acceptodds
Under review as a conference paper at ICLR 2027

AdsPrompter: Bridging User and Designer Preferences for Multimodal Advertising Design Generation

Abstract

Generative image modeling has shown remarkable advances recently, yet generating advertising designs with multimodal elements remains an enduring challenge. Existing works optimize either design aesthetics according to design criteria or user behaviors measured by click-through rates, overlooking the inherent discrepancy between user and designer preferences. We observe that visually appealing advertisements do not necessarily yield high user engagement, while solely maximizing user engagement may negatively affect brand perception. To bridge this gap, we introduce AdsPrompter, a new framework that aligns designer intent with user engagement for multimodal advertisement generation. Given a product image and taglines as inputs, AdsPrompter learns to create aesthetically pleasing advertising designs of varying sizes while faithfully preserving the product semantics, textual messages, and intended engagement goals. At the core of our framework is User-Designer Preference Optimization (UDPO), a coarse-to-fine learning strategy with complementary reward models for prompt refinement. The refined prompt guides the inputs to a diffusion-based model for preference-aligned advertisement generation. Furthermore, we introduce a large-scale advertisement dataset with 200k samples featuring designer and user preferences, along with a comprehensive benchmark. Experimental results show that AdsPrompter consistently outperforms existing methods in user engagement and design aesthetics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Graphic Design Generation, Text-to-Image Generation, Vision Language Models

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_SjshOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes AdsPrompter, a framework aimed at generating multimodal advertising designs that simultaneously balance visual design aesthetics and commercial effectiveness. To bridge the gap between their conflicts, the authors introduce UDPO, a coarse-to-fine learning strategy. In the coarse stage, a prompt generator (SFT on Qwen3-VL-8B) and an image renderer (full fine-tuning on a DiT model) are trained. In the fine stage, the prompt generator is refined via policy optimization using two complementary reward models: a User Reward Model (capturing CTR preferences via pairwise ranking) and a Designer Reward Model (capturing visual layout and style using DINOv2 with ordinal loss). The framework employs an optimization algorithm to dynamically balance the two reward objectives. The authors also introduce a dataset of 200k product-advertisement pairs along with 10k CTR preference pairs. Experiments show improvements over baseline methods across both aesthetic and commercial metrics.

Critical strengths
  1. Important problem: Addressing the discrepancy between commercial performance (CTR) and visual aesthetic quality (design principles) in multimodal ad generation is a highly relevant, real-world research problem;

  2. Data contribution: the creation of a 200k multimodal dataset with structured prompt protocols and real-world CTR preference annotations represents a valuable contribution to the community.

Critical weaknesses
  1. The optimization pipeline requires training multiple large components (a 30B multimodal URM, a DINOv2 DRM, a full DiT fine-tune, and GRPO on an 8B VLM). The computational demand is significant (e.g., hundreds of GPU hours), which may limit accessibility and reproduction for resource-constrained groups.

  2. The proposed optimization algorithm, i.e., MDGA with EMA, is heuristic and slightly over-engineered. In stochastic RL settings (like GRPO), MGDA updates derived from mini-batch policy gradients are noisy and sensitive to gradient magnitudes rather than true objective priority. The addition of periodic updates and EMA smoothing essentially acts as an empirical patch to stabilize MGDA's inherent variance in policy gradient training. An ablation comparing MGDA+EMA against simple running normalized scalarization would clarify if MGDA's mathematical overhead actually delivers a meaningful gain over simpler dynamic heuristics.

Questions for authors

How sensitive is the MGDA weight during policy optimization? Does it stabilize quickly, or does it oscillate significantly across different product domains?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Sjsh

We thank the reviewer for recognizing the practical importance of jointly considering design quality and user engagement, as well as the value of the dataset. We sincerely appreciate the constructive feedback and provide the following responses to address the concerns raised. We will be happy to clarify any further concerns.

Q1: The optimization pipeline requires training multiple large components...which may limit accessibility and reproduction for resource-constrained groups.

A1: As we focus on a practical problem of high-quality advertisement generation, we agree with the reviewer that the complete training pipeline requires substantial computational resources. Such costs are common in the adaptation of large generative models, particularly with RL-based optimization involving repeated candidate generation and reward evaluation.

It is worthnoting that at inference time, our model can run on a single RTX 4090 GPU, taking approximately 40–50 seconds to generate an advertisement. We will release both the code and the pre-trained model to enable inference and further adaptation without repeating the complete training pipeline.

Q2: The proposed optimization algorithm is heuristic and slightly over-engineered...An ablation comparing MGDA+EMA against simple running normalized scalarization would clarify if MGDA's mathematical overhead actually delivers a meaningful gain over simpler dynamic heuristics.

A2: Thanks for the valuable suggestion. The motivation for MGDA+EMA in the proposed UDPO is to balance the optimization between the designer and user policy gradients. The fusion weight is updated every 50 steps, and the two policy gradients are computed at those update steps.

We first compare the proposed strategy with several fixed-weight strategies and uncertainty modulation in the table below, showing that the proposed adaptive weighting strategy generally outperforms the fixed-weight configurations.

