Creative4U: Explainable Ad Creative Selector
- Creative4U is an explainable ad creative selector that uses a natural language generation task to compare creatives based on product titles and user queries.
- It leverages the CreativePair dataset and a 10-step evaluation protocol to generate detailed, stepwise comparative rationales for selecting superior ad creatives.
- Empirical results demonstrate improved selection accuracy and explanation quality, with significant gains in key ad performance metrics like CTR and RPM.
Creative4U is an MLLMs-based advertising creative image selector with comparative reasoning, introduced as a system for explainable creative assessment and selection in e-commerce advertising (Lin et al., 18 Aug 2025). It addresses a setting in which advertisers can produce large quantities of creative images at minimal cost but still face the harder problem of deciding which image is superior for a particular product and user intent. Rather than treating the problem as conventional creative ranking, Creative4U integrates assessment and selection into a natural language generation task: given two candidate creative images, a product title, and user queries, it generates a stepwise comparative rationale and a final choice. Its core components are a comparative reasoning protocol, the CreativePair dataset, and a two-stage post-training procedure called Reason-to-Select RFT (Lin et al., 18 Aug 2025).
1. Problem Setting and Task Definition
Creative4U is motivated by a specific limitation in prior advertising systems: existing methods primarily focus on creative ranking, which fails to address the need for explainable creative selection (Lin et al., 18 Aug 2025). In the paper’s framing, the proliferation of AIGC increases the supply of candidate creatives, but selection remains consequential because the objective is not merely visual quality in isolation; it is to choose the creative image most likely to receive a user’s click for a given product and search intent.
The task is formulated as
where and are the two candidate images, and provide textual context, is the chain-of-thought reasoning, and is the final answer indicating which image is superior (Lin et al., 18 Aug 2025). This formulation is notable because it binds decision-making and explanation into a single output space.
A central implication is that Creative4U is not designed as a generic vision-language comparison model. It is a domain-specific selector that conditions on product title and high-frequency user queries reflecting real search intent, and it is explicitly optimized to produce protocol-following comparative judgments rather than unstructured preferences (Lin et al., 18 Aug 2025).
2. CreativePair and the Creative Evaluation Protocol
The empirical foundation of Creative4U is CreativePair, described as the first comparative reasoning-induced creative dataset (Lin et al., 18 Aug 2025). The dataset contains 8,817 human-annotated image pairs, with a final filtered split of 7,088 training samples and 1,729 test samples. Each sample includes two creative images for one product, a high-frequency product query, the product’s title, and a binary label indicating which image had significantly better CTR, with the filtering criterion specified as more than 60% difference and both images having more than 1,000 impressions (Lin et al., 18 Aug 2025).
The annotation process has three stages. First, pairs are strictly filtered from real-world ad data to ensure significance and confidence. Second, manual annotation is carried out by trained designers using a detailed protocol. Third, stepwise annotations are converted into a natural language chain-of-thought via a teacher MLLM, Qwen2.5-VL-72B (Lin et al., 18 Aug 2025). This combination of real traffic data, expert annotation, and model-assisted verbalization is central to the system’s explainability objective.
The Creative Evaluation Protocol is a 10-step process for pairwise comparison. Q1–Q2 perform similarity checks, including whether the images are identical. Q3–Q4 evaluate image-text consistency with respect to user query and product title. Q5–Q9 assess image quality attributes, including text in image, use of models or props, layout, product subject, and background design. Q10 asks the decisive question: “Which image is the user more likely to click on?” (Lin et al., 18 Aug 2025). Each step requires a comparative answer such as “A > B”, “A = B”, or “A < B”, together with reasoning.
This protocol matters because it operationalizes explainability as structured comparative assessment rather than post hoc justification. A plausible implication is that the protocol constrains the model to reason along advertising-relevant axes instead of relying on latent heuristics that are difficult to audit.
3. Model Architecture and Comparative Reasoning Behavior
Creative4U is built on Qwen2.5-VL-7B-Instruct and is presented as a comparative reasoning pipeline rather than a scalar scorer (Lin et al., 18 Aug 2025). Its inputs are pairs of candidate creative images, contextual product title, and high-frequency user queries; its outputs are a stepwise chain-of-thought followed by a final selection. The model is therefore used as a selector with explicit rationale generation rather than as a black-box CTR regressor.
The paper describes a high-level flow with four stages: systematic modeling of creative images leads to the Creative Evaluation Protocol; CreativePair supplies pairwise human annotations; Reason-to-Select RFT trains the MLLM to generate both selection and reasoning; and the deployed model produces the best creative together with an explicit justification (Lin et al., 18 Aug 2025). This makes the system’s “interface” inherently linguistic and comparative.
The emphasis on users’ interests is operationalized through the query field. The query is not ancillary metadata: it is part of the decision context, allowing the same pair of images to be interpreted relative to actual search intent. This distinguishes Creative4U from systems that assess creatives only in relation to product metadata or global visual quality.
