenated as the target response. This exposes the model to diverse operation combinations while leaving subsequent online preference optimization to identify modifications that genuinely improve CTR. Let ̃ y = ( ̃y 1 ,..., ̃y T ) denote the sampled response. The model is optimized using the standard autoregressive loss: L SFT =− T X t=1 logP θ ( ̃y t | I r ,q, ̃y <t ),(2) where ̃y t denotes the t-th token, ̃y <t denotes the preceding tokens, andT is the response length. After instruction tuning, the prompt model acquires the initial capability to generate executable CTR-oriented editing suggestions, which are fur- ther refined using real online click feedback. Preference Optimization As shown in Fig. 3(b), the instruction-tuned prompt model generates multiple editing suggestions for each source image. These suggestions are executed by the image-editing back- end, and the resulting creatives are deployed online. Their real exposure and click feedback are then used to construct preference pairs, allowing the model to learn which editing operations are more likely to improve CTR. Given an input X, let Y + and Y − denote suggestions with higher and lower post-editing CTR, respectively. We first adopt the DPO objective to align the model with these online click preferences: L DPO =− logσ β log π θ (Y + |X) π ref (Y + |X) − log π θ (Y − |X) π ref (Y − |X) , (3) where π θ and π ref denote the policy and frozen reference models, respectively. This preference-learning stage further refines the initial attribute-guided supervision according to real online outcomes. However, standard DPO may overfit to superficial pat- terns in editing suggestions rather than underlying editing semantics. Inspired by invariance learning in self-supervised learning (Chen et al. 2020), we identify a task-specific order invariance: reordering independent editing operations does not affect the final edited image, so the original and shuffled chosen suggestions should receive consistent preferences. We construct e Y + by randomly shuffling the independent op- erations in each chosen suggestion Y + while preserving its semantics and editing effect. As illustrated in Fig. 3(c), the consistency term encourages them to receive similar length- normalized log-likelihoods (Lu et al. 2024): L Reg = logπ θ (Y + |X) |Y + | − logπ θ ( e Y + |X) | e Y + | ! 2 , (4) where |Y| denotes the response length in tokens. The final objective is L CPO =L DPO + λL Reg ,(5) where λ controls the regularization strength. CPO retains the DPO objective derived from online CTR feedback while explicitly enforcing preference consistency under semantics- preserving order variations. Datasets To support the two training stages, we construct an instruction-tuning dataset for learning executable editing suggestions and an online preference dataset for reinforce- ment optimization. The former provides the model with an initial capability to identify and describe actionable modifi- cations, while the latter further aligns these suggestions with real user click preferences. As shown in Fig. 2(c), we first group sufficiently ex- posed creatives by product and construct image pairs with clear CTR differences. For each pair, the lower-CTR im- age is treated as the editing target, while the higher-CTR image serves as a reference that provides potentially benefi- cial visual cues. Guided by the 28 selected attributes, GPT- 5.5 (OpenAI 2026) compares the two creatives and generates executable editing suggestions covering visual factors such as composition, copywriting, product presentation, and back- ground design. The reference image and extracted attribute values are used only during data construction and are not included in the model input. Each instruction-tuning sample therefore consists of the lower-CTR creative and an editing instruction as input, with the generated editing suggestion as the target response. After data filtering and quality control, this process produces 101,699 training samples. After instruction tuning, the model generates multiple can- didate editing suggestions for each source image. These sug- gestions are executed by the image-editing backend to pro- duce corresponding creatives, which are then reviewed and deployed online to collect real exposure and click feedback. For the same source image, suggestions that lead to higher and lower post-editing CTRs are selected as chosen and re- jected responses, respectively. Each preference sample con- tains the same source image and instruction as input, together with one chosen suggestion and one rejected suggestion, en- suring that the preference signal reflects differences in editing strategies rather than input conditions. This process produces 68,704 preference pairs. The structures of both datasets are shown in Fig. 2(d). Experiments Experimental Setup Implementation Details. We use Qwen3-VL-8B-Instruct as the base multimodal model for generating editing sugges- tions. All models are trained using the SWIFT (Zhao et al. 2024) framework on 4 NVIDIA A800 GPUs. For instruction- guided editing learning, we apply LoRA to all linear layers with rank 8 and alpha 32, and train the model for 3 epochs with a learning rate of 1×10 −4 . For preference optimization, we initialize from the instruction-tuned adapter and optimize using the proposed CPO objective. This stage uses LoRA training for 4 epochs with a learning rate of 5× 10 −5 , and the regularization weight λ is set to 1. Online Evaluation Protocol. To evaluate the practical ef- fectiveness of the generated editing suggestions, we conduct a one-week online experiment on a large-scale e-commerce platform covering 44 product categories. For all compared methods, Gemini 3 Pro Image Preview (Google 2025) applies the generated suggestions to the original images, and the re- sulting creatives are deployed online. Before deployment, all edited images are reviewed by professional practitioners to ensure visual quality and foreground–background compati- bility. Their real exposure and click feedback are then used to determine the CTR effects of different editing suggestions. The experiment accumulates approximately 20 million im- pressions, providing reliable feedback for both preference construction and online evaluation. Evaluation Metrics. We evaluate each method using the aggregate click-through rate (CTR) gain between edited and original advertising creatives: ∆ CTR = CTR edit − CTR orig CTR orig × 100%.