MethodVLS ↑TLS ↑APS ↑STS ↑DRM ↑Con ↑Sen. Acc ↑URM ↑
Designer/User weight (0.3/0.7)8.4477.1987.8748.3870.7120.9260.8670.736
Designer/User weight (0.5/0.5)8.5187.2267.9808.4430.7370.9280.8660.731
Designer/User weight (0.7/0.3)8.5327.2917.9628.4520.7760.9250.8640.694
w/o uncertainty modulation8.4967.2387.8828.4110.7540.9230.8620.718
Adaptive Weighting (Ours)8.5617.3567.9718.4690.8210.9290.8700.742

We further conduct an ablation study without EMA (w/o EMA) to investigate the effect of smoothing. The table below records the fusion weight under the two settings, together with the standard deviation and range of ww over the last 1000 steps. We can see that EMA reduces the observed fluctuations of the applied fusion weight and keeps the optimization objective stable during training.

Training step050010001500200025003000Std (last 1000 steps) ↓Range (last 1000 steps) ↓
w/o EMA0.5000.6630.5480.6520.5710.6410.5560.0380.112
MGDA+EMA (Ours)0.5000.5760.5980.6090.6090.6120.6100.0010.003

We have added the results and discussion in the revised PDF (see Table 3 and L480-L499).

Q3: How sensitive is the MGDA weight during policy optimization? Does it stabilize quickly, or does it oscillate significantly across different product domains?

A3: The EMA-smoothed MGDA weight ww is initialized to 0.5, re-solved by MGDA every 50 steps, and smoothed by EMA with a coefficient of 0.9. During policy optimization, ww stabilizes quickly from its initial value and converges without significant oscillation.

As suggested, we further analyze the user and designer weights across different product domains. For five product categories in our dataset, we compute the agreement rate, i.e., the proportion of CTR preference pairs in which the higher-CTR advertisement also receives the higher designer score.

CategoryWomenswearSnacksAppliancesBeautyHome
Agreement rate0.640.620.640.630.61

As shown above, the preferences across these coarse-level product domains do not vary a lot. We believe that considering fine-grained product categories could be an interesting future work in this direction.

Reviewer_6xFHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AdsPrompter generates advertisements from product images, taglines, and target canvas sizes by learning prompts for a diffusion-based renderer. The framework first performs supervised training of a prompt generator and an advertisement renderer, then refines the prompt policy using complementary user and designer reward models. The user reward learns from same-product CTR preference pairs, while the designer reward predicts aesthetic quality. User-Designer Preference Optimization (UDPO) combines their advantages through an adaptive MGDA-based weight. The paper also introduces a corpus of approximately 200K product-advertisement pairs and over 10K CTR preference pairs, with a 5K-condition evaluation benchmark. Evaluation includes automatic metrics, independent VLM judges, reward-model validation, a human preference study, and online CTR measurements.

Critical strengths
  • S1. A useful problem formulation supported by substantial application data. Jointly considering design quality and click preference is well motivated for advertising, where visual appeal alone does not establish commercial effectiveness. The product-conditioned prompt-to-image pipeline addresses product preservation, promotional text, and layout together. The corpus described in Section 4 combines advertisement reconstruction data with aesthetic supervision and observed same-product CTR preferences, providing a useful basis for studying this problem.

  • S2. Evaluation extends beyond the optimized reward scores. Tables 1 and 2 report improvements in text accuracy, product consistency, and design metrics. Appendix C adds evaluations using two external VLM judges and validation of the reward models, including 0.824 URM pairwise accuracy and DRM correlations with professional designers' ratings (Table 6). Human evaluation and live CTR testing are also valuable additions. These are meaningful strengths, although the reporting issues below limit how confidently I can interpret the independent evidence.

Critical weaknesses

W1. The independent human and online evaluations need reconciliation

  • Evidence: In Section 5.3, lines 455-461, the full model is reported to win 61.2% against the baseline and 63.5% against the model without RL. Figure 7 instead displays 83.2% and 66.9% for those comparisons. The paper describes 200 comparison pairs and 30 participants, but does not explain the vote aggregation or the unit used in the binomial tests. For online evaluation, Appendix B.6 describes three disjoint traffic buckets, whereas Table 2 reports five variants and Table 5 reports nine methods. The assignment of methods to buckets or separate experimental rounds is not specified. The full model's CTR is also 5.51% in Table 2 and 5.439% in Table 5 without an explanation of whether these come from different experiments. Although a proportion z-test is mentioned, the tables do not provide comparison-specific uncertainty or test results.
  • Impact on recommendation: Human preference and live engagement are central evidence that the method improves outcomes beyond its learned reward functions. These discrepancies leave the effect sizes and comparability of the experiments unclear. They may be resolvable reporting issues, but a reconciled account of the human votes, online experimental arms, and statistical uncertainty is necessary to assess the claimed improvements.

