acceptodds
Under review as a conference paper at ICLR 2027

CommerceVibe: Learning to Design E-Commerce Creatives as Executable Visual Code via Dual-Feedback Reinforcement Learning

Abstract

High-quality e-commerce creatives are essential for presenting products and conveying marketing messages. Recent diffusion models enable scalable creative generation and produce visually compelling images, but their flattened raster outputs often contain distorted text and inconsistent product details, requiring refinement before deployment. Moreover, without explicit structure, the resulting creatives are difficult to edit and reuse, while complex design requirements remain challenging to encode as verifiable training signals. To address these challenges, we present CommerceVibe, which represents creatives as executable visual code and formulates generation as conditional HTML/CSS program synthesis. Given product images, design requirements, and product information, it produces renderable, editable, and reusable creatives. We further introduce dual-feedback reinforcement learning, in which rule-based feedback evaluates rendered programs for text readability, product visibility, and layout validity, while visual feedback from a vision-language model (VLM) assesses rendered creatives against input specifications across six perceptual and commercial dimensions. Together, these complementary feedback signals improve both constraint satisfaction and perceptual quality. We perform supervised fine-tuning (SFT) of Qwen3.5-9B on over 28,000 e-commerce examples, followed by dual-feedback reinforcement learning. On a 1,300-case benchmark, the optimized CommerceVibe model achieves a weighted score of 94.0/100, compared with 87.3 for the SFT-only variant, and outperforms the three evaluated external models under the same one-pass HTML/CSS generation protocol. A blind evaluation by five e-commerce design experts provides independent evidence of improved output quality. CommerceVibe supports controllable and editable e-commerce creative generation through executable HTML/CSS.

Primary area:
generative models
Keywords:
multimodal generation, structured visual generation, HTML/CSS program synthesis, reinforcement learning with verifiable rewards

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nysUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies e-commerce creative generation: given product images, design requirements, and product information, the model generates renderable and editable HTML/CSS. The method first performs supervised fine-tuning of Qwen3.5-9B, followed by GRPO optimization using two types of feedback: checks of text, products, and layout based on the rendered structure, and VLM assessments of visual and commercial quality. The paper constructs a benchmark containing 1,300 cases and reports that the combined-feedback model improves the automatic score from 87.3 for SFT to 94.0, and the score from a blind evaluation by five experts from 81.6 to 90.0.

Critical strengths

1.The problem has clear practical value, and the representation is well suited to the task. Directly referencing product images and using native HTML text helps avoid errors caused by regenerating product appearance and text, while supporting element-level editing and structural checks.

2.The experiments support the effectiveness of combined feedback. The three RL variants share the same initialization and training configuration, with the feedback source being the primary difference. Table 1 shows that combined feedback outperforms both single-feedback variants, with larger gains on the Multi subset, providing evidence that structural constraints and visual preferences are complementary.

Critical weaknesses

1.The core technical novelty is limited, with the contribution primarily consisting of engineering integration and task-specific adaptation of existing methods. DesignAsCode already formulates design as HTML/CSS synthesis; AeSlides already uses rendered geometric constraints and GRPO; AesthetiQ already performs aesthetic preference optimization; and Learning to Present already combines structural, rendering, and screenshot-based aesthetic rewards. In addition, the supplementary experiments of PosterReward already combine text correctness and visual preference rewards. This paper adopts standard GRPO and an equally weighted reward combination, with the additions primarily involving e-commerce constraints and Single/Multi rule adaptation. The experiments demonstrate that this combination is effective, but do not establish technical contributions or research insights beyond combining existing feedback signals and adapting them to the task. I therefore consider the novelty insufficient for an ICLR-level research contribution. 2.The choice of comparison methods is insufficient to establish advantages over existing creative generation approaches. The authors propose an RL-based fine-tuning method, yet Table 1 compares general-purpose large models and variants of their own training approach without including the most relevant alternative methods for commercial design. Consequently, the comparison does not establish whether the proposed method outperforms existing mainstream optimization approaches, and it also weakens the novelty argument. The authors should include adaptable methods mentioned in Weakness 1 as baselines, or specifically explain why they are not comparable. The current comparison is insufficient to establish the method’s advantages over existing creative generation approaches. 3.The coverage and reliability of the custom evaluation framework are insufficiently validated. Automatic evaluation reuses the rule dimensions and weights employed during training. Although the model used for visual evaluation is different, the scoring dimensions remain the same; expert scoring also follows the corresponding evaluation framework. Public resources such as Graphic-Design-Bench, PosterReward/PosterBench, and ProductWebGen already provide evaluations of design quality or instruction following, but the paper does not use these resources for cross-validation. Developing task-specific e-commerce metrics is reasonable; however, the paper lacks validation of detection accuracy and the robustness of method rankings to thresholds and weights. The evaluation may therefore favor models specifically optimized for this reward framework. It does not provide a sufficient, independent, and fair comparison of overall creative quality, nor does it comprehensively and concretely establish whether the method delivers strong results in practice.