(6) Here, CTR edit and CTR orig denote the impression-weighted CTRs of the edited and original creatives, respectively. Online Results Main Online Results. Tab. 1 reports the main online com- parison in terms of CTR gain. ECC achieves the highest im- provement of 6.86%, substantially outperforming the exist- ing CTR-oriented method CAIG (Chen et al. 2025b), which achieves only 0.20%. ECC also surpasses GPT and Gemini by 3.61 and 1.34 percentage points, respectively. DPO S1 and DPO S2 indirectly exploit order invariance through sentence- order shuffling on one-third and all training samples, whereas CPO enforces it directly. Vanilla DPO and its shuffled- response variants yield only limited gains. Under the cur- rent traffic scale, ECC is expected to generate approximately RMB 117 million in additional annual revenue. These re- sults demonstrate the effectiveness of ECC in generating CTR-oriented editing suggestions and validate the proposed preference optimization strategy. MethodCTR Gain↑ CAIG (Chen et al. 2025b)+0.20% DPO+0.68% DPO S1 +1.12% DPO S2 +1.29% GPT+3.25% Gemini+5.52% ECC+6.86% Table 1: Main online results in terms of CTR gain. Analysis Across Product Categories and Advertising Sce- narios. As shown in the Product Category and Advertising Scenario blocks of Tab. 2, ECC consistently improves CTR across all evaluated categories and placements. The gains are particularly strong in Home Appliances, Computers, and Beauty, as well as on the Order Page and Product Detail Page. These results demonstrate that ECC generalizes across diverse products and traffic contexts rather than benefiting only specific categories or scenarios. SettingCTR Gain↑ Product Category Home Appliances+14.01% Computers+11.61% Beauty+11.37% Digital Products+3.92% Personal Care+3.28% Advertising Scenario Order Page+13.32% Product Detail Page+12.79% Guess You Like+5.53% Homepage Recommendation Feed+5.23% Shopping Cart Page+4.23% All+6.86% Table 2: CTR gains across different product categories and advertising scenarios. Aesthetic Evaluation We evaluate edited-image quality using LAION Aes- thetic (Schuhmann et al. 2022), Aesthetic Predictor v2.5 (Zhai et al. 2023; discus0434 2024), PickScore (Kirstain et al. 2023), and GPT-5.5. For each image group, methods are ranked by each evaluator and averaged across groups. The overall score is the mean of the four evaluator-specific ranks, with lower values indicating better performance. As shown in Tab. 3, ECC achieves the best overall ranking, suggesting that CTR-oriented editing preserves visual quality. However, evaluator rankings are inconsistent and do not fully align with the online CTR gains in Tab. 1, indicating that aesthetic quality and advertising effectiveness are not equivalent. Order Robustness Analysis To evaluate CPO, we perform inference on approximately 1,000 editing samples using models trained with different Method LAION Predictor PickScore VLM Avg.↓ ECC2.472.412.362.97 2.55 DPO S2 2.532.342.623.01 2.63 DPO S1 2.612.452.642.92 2.66 DPO3.342.593.202.98 3.03 Gemini 3.663.722.952.73 3.27 GPT3.723.703.402.78 3.40 Original 2.683.793.823.62 3.48 Table 3: Average rankings under four aesthetic evaluators. preference optimization strategies. For each chosen sugges- tion, we construct a shuffled variant by permuting its editing operations without changing its content or editing effect. Following the DPO implicit reward formulation, we com- pute length-normalized scores from the policy and reference models. We report three margins: ∆ c→r between chosen and rejected responses, ∆ c→s between chosen and shuffled chosen responses, and ∆ s→r between shuffled chosen and re- jected responses. Larger ∆ c→r and ∆ s→r indicate stronger preference discrimination, while smaller ∆ c→s indicates bet- ter order consistency. As shown in the Method Comparison block of Tab. 4, DPO exhibits similar ∆ c→r and ∆ c→s values, with ∆ s→r close to zero, indicating sensitivity to sentence-order variations. DPO S2 improves order consistency through shuffled data augmentation, while ECC achieves the lowest ∆ c→s and the highest ∆ s→r while maintaining strong chosen-rejected discrimination. These results demonstrate that CPO more effectively learns order-invariant preferences. Method∆ c→r ↑ ∆ c→s ↓ ∆ s→r ↑ Method Comparison DPO0.0213 0.0210 0.0003 DPO S2 0.0316 0.0126 0.0191 ECC0.0316 0.0119 0.0197 Ablation on Regularization Weight ECC (λ = 0.75) 0.0216 0.0121 0.0104 ECC (λ = 1.25) 0.0225 0.0091 0.0034 ECC (λ = 1.00) 0.0316 0.0119 0.0197 Table 4: Order robustness under DPO implicit rewards. Validation of CTR-Sensitive Visual Attributes To validate the discovered CTR-sensitive attributes, we for- mulate CTR estimation as a pairwise classification task: given two images of the same product, the model predicts which one has a higher CTR. We retain image pairs with suf- ficient exposure, resulting in 37,020 training pairs and 9,887 test pairs. Presenting each pair in both orders yields 74,040 training samples and 19,774 test samples. We evaluate general-purpose MLLMs and a task-specific CTR prediction model under three input settings: image pairs only, image pairs with prompts describing all 51 predefined attributes, and image pairs with prompts describing the se- lected 28 attributes. As shown in Tab. 5, the task-specific Product 1Product 2 Product 5Product 6 Product 3Product 4 Product 7Product 8 bad_ctr:0.0024 optimized_ctr:0.0782 bad_ctr:0.0011 optimized_ctr:0.0120 bad_ctr:0.0041 optimized_ctr:0.0248 bad_ctr:0.0126 optimized_ctr:0.1986 bad_ctr:0.0177 optimized_ctr:0.1781 bad_ctr:0.0060 optimized_ctr:0.0662 bad_ctr:0.0013 optimized_ctr:0.0147bad_ctr:0.0052 optimized_ctr:0.0319 Figure 4: Qualitative examples of CTR-driven edits from original images on the left to optimized images on the right. model consistently outperforms general-purpose MLLMs. Attribute prompts generally improve the image-only setting, while the selected 28 attributes achieve the highest accuracy of 59.56%, indicating that they retain CTR-relevant informa- tion while excluding redundant or unstable factors. Model Input Setting Image Only +51 Attr. +28 Attr. GPT47.5849.5049.86 Gemini48.9348.2449.44 Ours58.4959.4659.56 Table 5: Pairwise CTR prediction accuracy (%) under differ- ent input settings. Ablation Study Ablation on Preference Optimization. Tab. 6 compares standard DPO, partial and full response-order shuffling aug- mentation, and the proposed Consistent Preference Opti- mization Strategy. DPO brings only limited CTR improve- ment, indicating that directly fitting offline preference pairs is insufficient for learning robust editing preferences. Although shuffled-response augmentation slightly improves CTR gain by exposing the model to different operation orders, it re- mains much weaker than CPO. CPO explicitly constrains se- mantically equivalent editing suggestions to receive