W2. The benefit of the proposed adaptive optimization is not isolated

  • Evidence: Section 3.3 introduces uncertainty-modulated aesthetic advantages and MGDA-based adaptive reward fusion. Table 2 compares supervised training, single-reward variants, and the full model, which supports using both rewards, but it does not compare adaptive fusion against fixed-weight combinations under the same training budget. I could not find an ablation of the uncertainty modulation in the supplied PDF. The comparison among alternative prompt generators under a shared renderer is qualitative only (Appendix D.2 and Figure 12). Moreover, the text accompanying Equation 14 introduces an uncertainty-modulated advantage without specifying how the uncertainty in Equation 9 transforms the standardized aesthetic advantage. Neither Section 3.3 nor Appendix A.2.3 specifies the MGDA update interval or EMA coefficient.
  • Impact on recommendation: The results support the overall trained system, but do not establish how much the specific UDPO design contributes beyond supervised adaptation and joint reward optimization with a simpler weighting rule. This limits the methodological contribution, especially since both the prompt generator and renderer receive domain-specific training. A controlled comparison using the same supervised checkpoints, renderer, and rollout budget, together with a complete optimization specification, would materially strengthen the case.
Questions for authors
  1. Clarification of W1. Which human win rates are correct, and how are votes aggregated across participants and image pairs for the reported tests? Please also map the methods in Tables 2 and 5 to the online buckets or experimental rounds, explain the two full-model CTR values, and provide per branch impressions/clicks and confidence intervals for the relevant differences. A consistent account demonstrating reliable gains would substantially reduce this concern.

  2. Evidence and implementation details for W2. With the supervised prompt checkpoint, renderer, data, and rollout budget held fixed, how does UDPO compare with equal-weight and validation-tuned fixed-weight reward fusion, and with uncertainty modulation removed? Please define the uncertainty transformation, MGDA update frequency, and EMA coefficient. Evidence that these choices improve the design-engagement trade-off would materially strengthen my assessment of the technical contribution

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6xFH (Part 1)

Thanks a lot for your time and constructive feedback. We are encouraged by your recognition of the problem formulation, substantial data, and extensive evaluations of our work. We provide point-by-point responses below, and have also incorporated the corresponding changes into the revised PDF. We will be happy to clarify any further concerns.

Q1: In Section 5.3, lines 455-461, the full model is reported to win 61.2% against the baseline and 63.5% against the model without RL...Which human win rates are correct, and how are votes aggregated across participants and image pairs for the reported tests?

A1: Sorry for the confusion. Our user study involves two subgroups, i.e., 24 novice users and 6 professional designers. The two sets of rates correspond to different participant groups within the same study. The original Figure 7 reports preferences from 6 professional designers, whereas Section 5.3 reports results across all 30 participants. The group-specific and complete results of the user study are shown below. Each percentage denotes preference for the first method over the second.

GroupOurs vs. BaselineOurs vs. w/o RLw/o RL vs. Baseline
Professional Designers (6)83.2%66.9%76.4%
Novice Users (24)55.7%62.7%59.3%
All Participants (30)61.2%63.5%62.7%

For each comparison (Ours vs. Baseline, Ours vs. w/o RL, or w/o RL vs. Baseline), participants viewed the same input and two generated advertisements side by side, with randomized left/right placement. They selected the preferred advertisement based on textual fidelity, visual quality, layout design, and commercial appeal. Each of the three method comparisons contains 200 image pairs, with every pair evaluated by all participants. The win rate is calculated as the number of valid judgments favoring the first method divided by the total number of valid judgments for that comparison.

We have revised Figure 7 to show the results of all participants, and added the detailed user study process and the group-specific results (Figure 12) in the Appendix (see Appendix B.7).

Q2: For online evaluation, Appendix B.6 describes three disjoint traffic buckets, whereas Table 2 reports five variants and Table 5 reports nine methods...Please also map the methods in Tables 2 and 5 to the online buckets or experimental rounds, explain the two full-model CTR values.

A2: The online A/B tests were conducted on a large-scale E-commerce platform to validate the CTR performance of the generated ads. The three traffic buckets do not correspond one-to-one to methods. They are disjoint traffic partitions on the platform to allocate experimental traffic and expand exposure coverage. Each bucket contains all methods evaluated in the corresponding round.

Note that the platform supports at most ten creatives per product in a single exposure experiment. Table 2 and the original Table 5 (Table 6 in the revision) report two separate online A/B testing rounds with different objectives. The goal of Table 2 is to evaluate the effectiveness of individual components of our work, while Table 5 compares our work with external methods. The full-model CTRs were measured in different deployment windows and are therefore reported separately. Our comparisons are made within each experiment. Our conclusions are based on differences between methods evaluated during the same round and across the same traffic partitions, rather than on comparisons of absolute CTR across rounds.

We have clarified the details of the online A/B tests in the Appendix (see Appendix B.6 and C).

Q3: Although a proportion z-test is mentioned, the tables do not provide comparison-specific uncertainty or test results...Provide per branch impressions/clicks and confidence intervals for the relevant differences.