Questions for authors

1.Regarding novelty: Prior works including DesignAsCode, AeSlides, AesthetiQ, Learning to Present and PosterReward have explored HTML generation, GRPO, aesthetic rewards and multi-signal reward fusion. Beyond e-commerce adaptation and Single/Multi rules, what unique technical designs or new insights does this paper introduce? 2.Regarding baselines: The experiments only compare general LLMs and internal training variants, without the closely related creative-generation baselines listed above. Could you add comparisons against these works, or clearly explain why they are not suitable for a fair comparison?

2 comments
AuthorsOpenReview ↗

Methodological Novelty, Comparative Evaluation, and Evaluation Reliability

Thank you for recognizing the practical value of our work and the effectiveness of dual-feedback optimization. We appreciate your comments on methodological novelty, comparisons, and evaluation reliability.

  1. Methodological novelty (W1 & Q1). While HTML/CSS generation, GRPO, and multi-reward optimization are established techniques, our research investigates efficient, high-quality, and cost-effective e-commerce creative generation at scale. We formulate a design task involving both deterministic structural constraints and subjective visual preferences as a generation task optimized through reinforcement learning. Our technical contributions include: (1) Executable HTML/CSS that directly references supplied product assets, avoiding appearance changes caused by image regeneration and enabling programmatic detection of incorrect product references, copy violations, occlusion, and layout issues. (2) Input-conditioned rule verification, with checks, thresholds, weights, and critical violations tailored to the number of product images and e-commerce requirements. (3) Hard gates and critical-error penalties combined with holistic visual preference feedback for policy optimization. Equal weighting is merely the final reward aggregation step. More importantly, our experiments reveal stronger complementarity between structural verification and visual preference feedback in more complex design scenarios. Our contribution therefore extends beyond applying existing techniques to e-commerce, encompassing task formulation, feedback design, and empirical insights into optimizing complex visual generation tasks.
  2. Comparison with related methods (W2 & Q2). DesignAsCode employs iterative visual reflection and code refinement; AeSlides and Learning to Present focus on presentation generation; AesthetiQ and PosterReward emphasize aesthetic preference or poster quality optimization. These methods differ from ours in task constraints, representations, objectives, or inference procedures. Our external comparisons follow a unified one-pass HTML/CSS generation protocol, while controlled ablations compare feedback mechanisms under identical initialization and training configurations. Comparisons with these specialized methods require appropriate task adaptation and would provide complementary evidence.
  3. Evaluation reliability (W3). Our evaluation criteria reflect predefined e-commerce design requirements. Five design experts with different aesthetic preferences conduct randomized blind assessments without access to model identities or automatic scores. Their agreement reaches ICC(A,5) = 0.811, and combined feedback consistently improves both automatic and expert scores. Further evaluation beyond the current task-specific framework would provide additional evidence. During rule development, individual checks were manually reviewed on a large internal dataset. Their weights reflect the relative importance of different requirements in practical e-commerce production. Appendices D–F detail the rules, thresholds, weights, and reward computation. We plan to investigate ranking sensitivity to different thresholds and weights as resources permit. Graphic-Design-Bench, PosterBench, and ProductWebGen offer complementary evaluation perspectives but do not fully address product asset preservation, required copy completeness, and structured layout constraints. Our task-specific benchmark directly assesses these requirements. Cross-benchmark generalization remains an important research direction. 4.Manuscript revisions. We will clarify our contributions in task formulation, conditional verification, and reward construction; elaborate on feedback complementarity across input scenarios; and explain the rationale for rule weighting, manual verification, and evaluation scope.

Thank you for the detailed response. I have read the rebuttal and the other reviews.

  1. Novelty. The rebuttal restates the components, but does not explain what is technically different from prior work that already combines structural/rendering rewards with aesthetic rewards under GRPO (e.g., AeSlides, Learning to Present, PosterReward). The finding that the two feedback types complement each other more on the Multi subset is useful, but it is already in Table 1 and is an empirical observation, not a new method or insight that carries over to other tasks.

  2. Baselines. I understand that the specialized methods use different pipelines. Still, at least one of them could be compared under a controlled setting. Without such a comparison, the paper cannot show that its feedback design beats existing alternatives, and not only general-purpose LLMs and its own variants.

  3. Evaluation reliability. The high ICC shows that the experts agree with each other, but it does not address my concern: the training reward, the automatic metric, and the expert rubric all share the same dimensions.

Since no analyses were provided for these points, I will keep my original rating.