consis- tent preferences, helping the model focus on editing seman- tics rather than superficial ordering patterns. This provides a more direct way to encode task-specific invariance than simply expanding the training data with shuffled variants. As a result, ECC achieves the highest CTR gain of 6.86%, validating the proposed preference optimization design. Sensitivity Analysis of the Regularization Weight. To examine the sensitivity of CPO to the regularization weight λ, we evaluate ECC with several weight settings. As shown in the Ablation on Regularization Weight block of Tab. 4, all tested settings improve order robustness over DPO, con- firming that CPO remains effective across different weights VariantCTR Gain DPO+0.68% DPO S1 +1.12% DPO S2 +1.29% ECC w/ CPO +6.86% Table 6: Ablation results on preference optimization. and is not overly sensitive to a specific configuration. Among them, a weight of 1.00 provides the best overall trade-off be- tween order invariance and preference discrimination, and is therefore adopted in the main experiments. Qualitative Analysis Fig. 4 shows representative CTR-driven editing examples, with original and edited creatives shown on the left and right, respectively. The edited images feature cleaner layouts, more suitable backgrounds, clearer product presentation, and more focused copy, resulting in higher online CTR. Additional ex- amples with corresponding editing suggestions are provided in the supplementary material. Conclusion In this paper, we propose ECC, a CTR-driven framework that improves existing ad images through fine-grained edit- ing rather than generation from scratch. Using large-scale online data and statistical analysis, ECC identifies 28 CTR- sensitive attributes from 51 predefined dimensions and uses them to construct actionable editing supervision for targeted creative refinement. It first learns to generate executable edit- ing suggestions through instruction tuning and is then further refined through preference optimization based on real click feedback. We further propose Consistent Preference Opti- mization (CPO) as ECC’s preference optimization strategy, augmenting DPO with a consistency regularizer to mitigate overfitting to spurious ordering patterns and improve pref- erence generalization. Large-scale online experiments show that ECC outperforms strong MLLM-based baselines and achieves the highest aggregate CTR gain across diverse prod- uct categories and advertising scenarios, demonstrating the effectiveness and practical value of fine-grained attribute editing for creative optimization. References Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; Ge, W.; Guo, Z.; Huang, Q.; Huang, J.; Huang, F.; Hui, B.; Jiang, S.; Li, Z.; Li, M.; Li, M.; Li, K.; Lin, Z.; Lin, J.; Liu, X.; Liu, J.; Liu, C.; Liu, Y.; Liu, D.; Liu, S.; Lu, D.; Luo, R.; Lv, C.; Men, R.; Meng, L.; Ren, X.; Ren, X.; Song, S.; Sun, Y.; Tang, J.; Tu, J.; Wan, J.; Wang, P.; Wang, P.; Wang, Q.; Wang, Y.; Xie, T.; Xu, Y.; Xu, H.; Xu, J.; Yang, Z.; Yang, M.; Yang, J.; Yang, A.; Yu, B.; Zhang, F.; Zhang, H.; Zhang, X.; Zheng, B.; Zhong, H.; Zhou, J.; Zhou, F.; Zhou, J.; Zhu, Y.; and Zhu, K. 2025a. Qwen3-VL Technical Report. arXiv preprint arXiv:2511.21631. Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J. 2025b. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923. Breiman, L. 2001. Random forests. Machine learning, 45(1): 5–32. Brooks, T.; Holynski, A.; and Efros, A. A. 2023. Instruct- pix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 18392–18402. Cao, M.; Zhang, X.; Zheng, Y.; and Xia, Z. 2024. Instruction- based image manipulation by watching how things move. arXiv preprint arXiv:2412.12087. Chen, B.; Zhong, C.; Xiang, W.; Geng, Y.; and Xie, X. 2024. Virtualmodel: Generating object-id-retentive human-object interaction image by diffusion model for e-commerce mar- keting. arXiv preprint arXiv:2405.09985. Chen, H.; Zhou, M.; Jiang, J.; Chen, J.; Lu, Y.; Xiao, B.; Ge, T.; and Zheng, B. 2025a. PAID: A Framework of Product-Centric Advertising Image Design. arXiv preprint arXiv:2501.14316. Chen, J.; Xu, J.; Jiang, G.; Ge, T.; Zhang, Z.; Lian, D.; and Zheng, K. 2021. Automated creative optimization for e-commerce advertising. In Proceedings of the Web Confer- ence 2021, 2304–2313. Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. 2020. A simple framework for contrastive learning of visual repre- sentations. In International conference on machine learning, 1597–1607. PmLR. Chen, X.; Feng, W.; Du, Z.; Wang, W.; Chen, Y.; Wang, H.; Liu, L.; Li, Y.; Zhao, J.; Li, Y.; et al. 2025b. CTR-driven advertising image generation with multimodal large language models. In Proceedings of the ACM on Web Conference 2025, 2262–2275. Chowdhury, S. R.; Kini, A.; and Natarajan, N. 2024. Provably robust dpo: Aligning language models with noisy feedback. arXiv preprint arXiv:2403.00409. discus0434. 2024. Aesthetic Predictor V2.5: A SigLIP-Based Aesthetic Score Predictor. Accessed: 2026-07-17. Du, Z.; Feng, W.; Wang, H.; Li, Y.; Wang, J.; Li, J.; Zhang, Z.; Lv, J.; Zhu, X.; Jin, J.; et al. 2024. Towards reliable adver- tising image generation using human feedback. In European Conference on Computer Vision, 399–415. Springer. Fisher, R. A. 1970. Statistical methods for research workers. In Breakthroughs in statistics: Methodology and distribution, 66–70. Springer. Ge, T.; Zhao, L.; Zhou, G.; Chen, K.; Liu, S.; Yi, H.; Hu, Z.; Liu, B.; Sun, P.; Liu, H.; et al. 2018. Image matters: Visually modeling user behaviors using advanced model server. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, 2087–2095. Google. 2025. Gemini 3 Pro Image Preview. Accessed: 2026-07-15. Google. 2026. Gemini 3.1 Pro Preview. Accessed: 2026-07- 15. Guo, H.; Tang, R.; Ye, Y.; Li, Z.; and He, X. 2017. DeepFM: a factorization-machine based neural network for CTR pre- diction. arXiv preprint arXiv:1703.04247. Huang, Y.; Xie, L.; Wang, X.; Yuan, Z.; Cun, X.; Ge, Y.; Zhou, J.; Dong, C.; Huang, R.; Zhang, R.; et al. 2024. Smartedit: Exploring complex instruction-based image edit- ing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 8362–8371. Hui, M.; Yang, S.; Zhao, B.; Shi, Y.; Wang, H.; Wang, P.; Zhou, Y.; and Xie, C. 2024. Hq-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990. Kirstain, Y.; Polyak, A.; Singer, U.; Matiana, S.; Penna, J.; and Levy, O. 2023. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in neural information processing systems, 36: 36652–36663. Ku, Y.-N.; Kuznetsov, M.; Mishra, S.; and de Juan, P. 2023. Staging e-commerce products for online advertising using retrieval assisted image generation. arXiv preprint arXiv:2307.15326. Li, F.; Liu, A.; Feng, W.; Zhu, H.; Li, Y.; Zhang, Z.; Lv, J.; Zhu, X.; Shen, J.; Lin, Z.; et al. 2023. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 1249–1258. Lian, J.; Zhou, X.; Zhang, F.; Chen, Z.; Xie, X.; and Sun, G. 2018. xdeepfm: Combining explicit and implicit feature interactions for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 1754–1763. Lin, K.; Zhang, X.; Li, F.; Wang, P.; Long, Q.; Deng, H.; Xu, J.; and Zheng, B. 2022. Joint optimization of ad ranking and creative selection. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2341–2346. Lin, Y.; Seto, S.; Ter Hoeve, M.; Metcalf, K.; Theobald, B.-J.; Wang, X.; Zhang, Y.; Huang, C.; and Zhang, T. 2024. On the limited generalization capability of the implicit reward model induced by direct preference optimization. In Findings of the Association for Computational Linguistics: EMNLP 2024, 16015–16026. Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual instruction tuning. Advances in neural information processing systems, 36: 34892–34916. Lu, J.; Li, J.; An, S.; Zhao, M.; He, Y.; Yin, D.; and Sun, X. 2024. Eliminating biased length reliance of direct preference optimization via down-sampled kl divergence. In Proceed- ings of the 2024 Conference on Empirical Methods in Natural Language Processing, 1047–1067. Mishra, S.; Verma, M.; Zhou, Y.; Thadani, K.; and Wang, W. 2020. Learning to create better ads: Generation and ranking approaches for ad creative refinement. In Proceedings of the 29th ACM international conference on information & knowledge management, 2653–2660. OpenAI. 2026. GPT-5.5 System Card. Accessed: 2026-07- 15. Qian, Y.; Bocek-Rivele, E.; Song, L.; Tong, J.; Yang, Y.; Lu, J.; Hu, W.; and Gan, Z. 2026. Pico-banana-400k: A large- scale dataset for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 37226–37235. Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Er- mon, S.; and Finn, C. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 53728–53741. Schuhmann, C.; Beaumont, R.; Vencu, R.; Gordon, C.; Wightman, R.; Cherti, M.; Coombes, T.; Katta, A.; Mullis, C.; Wortsman, M.; et al. 2022. Laion-5b: An open large- scale dataset for training next generation image-text mod- els. Advances in neural information processing systems, 35: 25278–25294. Sheynin, S.; Polyak, A.; Singer, U.; Kirstain, Y.; Zohar, A.; Ashual, O.; Parikh, D.; and Taigman, Y. 2024. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8871–8879. Wang, P.; Bai, S.; Tan, S.; Wang, S.; Fan, Z.; Bai, J.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Fan, Y.; Dang, K.; Du, M.; Ren, X.; Men, R.; Liu, D.; Zhou, C.; Zhou, J.; and Lin, J. 2024. Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191. Wang, S.; Liu, Q.; Ge, T.; Lian, D.; and Zhang, Z. 2021. A Hybrid Bandit Model with Visual Priors for Creative Rank- ing in Display Advertising. In Proceedings of the Web Con- ference 2021, 2324–2334. Wei, P.; Yang, X.; Liu, S.; Wang, L.; and Zheng, B. 2022. CREATER: CTR-driven advertising text generation with controlled pre-training and contrastive fine-tuning. In Pro- ceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, 9–17. Xu, Y.; Kong, J.; Wang, J.; Pan, X.; Lin, B.; and Liu, Q. 2025. Insightedit: Towards better instruction following for image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2694–2703. Yang, H.; Yuan, J.; Yang, S.; Xu, L.; Yuan, S.; and Zeng, Y. 2024. A new creative generation pipeline for click-through rate with stable diffusion model. In Companion Proceedings of the ACM Web Conference 2024, 180–189. Yu, Q.; Chow, W.; Yue, Z.; Pan, K.; Wu, Y.; Wan, X.; Li, J.; Tang, S.; Zhang, H.; and Zhuang, Y. 2025. Anyedit: Mastering unified high-quality image editing for any idea. In Proceedings of the Computer Vision and Pattern Recognition Conference, 26125–26135. Zhai, X.; Mustafa, B.; Kolesnikov, A.; and Beyer, L. 2023. Sigmoid loss for language image pre-training. In Proceed- ings of the IEEE/CVF international conference on computer vision, 11975–11986. Zhang, K.; Mo, L.; Chen, W.; Sun, H.; and Su, Y. 2023. Mag- icbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36: 31428–31449. Zhao, Y.; Huang, J.; Hu, J.; Wang, X.; Mao, Y.; Zhang, D.; Jiang, Z.; Wu, Z.; Ai, B.; Wang, A.; Zhou, W.; and Chen, Y. 2024. SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning. arXiv:2408.05517. Evolving Creatives with Clicks: CTR-Driven Advertising image Editing via Attribute Disentanglement and Preference Alignment Supplementary Material Overview This supplementary material provides additional details and qualitative results of ECC. We organize the supplementary material into three parts. First, we introduce the CTR-aware visual attribute mining process, which establishes a fine-grained attribute space for analyzing the relationship between advertising creatives and CTR. Second, we present representative instruction-tuning examples, illustrating how CTR-sensitive attributes are used to construct executable editing suggestions from same-product creative pairs. Third, we provide additional qualitative results to demonstrate the effectiveness and diversity of CTR-driven image editing across different products and editing scenarios. A. CTR-Aware Visual Attribute Mining This section provides additional details of the CTR-aware visual attribute mining process. We construct a visual attribute schema with 51 attributes for large-scale advertising creative analysis. These attributes are extracted by Qwen3-VL-32B and further analyzed through statistical evaluation and contribution modeling. After preliminary screening, 37 candidate attributes are evaluated using one-way ANOVA and Random Forest contribution. Considering statistical significance, attribute reliability, semantic redundancy, and practical editability, we finally identify 28 CTR-sensitive attributes. A.1 Visual Attribute Schema To systematically characterize the visual factors of advertising creatives, we construct a hierarchical attribute schema covering composition and layout, copywriting and text design, background and visual style, product presentation, model and interaction, and logo-related information. For each attribute, we define a candidate label space to reduce semantic fragmentation during large-scale extraction. Most attributes follow the Strict setting, where Qwen3-VL-32B selects values from predefined candidates. Attributes with the Extensible setting allow additional values when predefined candidates cannot sufficiently describe the visual content. The complete attribute schema and candidate values are presented in Tab. 7. Table 7: Visual attribute schema. AttributeLabel Space / Candidate Values Composition and Layout