A3: Thanks for the valuable suggestion. For the online A/B tests, each creative receives 500–600 impressions, and each bucket accumulates more than 1.2M impressions and 60K clicks every half day. The impressions, clicks, and CTR of each method, together with the CTR differences, z statistics, and 95% confidence intervals, are shown below, demonstrating that the CTR improvements of the full model are statistically significant.

MethodImpressions (M)Clicks (M)CTR (%)CTR Difference (pp)95% CI (pp)z
Baseline56.12.875.12+0.390[0.382, 0.398]91.62
w/o RL56.22.995.33+0.180[0.172, 0.188]41.90
w/o Designer Reward56.13.065.46+0.050[0.042, 0.058]11.57
w/o User Reward55.02.945.35+0.160[0.152, 0.168]37.02
AuthorsOpenReview ↗

Response to Reviewer 6xFH (Part 2)

Q4: Section 3.3 introduces uncertainty-modulated aesthetic advantages and MGDA-based adaptive reward fusion...How does UDPO compare with equal-weight and validation-tuned fixed-weight reward fusion, and with uncertainty modulation removed?

A4: Thanks for the valuable suggestion. We conduct a controlled comparison against fixed-weight fusion of the two rewards and without uncertainty modulation (w/o uncertainty modulation). Three different fixed-weight combinations, i.e., designer/user weights of 0.3/0.7, 0.5/0.5, and 0.7/0.3, are added under the same training setup. We further remove the uncertainty modulation and keep MGDA and all other settings unchanged.

As shown below, the proposed adaptive weighting strategy outperforms both the fixed-weight configurations and the w/o uncertainty modulation, supporting its benefit in bridging the gap between designer intent and user engagement. We have added the results and discussion in the revised PDF (see Table 3 and L480-L499).

MethodVLS ↑TLS ↑APS ↑STS ↑DRM ↑Con ↑Sen. Acc ↑URM ↑
Designer/User weight (0.3/0.7)8.4477.1987.8748.3870.7120.9260.8670.736
Designer/User weight (0.5/0.5)8.5187.2267.9808.4430.7370.9280.8660.731
Designer/User weight (0.7/0.3)8.5327.2917.9628.4520.7760.9250.8640.694
w/o uncertainty modulation8.4967.2387.8828.4110.7540.9230.8620.718
Adaptive Weighting (Ours)8.5617.3567.9718.4690.8210.9290.8700.742

Q5: The comparison among alternative prompt generators under a shared renderer is qualitative only.

A5: As suggested, we conduct a quantitative comparison among alternative prompt generators under a shared renderer. The experiments are conducted under the same setup as Figure 13 (Figure 12 in the original version). As shown below, our work achieves the best performance on all metrics, indicating that the performance gap mainly comes from the prompt generated by our model.

Prompt GeneratorVLS ↑TLS ↑APS ↑STS ↑DRM ↑Con ↑Sen. Acc ↑URM ↑
GPT-4o8.2987.0327.8148.2360.6840.8970.8220.571
Gemini-38.3176.9867.7928.2050.6720.9060.8040.579
Qwen3-VL-235B-A22B-Instruct8.2847.1627.7688.1930.6630.9040.7710.566
AdsPrompter (Ours)8.5617.3567.9718.4690.8210.9290.8700.742

We have added the results and discussion in the Appendix (see Appendix D.2 and Table 8).

Q6: The text accompanying Equation 14 introduces an uncertainty-modulated advantage without specifying how the uncertainty in Equation 9 transforms the standardized aesthetic advantage...Please define the uncertainty transformation.

A6: We use inverse-uncertainty weighting followed by group-wise re-centering. Conditioned on the product image and the instruction condition, the proposed UDPO strategy generates a group of candidates. For the k-th candidate, URM produces the functional advantage AkfuncA_k^{\mathrm{func}} , while DRM produces the aesthetic advantage AkaesA_k^{\mathrm{aes}} . AkaesA_k^{\mathrm{aes}} is converted into A~kaes\widetilde{A}_k^{\mathrm{aes}} based on the uncertainty modulation with uku_k in Equation 9 as follows.

ρk=1/uk1K∑i=1K1/ui,A~kaes=ρkAkaes−1K∑i=1KρiAiaes\rho_k=\frac{1/u_k}{\frac{1}{K}\sum_{i=1}^{K}1/u_i}, \qquad \widetilde A_k^{\mathrm{aes}}=\rho_k A_k^{\mathrm{aes}}-\frac{1}{K}\sum_{i=1}^{K}\rho_i A_i^{\mathrm{aes}}

Candidates with higher uncertainty receive smaller pre-centering weights. The fused advantage for the k-th sample is then computed using Equation 14 in the original version (Equation 15 in the revision).

We have added the description in the revised PDF to clarify how the uncertainty transforms the standardized aesthetic advantage (see L260-L268).

Q7: Neither Section 3.3 nor Appendix A.2.3 specifies the MGDA update interval or EMA coefficient.

A7: The MGDA update interval is set to 50 steps. The EMA coefficient is set to 0.9, and the initial fusion weight is set to 0.5. We have added the details in the revised PDF (see L481-L483).

Thank you for your rebuttal. It has addressed all of my concerns, and I will keep my original positive rating of 3.