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

The paper proposes CommerceVibe, a method for converting e-commerce creative posters or product display pages into HTML files. The model generates a complete, executable HTML/CSS file in a single pass, directly referencing input product images and rendering text using native HTML. The method improves structural compliance and overall visual quality through an “SFT + dual-feedback reinforcement learning” training paradigm, combining rule-based feedback and VLM preference feedback via GRPO. Evaluated on a 1,300-example benchmark, CommerceVibe achieves an automatic total score of 94.0 and an expert blind-evaluation score of 90.0, outperforming closed-source models such as GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash.

Critical strengths

The proposed method is reasonable and effective. The “SFT + GRPO” training paradigm effectively improves the model’s code generation ability as well as the visual layout quality of rendered outputs. Compared with the baseline, the model shows substantial performance gains.

Critical weaknesses
  1. The evaluation scope is relatively narrow. The authors primarily assess visual quality based on HTML-rendered outputs. However, given the task formulation and evaluation criteria, image generation models (e.g., GPT-image-2 and similar systems) could also satisfy these 2. requirements. The paper lacks comparative experiments with image generation baselines, and it is possible that such models could outperform both the proposed method and other understanding-based models.
  2. Although many benchmarks exist for evaluating poster visual quality, the paper does not report experiments on alternative benchmarks, limiting the generalizability of the evaluation.
Questions for authors
  1. Did you consider alternative objectives during the SFT stage? What motivated the choice of token-level negative log-likelihood?
  2. How is the reference HTML/CSS training target constructed?
  3. How is reward normalization implemented? Is it performed using a model, a programmatic rule, or manual annotation?
1 comment
AuthorsOpenReview ↗

Baseline Selection, Benchmark Design, and Training Details

Thank you for your constructive feedback. We would like to further clarify our task formulation, practical production requirements, evaluation scope, and training details.

  1. Comparison with image generation models. We agree that image generation models such as GPT-image-2 can produce visually appealing e-commerce advertisements. Our decision not to include them in the current comparison protocol is based on three main considerations: (1) Evaluation requirements. Beyond visual quality, e-commerce advertisements must faithfully represent the actual products and accurately and completely present the required copy. Image generation models typically produce raster images without directly accessible product image references, native text, or element structures, making it difficult to apply programmatic verification directly. (2) Output dimensions. Our benchmark includes long-form product detail images exceeding 7,500 pixels in height, which is beyond the native output dimensions of many mainstream image generation models. In contrast, HTML/CSS supports the generation of such long-form creatives through structured layouts. (3) Production costs. Generation costs are an important consideration in large-scale e-commerce production, where a complete set of advertising creatives may contain dozens of images. The estimated production costs of different approaches are presented below (in CNY). Under this estimation, our approach reduces generation costs by approximately 65.6%–94.9%. Generation approach Cost for 10 images GPT-image-2 (Low) ¥0.66 Qwen-image-pro-3 ¥1.35 GPT-image-2 (Medium) ¥4.50 CommerceVibe ¥0.23 Accordingly, our baseline selection considers not only visual quality but also the verifiability of task constraints, output dimensions, and practical production costs. Our current experiments primarily assess model performance under a unified HTML/CSS generation protocol rather than comparing generated images solely in terms of visual quality.
  2. Evaluation on other public benchmarks. E-commerce creative generation differs from general poster or product webpage generation in both task constraints and evaluation priorities. Existing public benchmarks do not jointly cover requirements such as the preservation of supplied product assets, completeness of required copy, product visibility, text readability, and layout validity. We therefore construct a task-specific benchmark that is disjoint from the training data at the product level. We recognize the complementary value of other public benchmarks and plan to investigate cross-benchmark generalization.
  3. Choice of the SFT objective. We considered alternative objectives, including sequence-level loss and weighted cross-entropy, and ultimately adopted token-level negative log-likelihood (NLL) because it directly corresponds to the supervised learning objective for autoregressive HTML/CSS sequence generation. In our framework, SFT primarily enables the model to learn code structures and design representations, while product presentation, text readability, layout validity, and overall visual quality are subsequently optimized through rendering-based feedback during GRPO.
  4. Construction of reference HTML/CSS documents. The product images, product information, and design requirements used as SFT inputs are collected from real-world e-commerce design scenarios. The corresponding reference HTML/CSS documents are generated by large language models, manually screened for quality, and iteratively refined before being used as supervision targets. These documents contain information such as product image references, fonts, and layouts. The resulting dataset comprises 28,568 training examples and 3,174 validation examples. Further details on data construction and request normalization are provided in Appendix B (SFT Data Construction).
  5. Reward normalization. The rule reward is computed programmatically using predefined checks, weights, and penalty mechanisms. The preference reward is obtained from a fixed Qwen3-VL-Plus Judge that evaluates six dimensions, followed by normalization through a predefined mapping. The two rewards are combined with equal weights. During GRPO, the relative advantages of candidates sampled for the same input are computed using the mean and standard deviation of their rewards within the group. Both reward mapping and within-group standardization are performed programmatically, without requiring an additional normalization model or manual annotations. Further details are provided in Section 3.4 (Eq. 3) and Appendices D–F.
  6. Manuscript revisions. We will elaborate on the cost motivation for structured generation in multi-module e-commerce creative production in the Introduction, clarify the rationale for adopting NLL in Section 3.1, and provide additional details on the generation and quality screening of reference HTML/CSS documents in Appendix B.
Reviewer_vK8xOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CommerceVibe, which generates e-commerce creatives as executable visual code by formulating the task as conditional HTML/CSS program synthesis given product images, design requirements, and product information. It further introduces dual-feedback reinforcement learning that uses feedback across different dimensions. Experiments on a 1,300-case benchmark show improvements over external models.