Image Aspect RatioStrict: 1:1, 16:9, 9:16, 3:4, 4:3, 3:5, 4:5. Composition SymmetryStrict: central symmetry, rule of thirds, horizontal symmetry, vertical symmetry, asymmetric balance, diagonal composition, radial composition. Element DensityStrict: sparse, medium, dense. Visual BalanceExtensible: balanced, cluttered, visually unbalanced, or other layout balance descriptions. Element AlignmentStrict: left-aligned, centered, right-aligned, top-aligned, bottom-aligned. Number of Visual CentersStrict: no clear center, single center, double centers, multiple centers. Visual LayeringStrict: single-layer flat layout, foreground-background layering, multi-layer composition. Shooting AngleStrict: eye-level, high angle, low angle, full view, flat lay, close-up, macro. Copywriting and Text Design Copy ContentOpen-ended: promotional copy extracted from the image. Copy DensityNumeric: number of promotional text blocks. Copy LanguageStrict: Chinese, English, mixed language. Text–Product ConsistencyExtensible: consistent with product, intentional contrast, neutral, severe mismatch, or other consistency descriptions. Copy AlignmentStrict: left, right, center, justified, vertical, outline. Copy PositionStrict: top-left, top-right, bottom-left, bottom-right, left-center, right-center, top-center, bottom- center, center. Character SpacingStrict: large, medium, small. Font SizeStrict: extra large, large, medium, small, tiny. Continued on next page AttributeLabel Space / Candidate Values Font TypeExtensible: minimalist, handwritten, retro, calligraphic, rounded, bold, artistic, or other font styles. Font Style & EffectsExtensible: no effect, gradient, outline, shadow, glow, metallic texture, neon effect, 3D effect, or other text effects. Copy Hierarchy ConsistencyStrict: yes, no. Copy AppropriatenessStrict: appropriate, inappropriate. Text–Subject LayeringStrict: no overlap, text in front of subject, text behind subject. Background, Scene, and Visual Style Background TypeExtensible: indoor home, outdoor nature, urban street, office, industrial or factory, studio, virtual background, none, or other background types. Scene TypeExtensible: solid color, gradient, AI or virtual scene, abstract texture, geometric background, 3D space, none, or other scene types. Background Color ToneExtensible: warm, cool, neutral, low-saturation, soft, high-contrast, or other color-tone descrip- tions. Visual StyleExtensible: fresh and natural, colorful, minimalist, cinematic, cyberpunk, premium metallic, matte, gothic, retro, film-grain, or other visual styles. Background–Product HarmonyExtensible: well matched, product highlighted by contrast, neutral, mismatched, or other harmony descriptions. Color RelationshipExtensible: analogous, monochrome, complementary, high contrast, background highlights sub- ject, or other color relationships. Depth of FieldStrict: sharp, slight blur, strong depth-of-field blur, motion blur, none. Lighting ConsistencyExtensible: natural lighting, lighting inconsistency, shadow error, missing shadow, floating object, or other lighting issues. Product Presentation Product QuantityStrict: product set, single item, 2–3 items, 4–5 items, large flat-lay assortment. Product Area RatioStrict: > 3/4 dominant, 1/2–3/4 main, 1/4–1/2 moderate, 1/8–1/4 small, < 1/8 decorative. Product PositionStrict: left, right, center, top, bottom, top-left, top-right, bottom-left, bottom-right, none. Product OrientationStrict: front view, 45-degree tilt, top-down view, bottom-up view. Product Display StateStrict: static display, in use. Product CompletenessStrict: fully visible, partially cropped, edge only, heavily occluded. Product OverlapStrict: no overlap, slight overlap, heavy stacking. Product ShadowStrict: no shadow, natural shadow, reflection. Model and Interaction Model PresenceStrict: present, absent. Number of ModelsStrict: 1, 2, more than 2, none. Model TypeExtensible: human model, non-human model, virtual or AI model, or other model types. Model AgeStrict: infant, child, young adult, middle-aged, elderly, none. Model GenderStrict: male, female, neutral, mixed, none. Model OutfitExtensible: business, casual, sportswear, techwear, retro, ethnic, homewear, pajamas, minimalist, premium, formalwear, eveningwear, none, or other outfit styles. Model Gaze DirectionStrict: looking at camera, looking at product, looking elsewhere, eyes closed or sunglasses, none. Model ActionExtensible: standing, sitting, walking, running, product demonstration, sports or fitness action, facial close-up, none, or other actions. Model–Product InteractionExtensible: physical contact, gaze interaction, functional demonstration, no interaction, or other interaction types. Inter-Model Physical Interaction Extensible: no contact, handshake, hug, leaning, side-by-side, object passing, interactive demon- stration, none, or other interactions. Logo Logo PresenceStrict: present, absent. Logo PositionStrict: left, right, center, top, bottom, top-left, top-right, bottom-left, bottom-right, none. Logo SizeStrict: large, medium, small, none. Logo CountStrict: 1, 2, 3 or more, none. A.2 Preliminary Attribute Screening Before statistical analysis, we exclude attributes that cannot provide reliable and independent CTR attribution. The main reasons include unstable large-scale extraction, highly subjective labels, severe distribution imbalance, semantic redundancy, and insufficient valid samples. Tab. 8 summarizes the 14 pre-excluded attributes and their exclusion rationales. Table 8: Pre-excluded visual attributes and their exclusion rationale. Pre-excluded AttributeExclusion Rationale Image Aspect RatioThe production scenario enforces square advertising creatives, making aspect ratio uninformative for CTR attribution. Copy ContentOn-image copy is often constrained by platform review, legal requirements, and business rules. Therefore, free-form copy content is excluded as a stable statistical factor, while copy-related visual properties are retained. Copy AlignmentA single global alignment label is unreliable because different copy blocks within the same creative may follow different alignment patterns. Character SpacingCharacter spacing is visually subtle in most creatives and is difficult to extract consistently at large scale. Font SizeFont size is highly relative to image scale, resolution, and layout context, making stable automatic annotation difficult across heterogeneous creatives. Font TypeThis attribute substantially overlaps with Font Style & Effects and may introduce redundant statistical factors. Copy PositionCopy position is better represented by text–product layout and overall composition rather than