Ad creatives stop being redrawn whole: editing only the attributes that drag clicks lifts CTR 6.86% over 20 million impressions

Ad image optimization no longer treats the whole creative as one unit to be redone; it edits only the visual attributes that hurt click-through rate and keeps the ones that help. Earlier methods labeled a prompt describing the entire ad image as positive or negative, so every visual attribute in an underperforming image was penalized equally, including attributes that were actually good for clicks. In a one-week online experiment, edited creatives raised aggregate CTR by 6.86% over the originals, while the CTR-oriented CAIG method reached only 0.20%, and GPT and Gemini were surpassed by 3.61 and 1.34 percentage points.

The system first finds which visual attributes actually correlate with click-through rate, then teaches a model to emit executable editing suggestions, and finally uses real click feedback to keep the suggestions that work. Attribute selection compresses a vast editing space into a structured set of actionable dimensions instead of letting the model generate freely over the whole image. From more than 550,000 online ad images, Qwen3-VL-32B extracted 51 predefined visual attributes; ANOVA and Random Forest contribution analysis narrowed these to 28 click-sensitive attributes, which supported 101,699 instruction-tuning samples and 68,704 preference pairs.

The preference stage adds a consistency constraint so that semantically identical editing suggestions receive the same preference even when their operations are shuffled. Standard DPO tends to overfit to superficial ordering patterns in the suggestions rather than to the editing semantics itself. On about 1,000 editing samples, ECC reached an order-consistency margin of 0.0197, above DPO's 0.0003 and DPOS2's 0.0191, while keeping strong chosen-versus-rejected discrimination.

The report covers one concrete setting: e-commerce ad images. The system edits existing creatives at fine granularity rather than regenerating a whole creative from a product image. The online experiment ran for one week across 44 product categories and accumulated about 20 million impressions, and edited images were reviewed by professional practitioners for visual quality and foreground-background compatibility before deployment. The metric is the impression-weighted aggregate CTR gain between edited and original creatives, reported separately for categories such as home appliances, computers, and beauty and for placements such as the order page and product detail page. Readers interested in generic image editing or non-e-commerce creative settings are outside the demonstrated scope.

Attribute extraction relies on Qwen3-VL-32B, and 14 attributes were excluded before statistical analysis for unstable extraction, subjective semantics, or skewed distributions, so the retained 28 still carry those screening judgments. The preference pairs come from real exposure and click feedback, yet the report notes that predicting CTR from images remains difficult, with prior reward models reaching only about 58% accuracy, so the preference signal is not perfectly clean. The four aesthetic evaluators disagree in their rankings and do not fully align with the online CTR gains, indicating that aesthetic quality and advertising effectiveness are not equivalent. In the pairwise attribute-validation task, the best accuracy with the 28 attributes was 59.56%, a limited margin above chance.