We're glad that we could address your questions and concerns. Thank you very much for your feedback and support for our paper.

Reviewer_8c3POpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2(+1)
Soundness
1
Read review
Paper summary

This paper consider the problem of generating an ad design. The paper argues that both design aesthetics and click through rate are important in ad designs. The authors propose a pipeline that has several sub-components many of which are fine tuned. Ultimately, the paper proposes a method that optimizes both for design criterion and user click through rates. Experimental results compare the proposed method against several baselines.

Critical strengths

Practically useful paper.

Critical weaknesses
  1. The novelty of the paper is limited. It maybe a good applied engineering paper, but the research component seems quite limited. It is a pipeline put together based on well known techniques such as GRPO, Bradley-Terry model etc.

  2. The overall method consists of fine tuning several models as described in the paper. However, how each component was trained, what data was used for it, how the labels were obtained etc. are not discussed in the paper.

  3. While the main part of the paper including the abstract announce a "200K" sized dataset, the authors in the appendix say "we will release a 5K sample test set built with the same protocol". So there is a huge gap between what is claimed in the main paper and what the authors are likely to release later. So, none of the results in the paper are likely to be reproducible. This is a huge red flag.

  4. No statistical significance is given in the results except for the small human study.

  5. Overall writing is poor and many questions are left unanswered.

Questions for authors

As pointed out in the weaknesses section, none of the sections describe what training data was used for fine tuning each component. These details matter and it is a major miss.

In Figure 1, How were DI and UE measured and reported? While the text mentions an aesthetic score and CTR-based engagement score, no further details are given.

In the ablation, removing the User Reward reduces DRM and further removing the Designer reward reduces URM. What is the explanation for this? If we are optimizing two rewards simultaneously, shouldn't removing one in fact improve the other?

In eq. 9, the authors do not say what M is, why we need the cumulative probability and I am not sure where r_{aes} and U_{aes} are used further in the paper. Also, "cumulative probability" would require monotonicity and sum to one, how is the sum to one property enforced?

Both the fine tuned model and the VLM as a judge are from Qwen family of models.

Since the model is optimized for exactly the same metrics that are reported along with a few other metrics, is it really surprising that the model does well on those?

Which of the results in the main paper are from A/B tests? Are the CTRs reported from A/B tests or based on a model estimation?

7 comments
AuthorsOpenReview ↗

Response to Reviewer 8c3P (Part 1)

We thank the reviewer for recognizing the practical usefulness of our work. We address the concerns below and welcome any follow-up questions.

Q1: The novelty of the paper is limited.

A1: Our contributions can be assessed via the motivation, problem formulation, learning strategy, dataset, and experimental findings, which are also recognized by other reviewers.

Advertising images are different from natural images in terms of multimodal elements, layout structures, design rules, and user preferences. Without considering domain knowledge in advertisements, a naive combination of existing components, such as GRPO and Bradley-Terry model, cannot be effective in practical applications.

Our key novelty is that we, for the first time, bridge the gap between user and designer preferences for multimodal advertising image generation. This is a practical but less-explored problem, and we incorporate advertising domain knowledge in the proposed UDPO strategy for coarse-to-fine prompt optimization and advertisement generation. We also construct a large-scale advertising dataset to facilitate research in this direction.

We will release the dataset, code, and pre-trained model to support reproducibility and future research in the community.

Q2: How each component was trained, what data was used for it, how the labels were obtained etc. are not discussed in the paper.

A2: Due to the page limitation, the training procedures and dataset details were described in the Appendix. We summarize the correspondence below.

ComponentDataLabelsAppendix
Prompt Generator (SFT)Product condition (Ip,T,S)(I_p,T,S)Prompts generated by a VLM from designed advertisementsA.1.1, B.1
Image Renderer (Full FT)Triplets (Ip,y,Iposter)(I_p,y,I_{\text{poster}})Target advertisements IposterI_{\text{poster}}A.1.2, B.1
URM (LoRA)CTR preference pairsHigher-CTR creative of the same productA.2.1, B.3
DRM (Ordinal)AdvertisementsVLM scores reviewed by designersA.2.2, B.2
UDPO (GRPO)Product condition (Ip,T,S)(I_p,T,S)Rewards from frozen URM and DRMA.2.3

In particular, the training details of each component were described in Appendix A, including the prompt generator and image renderer (Appendix A.1), the user and designer reward models (Appendix A.2.1 and A.2.2), and the UDPO policy optimization (Appendix A.2.3). Table 4 provides the component-specific training configurations. The data construction details were described in Appendix B, including prompt annotation (B.1), designer score annotation (B.2), CTR preference pairs (B.3), and data quality control (B.5).

Q3: While the main part of the paper including the abstract announce a "200K" sized dataset, the authors in the appendix say "we will release a 5K sample test set built with the same protocol"...

A3: We apologize for the ambiguous description. To support the proposed framework, we construct a large-scale advertisement dataset with 200k samples considering designer and user preferences, along with a comprehensive benchmark (see L82-L84). The 5k sample test set mentioned here is the benchmark. We will release both the training dataset and the evaluation benchmark.