4. Reason-to-Select RFT
Reason-to-Select RFT is the paper’s two-stage post-training method. The first stage is supervised fine-tuning with chain-of-thought, or CoT-SFT. Its purpose is to teach the MLLM to generate detailed, domain-specific, stepwise reasoning together with the final predicted answer, using next-token prediction over the concatenation of comparative reasoning and answer (Lin et al., 18 Aug 2025). The prompting scheme instructs the model to respond question-by-question according to the Creative Evaluation Protocol.
The second stage is GRPO-based reinforcement learning. In this stage, for each input the model samples a group of responses, evaluates each response with a reward that combines protocol compliance and answer correctness, normalizes rewards into relative advantages, and updates the policy with grouped optimization and KL regularization relative to the SFT reference policy (Lin et al., 18 Aug 2025). The paper characterizes this as encouraging correct, protocol-compliant, and accurate reasoning while also promoting diversity and robustness.
The technical novelty claimed for R2S-RFT is threefold: fine-grained protocol-guided comparative reasoning, two-level reward shaping over format and accuracy, and grouped sampling for stable RL optimization (Lin et al., 18 Aug 2025). The result is not simply a stronger classifier. It is a selector trained to preserve the structure of explanation as part of the optimization target.
An important nuance emerges from the ablation results: explainability and raw accuracy are not identical objectives. The summary states that pure answer-only SFT reaches the highest raw accuracy, 76.46%, but provides no explainability; CoT-SFT improves Qwen-Score; and R2S-RFT delivers the best explainability with strong accuracy at 74.66% (Lin et al., 18 Aug 2025). This is a substantive design trade-off rather than an incidental observation.
5. Empirical Results
Creative4U is evaluated against VAM, Co-Instruct, DepictQA-Wild, and Q-Insight on CreativePair (Lin et al., 18 Aug 2025). Two headline metrics are reported: selection accuracy and Qwen-Score, where the latter reflects the helpfulness, relevance, and detail of the generated reasoning.
| Method | Accuracy (%) ↑ | Qwen-Score ↑ |
|---|---|---|
| VAM | 61.94 | — |
| Co-Instruct | 50.20 | 46.19 |
| DepictQA-Wild | 55.59 | 49.09 |
| Q-Insight | 72.93 | 58.50 |
| Creative4U | 74.66 | 71.03 |
These results show that Creative4U leads both on correct selection and on explanation quality (Lin et al., 18 Aug 2025). The gap in Qwen-Score is especially notable because it indicates that the model’s comparative rationales are materially more useful than those of the baselines, not merely longer or more stylistically polished.
The reported online evidence is also consequential. The summary states that large-scale online experiments were run on a leading e-commerce platform, with CTR, CVR, and RPM as key metrics. It reports gains of “up to +6.28% CTR, +6.97% RPM, +2.03% CVR” and “+3.63% CTR, +3.84% RPM, +1.46% CVR even when paired with the production online ranker” (Lin et al., 18 Aug 2025). These figures are presented as evidence that explainable selection is not only diagnostically useful but also operationally competitive.
A common misunderstanding would be to treat Creative4U as a ranking model with added verbalization. The ablation evidence argues against that interpretation: answer-only SFT attains higher raw accuracy but loses the system’s defining property, whereas R2S-RFT preserves structured explanation and still remains strong on selection accuracy (Lin et al., 18 Aug 2025).
6. Relation to Adjacent Creative-AI Systems
Creative4U occupies the creative selection layer of a broader advertising and design pipeline. In online advertising, CG4CTR optimizes for CTR directly during the creative generation stage by integrating Stable Diffusion inpainting with LoRA, a Prompt Model for individualized prompt selection, and a Reward Model for multimodal creative ranking inside a self-cyclic pipeline (Yang et al., 2024). Creative4U addresses a different problem: explainable assessment and pairwise selection conditioned on user interests and comparative reasoning.
In graphic design generation, CreatiDesign targets multi-condition control over images, layouts, and texts through a unified multi-condition driven architecture with a multimodal attention mask mechanism, and it is trained on a dataset with 400,000 fully annotated graphic design samples (Zhang et al., 25 May 2025). CreatiPoster, by contrast, focuses on editable, multi-layer compositions using a protocol model that outputs a JSON specification for text and asset layers, together with a conditional background model and a copyright-free corpus of 100,000 multi-layer designs (Zhang et al., 12 Jun 2025). These systems are concerned with controllable creation and editability rather than advertising-specific comparative selection.
This suggests that Creative4U can be understood as complementary to generative creative systems. A plausible implication is a pipeline in which generation systems produce candidate creatives, design systems preserve editability and compositional control, and Creative4U performs user-aware, explainable final selection. Within that stack, Creative4U’s principal contribution is to convert creative choice from opaque scoring into protocol-governed comparative reasoning (Lin et al., 18 Aug 2025).