by an isolated spatial location label. Copy AppropriatenessThe judgment is highly subjective and lacks a stable objective boundary for large-scale automatic extraction. Logo PositionLogo-related information is analyzed at a coarser presence level to avoid sparse and unstable subgroups. Logo SizeLogo size is merged into the coarse logo-presence factor because fine-grained size labels provide limited independent statistical support. Logo CountLogo count is merged into logo presence since multiple-logo cases are relatively sparse in the dataset. Inter-Model Physical InteractionThis attribute only applies to rare multi-model cases and does not provide sufficient samples for robust statistical analysis. Product OverlapProduct overlap is difficult to label consistently when multiple product sets or dense layouts appear in the same creative. Text–Subject LayeringThe distribution is highly imbalanced because most creatives do not contain explicit text–subject overlap. A.3 Statistical Analysis and Final Selection For the remaining candidate attributes, we combine one-way ANOVA with Random Forest contribution. ANOVA measures the independent association between each attribute and normalized CTR, while the Random Forest captures its contribution under interactions with other attributes. As reported in Tab. 9, most attributes are initially selected if they exhibit either significant ANOVA results (p < 0.05) or a normalized Random Forest contribution above 5.0%. The complete procedure is summarized in Algorithm 1. Analysis of Variance. We employ analysis of variance (ANOVA) to examine whether different values of each visual attribute are associated with statistically significant differences in click-through rate (CTR). Since the raw CTR of an image can be influenced by product-level factors, such as product popularity, category, price, and traffic allocation, directly comparing CTR across different products may introduce considerable bias. Therefore, we normalize CTR within each product and use the normalized CTR as the response variable. For each visual attribute, images are grouped according to their corresponding attribute values. ANOVA evaluates whether the mean normalized CTR differs significantly among these groups by decomposing the total variance into between-group and within-group components. The between-group component captures the variation of group means caused by different attribute values, while the within-group component measures the residual variation among samples within the same group. The test statistic is defined as F = MS between MS within ,(7) Algorithm 1: CTR-Aware Visual Attribute Mining Require: DatasetD, initial attribute setA 0 ,|A 0 | = 51 Ensure: CTR-sensitive attribute setA ∗ ,|A ∗ | = 28 1: for each sample i∈D do 2: r i ← c i n i 3:ˆr i ← Norm(r i | s i ) 4: end for 5: A← PreExclude(A 0 ) 6: |A| = 37 7: for each a∈A do 8: G a ← Group(D,a) 9: p a ← ANOVA(ˆr i i∈G a ) 10: end for 11: X← Encode(D,A) 12: f RF ← RFRegress(X, ˆ r) 13: for each a∈A do 14: w a ← P j∈B(a) Imp f RF (j) 15: end for 16: A ∗ ← Select (A,p a ,w a ,R dist ,R sem ,R edit ) 17: return A ∗ where MS between and MS within denote the mean squares between and within attribute groups, respectively. A larger F value indicates that the differences among attribute groups are relatively larger than the variations within groups. The corresponding significance test evaluates the null hypothesis that all attribute-value groups share the same expected normalized CTR. Rejecting this hypothesis suggests that the attribute has a statistically significant association with relative click performance. Since ANOVA evaluates each attribute independently, it provides an interpretable measurement of the marginal relationship between individual visual attributes and CTR, but does not explicitly capture nonlinear effects or interactions among multiple attributes. Random Forest Regression. We further employ a random forest regressor to evaluate the contributions of visual attributes under a multivariate setting. Categorical attribute values are first converted into binary indicator features, such that each possible attribute value is represented by an individual feature branch. A random forest is composed of multiple regression trees. Each tree recursively partitions the feature space according to visual attribute values, with the objective of grouping images that have similar normalized CTR values. At each internal node, the model selects a feature and a splitting rule that produce the largest reduction in regression error. In this way, regression trees can model nonlinear relationships and conditional dependencies between visual attributes and CTR without requiring a predefined functional form. Different regression trees are constructed from randomized samples and feature subsets. Their predictions are aggregated to obtain the final CTR prediction, reducing the instability and overfitting of a single decision tree. The importance of each feature branch is calculated according to the cumulative reduction in regression impurity produced by splits using that branch across all trees. Since one categorical attribute may be expanded into multiple binary branches, we aggregate the importance scores of all branches belonging to the same original attribute: I(a) = X j∈B(a) I j ,(8) whereB(a) denotes the set of binary feature branches derived from attribute a, and I j denotes the importance score of the j-th branch. The aggregated importance scores are then normalized across all attributes: e I(a) = I(a) P a ′ ∈A I(a ′ ) ,(9) whereA denotes the complete set of visual attributes. Unlike ANOVA, which examines each attribute independently, random forest regression evaluates the predictive contribution of an attribute while considering the effects of other attributes. It can therefore identify attributes that influence CTR through nonlinear patterns or interactions, even when their individual marginal effects are relatively weak. The random forest analysis complements ANOVA by providing a multivariate estimation of the relative contribution of each visual attribute. Implementation details. CTR is normalized within each product to reduce differences in product popularity and user demand. For each candidate attribute, samples are grouped by their extracted values, and one-way ANOVA is applied to evaluate its association with normalized CTR. Categorical attributes are one-hot encoded before training a Random Forest regressor. The importance scores of feature branches belonging to the same attribute are summed to obtain attribute-level contributions. Final selection further considers distribution