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.
Figures & tables
Figure 1: Comparison of Existing Advertising Creative Optimization Paradigms.
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.
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.
Method
CTR 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%
Table 1: Main online results in terms of CTR gain.
Setting
CTR Gain ↑
Product Category
Home Appliances
+14.01%
Computers
+11.61%
Beauty
+11.37%
Digital Products
+3.92%
Personal Care
+3.28%
Table 2: CTR gains across different product categories and advertising scenarios.
Method
LAION
Predictor
PickScore
VLM
Avg. ↓
ECC
2.47
2.41
2.36
2.97
2.55
DPO S2
2.53
2.34
2.62
3.01
2.63
DPO S1
2.61
2.45
2.64
2.92
2.66
DPO
3.34
2.59
3.20
2.98
3.03
Gemini
3.66
3.72
2.95
2.73
3.27
GPT
3.72
3.70
3.40
2.78
3.40
Table 3: Average rankings under four aesthetic evaluators.
Figure 4: Qualitative examples of CTR-driven edits from original images on the left to optimized images on the right.
Method
Δc→r↑
Δc→s↓
Δs→r↑
Method Comparison
DPO
0.0213
0.0210
0.0003
DPO S2
0.0316
0.0126
0.0191
ECC
0.0316
0.0119
0.0197
Ablation on Regularization Weight
ECC ( λ=0.75 )
0.0216
0.0121
0.0104
Table 4: Order robustness under DPO implicit rewards.
Model
Input Setting
Image Only
+51 Attr.
+28 Attr.
GPT
47.58
49.50
49.86
Gemini
48.93
48.24
49.44
Ours
58.49
59.46
59.56
Table 5: Pairwise CTR prediction accuracy (%) under different input settings.
Variant
CTR Gain
DPO
+0.68%
DPO S1
+1.12%
DPO S2
+1.29%
ECC w/ CPO
+6.86%
Table 6: Ablation results on preference optimization.
Attribute
Label Space / Candidate Values
Composition and Layout
Image Aspect Ratio
Strict: 1:1, 16:9, 9:16, 3:4, 4:3, 3:5, 4:5.
Composition Symmetry
Strict: central symmetry, rule of thirds, horizontal symmetry, vertical symmetry, asymmetric balance, diagonal composition, radial composition.
Element Density
Strict: sparse, medium, dense.
Visual Balance
Extensible: balanced, cluttered, visually unbalanced, or other layout balance descriptions.
The production scenario enforces square advertising creatives, making aspect ratio uninformative for CTR attribution.
Copy Content
On-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 Alignment
A single global alignment label is unreliable because different copy blocks within the same creative may follow different alignment patterns.
Character Spacing
Character spacing is visually subtle in most creatives and is difficult to extract consistently at large scale.
Font Size
Font size is highly relative to image scale, resolution, and layout context, making stable automatic annotation difficult across heterogeneous creatives.
Font Type
This attribute substantially overlaps with Font Style & Effects and may introduce redundant statistical factors.
Table 8: Pre-excluded visual attributes and their exclusion rationale.
Attribute
p -value
RF Contr. (%)
Selected
Font Style & Effects
<0.001
9.45
Yes
Background Color Tone
0.001
8.29
Yes
Copy Density
<0.001
6.28
Yes
Visual Style
<0.001
6.08
Yes
Shooting Angle
0.019
6.04
Yes
Composition Symmetry
<0.001
4.90
Yes
Table 9: Statistical analysis of candidate visual attributes after preliminary screening.
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.
Figure 6: Qualitative examples of CTR-driven ad image modification.
Generating realistic and user-preferred advertisements is a key challenge in e-commerce. Existing approaches utilize multiple independent models driven by click-through-rate (CTR) to controllably create attractive image or text advertisements. However, their pipelines lack cross-modal perception and rely on CTR that only reflects average preferences. Therefore, we explore jointly generating personalized image-text advertisements from historical click behaviors. We first design a Unified Advertisement Generative model (Uni-AdGen) that employs a single autoregressive framework to produce both advertising images and texts. By incorporating a foreground perception module and instruction tuning, Uni-AdGen enhances the realism of the generated content. To further personalize advertisements, we equip Uni-AdGen with a coarse-to-fine preference understanding module that effectively captures user interests from noisy multimodal historical behaviors to drive personalized generation. Additionally, we construct the first large-scale Personalized Advertising image-text dataset (PAd1M) and introduce a Product Background Similarity (PBS) metric to facilitate training and evaluation. Extensive experiments show that our method outperforms baselines in general and personalized advertisement generation. Our project is available at https://github.com/JD-GenX/Uni-AdGen.
Recent advances in generative AI have substantially accelerated the creation of high-quality ad creatives, dramatically expanding the number of candidate variants per campaign. This shift increases the need for scalable dynamic creative optimization (DCO) systems that can match creatives to the most relevant audiences under stringent latency and cost constraints. We present PinDCO, a production DCO system for ad creative retrieval and selection on Pinterest, a billion-scale visual discovery platform. PinDCO is built around a Creative Component Fusion Network (CCFN) that performs dynamic creative scoring by modeling each creative component (e.g., image, title, layout) with a dedicated tower, using component-specific hyperparameters to account for differing modeling complexity. The component representations are fused to predict a creative-level score conditioned on the ad-level prediction, and we improve training data quality via an exploration-exploitation strategy. To account for Pinterest's waterfall grid layout, where a creative's rendered size affects nearby content and session-level engagement, we introduce a Pixel-aware Adjustment Module(PAM) that adjusts scores based on creative size to encourage efficient screen real-estate utilization and better whole-page outcomes. To support the large volume of creative candidates, we further employ a lightweight pre-selection model for early pruning, and optimize serving efficiency through caching and dynamic batching. Extensive offline analyses and online A/B experiments demonstrate the effectiveness of PinDCO, yielding a +3.09% lift in ad Click-Through Rate(CTR) with positive whole-page metrics. With the strong performance, we launched PinDCO in the Pinterest Ads platform.
Diffusion models enable multiple advertisers to compete for exposure within a shared generated image. Yet existing approaches are computationally costly when generating images under varying target-share configurations, and it remains unclear how the resulting outcomes should be evaluated and priced. To address these gaps, we propose DiffAuction, a framework integrating controllable generation, click-intent evaluation, and auction design. First, our training-free Residual Share Calibration (RSC) decomposes joint-attention updates into product-specific residuals and rescales them according to target shares. We further introduce Share-Conditioned Prompting (SCP) to adapt the viewpoint to each desired allocation. Second, we develop an evaluator that uses a VLM to extract visual evidence and an LLM to assess Click-Through Intent (CTI), a proxy for click propensity, along four dimensions: Product Prominence, Brand Recognizability, Scene Integration, and Contextual Value Communication. We validate its assessments against judgments from 10 trained annotators on 1000 images. Third, we formulate this advertising setting as an auction over jointly feasible generation configurations. Each configuration has a public expected CTI vector, while advertisers hold configuration-specific private values per unit of CTI that map click intent into commercial value. We adopt a learnable affine maximizer auction (AMA), which jointly selects a shared generation configuration and determines advertiser-specific payments to maximize expected platform revenue. The resulting allocation and payment rules guarantee dominant-strategy incentive compatibility (DSIC) and individual rationality (IR). Experiments show that our generation method strengthens the monotonic relationship between target shares and CTI relative to baselines, while the learned AMA achieves higher platform revenue in the resulting auction setting.
Yijie Su, Ruiyang Si, Tonghan Wang
Tongji University · Tsinghua University · Shanghai Qi Zhi Institute