AuthorsOpenReview ↗

Response to Reviewer 8c3P (Part 2)

Q4: No statistical significance is given in the results except for the small human study.

A4: Note that our human study evaluation contains two separate parts. One is the controlled user study, and the other is the large-scale online A/B tests measuring user engagement.

For the controlled user study, we have recruited 30 participants (6 professional designers and 24 novice users) with 200 comparison cases. The scale is comparable to, or larger than, the user studies in recent ICLR papers. For example, TweedieMix (Kwon & Ye, 2025) (20 participants) and CubeDiff (Kalischek et al., 2025) (28 participants). Each of the three method comparisons contains 200 image pairs, with every pair evaluated by all participants. We conduct a two-sided binomial test for the reported preference comparisons, confirming that the improvements in human preference are statistically significant.

For the large-scale online A/B tests, we use 100K products and one week of deployment. Each creative receives 500–600 impressions, and each bucket accumulates more than 1.2M impressions and 60K clicks every half day. The impressions, clicks, and CTR of each method, together with the CTR differences, z statistics, and 95% confidence intervals are shown below, demonstrating that the CTR improvements of the full model are statistically significant.

MethodImpressions (M)Clicks (M)CTR (%)CTR Difference (pp)95% CI (pp)z
Baseline56.12.875.12+0.390[0.382, 0.398]91.62
w/o RL56.22.995.33+0.180[0.172, 0.188]41.90
w/o Designer Reward56.13.065.46+0.050[0.042, 0.058]11.57
w/o User Reward55.02.945.35+0.160[0.152, 0.168]37.02

Q5: Overall writing is poor and many questions are left unanswered.

A5: Sorry for the confusion. We have proofread the manuscript thoroughly and made necessary revisions based on the comments from all reviewers. We would be happy to discuss any remaining issues.

Q6: In Figure 1, how were DI and UE measured and reported?...

A6: The DI and UE shown in Figure 1 are computed by averaging the values of the metrics described in Section 5.1 (reported in Table 1) in terms of Designer Intent and User Engagement. We have revised the caption of Figure 1 (see L60-L63) to specify how designer intent (DI) and user engagement (UE) are measured for the arrows.

Q7: Removing the User Reward reduces DRM and further removing the Designer reward reduces URM...If we are optimizing two rewards simultaneously, shouldn't removing one in fact improve the other?

A7: In fact, this is an interesting observation that highlights the motivation of jointly modeling designer and user preferences in multimodal advertising generation. Although these preferences can diverge, they are not necessarily conflicting for every advertisement. Instead, they are complementary factors that work together to generate ads with high commercial performance and visual aesthetic quality. Readability, clear product presentation, and coherent layouts can benefit both criteria. Removing one reward does not guarantee improved performance on the remaining reward.

AuthorsOpenReview ↗

Response to Reviewer 8c3P (Part 3)

Q8: In eq. 9, the authors do not say what M is...how is the sum to one property enforced?

A8: MM denotes the number of thresholds used to model the aggregated aesthetic score. The jj-th output pj(I^)p_j(\hat{I}) estimates the probability that the aesthetic score exceeds threshold jj. These probabilities are not mutually exclusive, and do not need to sum to one. We aggregate them to obtain the aesthetic score in Equation 9.

Conditioned on the product image and the instruction condition, the proposed UDPO strategy generates a group of candidates. For the k-th candidate, URM produces the functional advantage AkfuncA_k^{\mathrm{func}}, while DRM produces the aesthetic advantage AkaesA_k^{\mathrm{aes}}. AkaesA_k^{\mathrm{aes}} is converted into A~kaes\widetilde{A}_k^{\mathrm{aes}} based on the uncertainty modulation with uku_k in Equation 9 as follows.

ρk=1/uk1K∑i=1K1/ui,A~kaes=ρkAkaes−1K∑i=1KρiAiaes\rho_k=\frac{1/u_k}{\frac{1}{K}\sum_{i=1}^{K}1/u_i}, \qquad \widetilde A_k^{\mathrm{aes}}=\rho_k A_k^{\mathrm{aes}}-\frac{1}{K}\sum_{i=1}^{K}\rho_i A_i^{\mathrm{aes}}

Candidates with higher uncertainty receive smaller pre-centering weights. The fused advantage for the k-th sample is then computed using Equation 15.

We have added the description in the revised PDF to clarify how the uncertainty transforms the standardized aesthetic advantage (see L260-L268).

Q9: Both the fine tuned model and the VLM as a judge are from Qwen family of models.

A9: In addition to the Qwen-family models, we also evaluated representative methods using GPT-5.5 and Gemini-3-Pro in Appendix C (see Table 6), showing the consistent superior performance of our method against existing works. The human study and online CTR evaluation additionally assess participant preferences and observed clicks without using a VLM judge.

Q10: Since the model is optimized for exactly the same metrics that are reported along with a few other metrics, is it really surprising that the model does well on those?