reliability, semantic redundancy, and practical editability. Table 9: Statistical analysis of candidate visual attributes after preliminary screening. Attributep-value RF Contr. (%) Selected Font Style & Effects< 0.0019.45Yes Background Color Tone0.0018.29Yes Copy Density< 0.0016.28Yes Visual Style< 0.0016.08Yes Shooting Angle0.0196.04Yes Composition Symmetry< 0.0014.90Yes Background–Product Harmony< 0.0014.69Yes Scene Type< 0.0013.82Yes Product Orientation0.0203.31Yes Copy Language< 0.0013.07Yes Depth of Field< 0.0013.01Yes Model–Product Interaction< 0.0012.96Yes Element Density< 0.0012.15Yes Product Position< 0.0012.11Yes Model Outfit< 0.0012.09Yes Copy Hierarchy Consistency< 0.0011.97No Product Display State< 0.0011.89Yes Model Type< 0.0011.56Yes Model Age< 0.0011.32Yes Model Gaze Direction< 0.0011.30Yes Model Action< 0.0011.22Yes Model Gender< 0.0011.15Yes Text–Product Consistency< 0.0010.92Yes Product Area Ratio< 0.0010.88Yes Background Type< 0.0010.84Yes Number of Models< 0.0010.53Yes Model Presence< 0.0010.14Yes Product Quantity0.2755.94Yes Logo Presence0.0332.55Yes Color Relationship0.9203.17No Product Shadow< 0.0012.26No Number of Visual Centers0.3451.74No Visual Layering< 0.0010.90No Element Alignment0.9230.57No Lighting Consistency0.6420.51No Product Completeness0.1780.35No Visual Balance0.0820.06No Selection exceptions. Most attributes follow the joint criterion of ANOVA significance and Random Forest contribution. A few attributes are manually adjusted according to semantic redundancy, distribution reliability, and practical editability. Product Quantity is retained despite its weak ANOVA significance because it has a high Random Forest contribution and directly corresponds to actionable product-presentation edits. Copy Hierarchy Consistency, Product Shadow, and Visual Layering are removed despite significant ANOVA results because they are either redundant with retained attributes or dominated by unstable value distributions. Color Relationship is removed because its ANOVA significance is weak and its semantics largely overlap with Background Color Tone and Background–Product Harmony. Product Completeness and Visual Balance are also removed due to weak model contribution and redundancy with Product Area Ratio, Product Display State, Composition Symmetry, and Element Density. B. Instruction-Tuning Dataset Examples Fig. 5 presents representative instruction-tuning samples. Each sample contains a low-CTR creative as the model input and its corresponding editing suggestion as the supervision target. The higher-CTR image of the same product is used only to construct and validate the editing suggestion and is not provided to the model during training. Colored phrases indicate the CTR-sensitive attributes involved in each editing operation. Replace background: Replace the blue-purple neon sci-fi corridor, the glowing mechanical rings on both sides, and the reflective floor overall with a dark gray matte studio gradient background, making the overall visual style more concise and product-focused. Adjust layout: Reorganize the composition so that the promotional message and the laptop are more clearly separated; change the laptop from a front-facing centered display to an angled display positioned toward the lower right, leaving clearer visual space for the main message. Replace text: Condense the original dense two-line copy into the main title 「移动战神芯霸主」 and the subtitle 「一键召唤凶猛战力」, reducing copy density while keeping the message consistent with the product's gaming-performance selling point.Add a short purple horizontal line decoration below the main message area, and keep the subtitle in clean white text to support the main visual focus. Adjust product: Change the laptop from a fully open front-facing view to a top-down three-quarter angled view, with the body tilted toward the upper right and the screen changed to a half-closed state, emphasizing the gray exterior shell and the purple-blue backlit keyboard. Remove the ROG STRIX image from the original screen, so the half-closed screen no longer shows a front-facing screen image. Add props: Add two black circular support bases beneath the laptop, paired with subtle grounding elements to enrich the composition without increasing visual clutter. Replace background: Replace the gray studio advertising background and gold display platform base with a realistic indoor window-side scene, including wooden flooring, beige walls, a large window on the right, a view of buildings outside the window, and natural daylight. Adjust layout: Change the original composited poster layout of “shipping box + massage chair + top and bottom advertising text” into a text-free real-photo composition, with the massage chair enlarged and occupying most of the left-center area of the image, while leaving space for the window on the right. Remove text: Remove the top 「舒适度 爆表！」, as well as the bottom gray-and-gold horizontal bar and all post-production promotional text such as 「小腿肚按摩」「头等舱级 全身舒展」. Adjust product: Replace the composited massage chair on the right of the image with a gray- black massage chair with a realistic photographed texture, remove the display platform, maintain a three-quarter front angle, and make it larger overall and closer to the floor. Remove props: Remove the Westinghouse cardboard box on the left, the transparent calf- massage demonstration piece at the bottom, and the transparent decorative element at the bottom right. Remove logo: Remove the orange crown icon in the upper right corner. Background Color Tone Background Type / Scene Type Visual Style Product Orientation Product Position Copy DensityFont Style & Effects Shooting Angle Product Display State Element Density Logo Presence Product Area Ratio Product Quantity Replace background: Replace the white-background red promotional layout in the image, the top red label, the bottom red-and-black price bar, and the coupon/free-gift card area with a light beige indoor wall and light wood tabletop scene. Add a white rectangular display stand beneath the product, and add natural green plant decorations on the left and right. Adjust layout: Change the dense multi-product promotional layout to a minimalist centered display. Keep one green cleanser tube as the only main subject, placed in the center of the original product area and standing on the white display stand. Replace text: Replace all original top livestream promotional copy with two lines of benefit copy. The first line 「平衡水油 清爽不紧绷」 should be in dark green text, and the second line 「净彻毛 孔 告别油腻清洁更彻底」 should be in black text. Both should have no background panel and remain in the original top text area. Adjust product: Remove the white cleanser tube on the left, the sample set on the right, the foaming net, cotton pads, storage bag, and other free gifts. Keep only the original green ORIGINS cleanser tube slightly left of center, and adjust it to face forward, centered, and clearly displayed. Remove logo: Remove the ORIGINS brand logo added in post-production at the top left, keeping only the text printed on the product packaging itself. Add decoration: Add a transparent glass bottle with green branches in the blank area to the left of the original product, and add a white potted green plant in the blank area to the right of the original product to create a fresh, natural atmosphere. Low CTR Image Reference High CTR Image Figure 5: Examples of the instruction-tuning dataset. Each row shows the low-CTR input creative, its target editing suggestion, and the same-product high-CTR reference used only during data construction. Colored phrases indicate the involved CTR- sensitive attributes. C. Additional Qualitative Results Fig. 6 presents additional examples of CTR-driven ad image editing. ECC generates executable suggestions covering background, layout, product presentation, visual style, and copywriting. The resulting creatives show clearer visual organization and improved online CTR, while preserving the identity and core content of the advertised products. Before EditingAfter Editing Replace background: Replace the study desk and chairs, computer, curtains, strong sunlight, and floor shadows in the image with a simple indoor background featuring gray-white walls and light wood flooring, with soft, even lighting. Remove text: Remove the orange rounded-corner label 「书房」 in the upper right corner of the image. Add model: Add a standing woman in the original desk-and-chair area on the left side of the image, visible from side profile to feet, wearing a white short-sleeve top and light gray pants, with her left hand holding the top edge of the bookshelf and her right hand holding a book. Replace background: Replace the light blue gradient blank advertising background with a warm beige home living room scene, adding wooden flooring, beige walls, wooden shelves on both sides, green plants, picture frames, and small ornaments. Change the lighting to warm, natural side light with soft shadows. Adjust layout: Move the tower fan, originally positioned to the right, to the center of the image, and scale it down and move it lower overall so that the base rests on the wooden floor. Remove text: Remove all advertising text in the upper left, including 「轻巧摆放节省空间」「加长 风道 细窄栅格」 and the three feature labels 「三种风类」「柔风低噪」「懒人福音」. Remove the gold circular 「本店热销」 badge in the lower left. Adjust product: Reduce the overall size of the OMO boxed laundry detergent that originally almost fills the frame, move it slightly upward and to the right, and leave foreground space at the bottom. Add props: Add a row of lemons, lemon slices, and green leaves in the foreground at the bottom of the product, partially covering the lower edge of the box to create a fresh fruit-surrounding effect. Add decoration: Add a few small semi-transparent bubbles near the upper-left area of the product to enhance the fresh feel. Replace background: Replace the original close-up warm-toned blurred background with a bright outdoor scene, with a light wooden tabletop at the bottom and blurred blue sky, white clouds, and distant mountains in the background. Adjust product: Change the single/partial peach that originally filled the frame into a display of multiple whole honey peaches piled together, while retaining the texture of some peaches with peeled- open flesh. Add props: Add a vintage wooden carrying basket/slatted crate around the peaches, with the peaches piled inside the crate, and the wooden crate featuring a slatted structure and a handle on top. Add limb: Add a hand entering from the upper left of the image and holding the wooden crate handle, with the arm entering the frame from the left side. Replace model pose: Change the woman who was originally bending over behind the sofa to arrange the cushions into a resting pose lying on the sofa seat cushion, with her head at the original left-side cushion position and her body extending horizontally to the right, wearing light beige loungewear and a relaxed expression. Adjust cushions: Change the two original brown cushions from standing upright in the middle to one lying flat at the left end of the sofa as a head pillow, and move the other to the front of the right- side backrest. Replace background: Replace the entire gray-white gradient studio background in Figure 1 with a warm brown wooden interior scene, add a wooden tabletop at the bottom, and use a blurred wooden plank wall as the background. Adjust layout: Change the product from being displayed large on the right side of Figure 1 to a smaller size placed in the center of the wooden table in the lower-middle area of the image, while keeping the title area at the top left. Remove text: Remove the small orange text at the top, 「助力消化吸收」,Remove the three feature boxes on the left, 「亮泽被毛」「燃脂配方」「塑形营养」,Remove the golden circular badge at the lower left, 「品质好物」. Keep the main title 「富含益生元 双肩发达」 in the original upper-left title area, change it to pure black bold text with no background, and move it slightly upward. Remove background decoration: Remove the horizontal light green smoke/light streak effect behind the product, restoring the background to a clean, bright curtain and tabletop scene. Adjust layout: Scale down the originally oversized main product overall and move it slightly down and to the right, freeing up space for text in the upper-left corner, while keeping the bottom on the tabletop area. Add text: Add two lines of large dark brown bold text 「柔声轻音」「精致小巧」 in the blank curtain area at the upper left of the original image, with no backing panel. Replace background: Replace the off-white abstract water-ripple rings and transparent liquid texture at the bottom in Image 1 with a warm beige nursery scene, with a crib on the left, a wooden cabinet and green plant on the right, and a light wood tabletop added at the bottom. Adjust layout: Change the product from its original diagonal placement in the lower half of the image to an upright position in the central area of the original product placement, with the bottom resting on the wooden tabletop. Adjust text: Keep the top main title content unchanged, change its color from brownish orange to a brighter orange-yellow, and keep its position in the original top area. Figure 6: Qualitative examples of CTR-driven ad image modification.