A10: First, the training rewards are NOT the same as the evaluation metrics we used. As illustrated in L341-L349, URM and DRM are the training rewards and are reported as internal optimization measures. All other metrics are from a variety of evaluation sources, including scores from different VLM judges (Qwen-3-VL-235B-A22B-Thinking, GPT-5.5, Gemini-3-Pro), product consistency (Con), text accuracy (Sen. Acc), online Click-Through Rate (CTR), and user studies. These outcomes are not directly optimized during training.

Second, the optimization objective of our model is to bridge the gap between designer intent and user engagement. Improving metrics from both aspects simultaneously is practical but not easy to achieve. Our proposed training pipeline consists of a coarse-to-fine learning strategy with complementary reward models. The extensive evaluations show that our model consistently outperforms existing methods in user engagement and design aesthetics.

Q11: Which of the results in the main paper are from A/B tests? Are the CTRs reported from A/B tests or based on a model estimation?

A11: All reported CTR results are from real online A/B tests on a large-scale E-commerce platform, rather than model estimation.

Reviewer_Wy7cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to align product advertisements with both design quality and user click preferences. AdsPrompter first fine-tunes a prompt generator and an image renderer for advertising, then trains a User Reward Model on same-product CTR preference pairs and a Designer Reward Model on four design dimensions. With the renderer fixed, User-Designer Preference Optimization (UDPO) optimizes the prompt policy using both rewards and MGDA-based adaptive weighting. The authors construct a dataset of product–advertisement pairs and report offline evaluations, human comparisons, and online CTR results.

Critical strengths

The approach is well-motivated and addresses a practical problem, since visual appeal does not necessarily translate into higher click-through rates. Optimizing the prompt policy while keeping the renderer fixed during preference optimization is a clean and reasonable design.

The experiments are extensive and cover external VLM judges, human studies, and online CTR tests alongside the main offline benchmark.

Critical weaknesses

-The current ablation in Table 2 does not isolate the benefit of MGDA-based adaptive weighting. Removing either reward also removes the need to balance the two objectives. The full model performs better than either single-reward variant, but this does not establish whether adaptive weighting improves over a fixed-weight combination. Supporting that claim requires a two-reward fixed-weight control under the same training setup; otherwise, the conclusion should be limited to the benefit of using both rewards.

-The human-preference win rates reported in Section 5.3 do not match those shown in Figure 7. Section 5.3 cites Figure 7 and reports win rates of 61.2% and 63.5% for the full model over the baseline and coarse-level generation (w/o RL), whereas Figure 7 displays 83.2% and 66.9% for Baseline vs Ours (Full) and w/o RL vs Ours (Full). The authors should explain this discrepancy and clarify whether the two sets of rates use different participant groups, subsets, or aggregation procedures. Raw win/loss counts and a description of how judgments are aggregated across raters would also help establish the unit of analysis.

Questions for authors

-Through which function does the uncertainty in Equation (9) enter the aesthetic advantage in Equation (14)? What are the MGDA update interval, the EMA coefficient, and the initial weight?

-How do the three traffic buckets in Appendix B.6 map to the five variants in Table 2 and the nine methods in Table 5? The full model's online CTR differs between the two tables. Do the results in Tables 2 and 5 come from different experimental rounds?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Wy7c (Part 1)

We thank the reviewer for recognizing the motivation, practical problem formulation, and extensive evaluation of our work. We appreciate the constructive feedback and provide point-by-point responses below. We are happy to clarify any follow-up questions.

Q1: The current ablation in Table 2 does not isolate the benefit of MGDA-based adaptive weighting...Supporting that claim requires a two-reward fixed-weight control under the same training setup.

A1: Thanks for the valuable suggestion. We conduct a controlled comparison against fixed-weight fusion of the two rewards. Three different fixed-weight combinations, i.e., designer/user weights of 0.3/0.7, 0.5/0.5, and 0.7/0.3, are added under the same training setup.

As shown below, the proposed adaptive weighting strategy generally outperforms the fixed-weight configurations, supporting its benefit in bridging the gap between designer intent and user engagement. We have added the results and corresponding discussion in the revised PDF (see Table 3 and L480-L499).

MethodVLS ↑TLS ↑APS ↑STS ↑DRM ↑Con ↑Sen. Acc ↑URM ↑
Designer/User weight (0.3/0.7)8.4477.1987.8748.3870.7120.9260.8670.736
Designer/User weight (0.5/0.5)8.5187.2267.9808.4430.7370.9280.8660.731
Designer/User weight (0.7/0.3)8.5327.2917.9628.4520.7760.9250.8640.694
Adaptive Weighting (Ours)8.5617.3567.9718.4690.8210.9290.8700.742

Q2: The human-preference win rates reported in Section 5.3 do not match those shown in Figure 7...Raw win/loss counts and a description of how judgments are aggregated across raters would also help establish the unit of analysis.

A2: Sorry for the confusion. Our user study involves two subgroups, i.e., 24 novice users and 6 professional designers. The two sets of rates correspond to different participant groups within the same study. The original Figure 7 reports preferences from 6 professional designers, whereas Section 5.3 reports results across all 30 participants. The group-specific and complete results of the user study are shown below. Each percentage denotes preference for the first method over the second.

GroupOurs vs. BaselineOurs vs. w/o RLw/o RL vs. Baseline
Professional Designers (6)83.2%66.9%76.4%
Novice Users (24)55.7%62.7%59.3%
All Participants (30)61.2%63.5%62.7%

For each comparison (Ours vs. Baseline, Ours vs. w/o RL, or w/o RL vs. Baseline), participants viewed the same input and two generated advertisements side by side, with randomized left/right placement. They selected the preferred advertisement based on textual fidelity, visual quality, layout design, and commercial appeal. Each of the three method comparisons contains 200 image pairs, with every pair evaluated by all participants. The win rate is calculated as the number of valid judgments favoring the first method divided by the total number of valid judgments for that comparison.

We have revised Figure 7 to show the results of all participants, and added the detailed user study process and the group-specific results in the Appendix (see Appendix B.7).

Q3: Through which function does the uncertainty in Equation (9) enter the aesthetic advantage in Equation (14)?

A3: We use inverse-uncertainty weighting followed by group-wise re-centering. Conditioned on the product image and the instruction condition, the proposed UDPO strategy generates a group of candidates. For the k-th candidate, URM produces the functional advantage AkfuncA_k^{\mathrm{func}} , while DRM produces the aesthetic advantage AkaesA_k^{\mathrm{aes}} . AkaesA_k^{\mathrm{aes}} is converted into A~kaes\widetilde{A}_k^{\mathrm{aes}} based on the uncertainty modulation with uku_k in Equation 9 as follows.

ρk=1/uk1K∑i=1K1/ui,A~kaes=ρkAkaes−1K∑i=1KρiAiaes\rho_k=\frac{1/u_k}{\frac{1}{K}\sum_{i=1}^{K}1/u_i}, \qquad \widetilde A_k^{\mathrm{aes}}=\rho_k A_k^{\mathrm{aes}}-\frac{1}{K}\sum_{i=1}^{K}\rho_i A_i^{\mathrm{aes}}

Candidates with higher uncertainty receive smaller pre-centering weights. The fused advantage for the k-th sample is then computed using Equation 14 in the original version (Equation 15 in the revision).

We have added the description in the revised PDF to clarify how the uncertainty transforms the standardized aesthetic advantage (see L260-L268).

AuthorsOpenReview ↗

Response to Reviewer Wy7c (Part 2)

Q4: What are the MGDA update interval, the EMA coefficient, and the initial weight?

A4: The MGDA update interval is set to 50 steps. The EMA coefficient is set to 0.9, and the initial fusion weight is set to 0.5. We have added the details in the revised PDF (see L481-L483).

Q5: How do the three traffic buckets in Appendix B.6 map to the five variants in Table 2 and the nine methods in Table 5? The full model's online CTR differs between the two tables. Do the results in Tables 2 and 5 come from different experimental rounds?

A5: The online A/B tests were conducted on a large-scale E-commerce platform to validate the CTR performance of the generated ads. The three traffic buckets do not correspond one-to-one to methods. They are disjoint traffic partitions on the platform to allocate experimental traffic and expand exposure coverage. Each bucket contains all methods evaluated in the corresponding round.

Note that the platform supports at most ten creatives per product in a single exposure experiment. Table 2 and the original Table 5 (Table 6 in the revision) report two separate online A/B testing rounds with different objectives. The goal of Table 2 is to evaluate the effectiveness of individual components of our work, while Table 5 compares our work with external methods. The full-model CTRs were measured in different deployment windows and are therefore reported separately. Our comparisons are made within each experiment. Our conclusions are based on differences between methods evaluated during the same round and across the same traffic partitions, rather than on comparisons of absolute CTR across rounds.

We have clarified the details of the online A/B tests in the Appendix (see Appendix B.6 and C).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Revisions

We thank the AC and all reviewers for their time and effort throughout the review process. We sincerely appreciate the valuable feedback and recognition of our work for the practical importance of jointly considering design quality and click preferences for multimodal advertisements. We have carefully considered each comment and provided point-by-point responses below.

As suggested, we have also revised the main text to address the key concerns, and incorporated details in the appendix. The revisions in the main text are summarized as follows:

  1. L60-L63: We clarify how designer intent (DI) and user engagement (UE) are measured in Figure 1, as suggested by Reviewer 8c3P.
  2. L260-L268: We add an equation and its accompanying descriptions to specify how the uncertainty in Equation 9 is used in Equation 15, as suggested by Reviewers Wy7c, 6xFH, and 8c3P.
  3. L442-L448: We show the win rates across all participants in Figure 7, and provide group-specific results and the evaluation protocol in Appendix B.7, as suggested by Reviewers Wy7c and 6xFH.
  4. L480-L499: We add Table 3 and a reward weight analysis subsection to compare adaptive fusion with three fixed-weight configurations and uncertainty modulation, as suggested by Reviewers Wy7c and 6xFH. We also clarify the MGDA update settings and sensitivity during policy optimization, as suggested by Reviewer Sjsh.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.