Critical strengths
  1. The motivation and method design are clear and solid. The paper starts from specific practical requirements and current challenges, then develops a method to address them.
  2. The experiments are solid and support the method design. Benchmark comparisons demonstrate improvements, which are further validated by five human design experts.
Critical weaknesses
  1. Methodological novelty is limited. The SFT–GRPO workflow and HTML/CSS representation are common choices. The paper could clarify its novelty beyond the specific application and reward design. The external comparisons focus on general-purpose models such as GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, providing less insight into its advantages over closely related methods.
Questions for authors

What is the main methodological novelty compared with related work discussed in the paper, such as PosterVerse and DesignAsCode?

1 comment
AuthorsOpenReview ↗

Methodological Contributions and Related Work

Thank you for your constructive feedback. We would like to further clarify the characteristics of our task, our methodological contributions, and the differences between our approach and related work.

  1. Methodological contributions. Compared with general poster or graphic design generation, e-commerce creative generation must satisfy more explicit constraints on product assets and content. These include faithfully presenting the supplied products, accurately and completely displaying the required copy, complying with platform content policies, and avoiding product occlusion and text overflow. These requirements motivate a different approach to creative generation, together with a corresponding framework for verifiable quality assessment. To this end, CommerceVibe formulates the task as HTML/CSS design generation conditioned on supplied product assets. We develop fine-grained rule checks tailored to the number of input product images to assess text, product, and layout constraints, thereby providing verifiable rewards for policy optimization. Meanwhile, we introduce VLM-based feedback to assess overall visual and commercial quality. Both feedback sources are jointly used for GRPO policy optimization. Our ablation studies demonstrate their complementarity, with combined feedback outperforming either feedback source alone. The detailed design is presented in Appendix D (Partially Conditional Rule Verifier).
  2. Differences from PosterVerse and DesignAsCode. PosterVerse introduces a commercial poster generation pipeline that combines background generation with HTML-based typography. DesignAsCode generates editable HTML/CSS designs and refines them through iterative rendering and visual reflection. In contrast, CommerceVibe focuses on preserving supplied product assets and satisfying e-commerce platform-specific constraints, which motivates different verification criteria and rule configurations. Furthermore, we incorporate rule-based and preference feedback into policy optimization during training, rather than repeatedly refining individual outputs at inference time, enabling one-pass HTML/CSS creative generation.
  3. Scope of comparison. We recognize the value of comparisons with specialized design generation systems. Our current experiments follow a unified one-pass HTML/CSS generation protocol, whereas PosterVerse and DesignAsCode employ different generation or iterative refinement pipelines. Direct comparisons between complete systems would therefore introduce additional differences in their generation procedures. We will clarify the scope of our comparisons in the revised manuscript.
  4. Manuscript revisions. We will clarify the rule verification mechanism conditioned on the number of input images and the dual-feedback GRPO optimization framework in the Introduction. We will also elaborate on the product asset constraints and verification criteria of e-commerce creative generation in Related Work, highlighting how CommerceVibe differs from PosterVerse and DesignAsCode in its use of feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author Responses and Key Research Contributions

Dear Senior Area Chair, We sincerely thank the reviewers for their constructive feedback and for recognizing the practical value of our work, the suitability of executable HTML/CSS for e-commerce creative generation, and the effectiveness of dual-feedback optimization. We have provided detailed responses to each reviewer and would like to summarize the key contributions of our work and address the common concerns raised during the review process.

  1. Research contributions and methodological novelty CommerceVibe investigates how to enable efficient, high-quality, and cost-effective e-commerce creative generation at scale. Unlike general poster or graphic design generation, this task requires not only visual appeal but also faithful product presentation, accurate and complete copy, compliance with platform requirements, and valid layouts. Our contribution lies in formulating these requirements as verifiable training signals through executable HTML/CSS, input-conditioned rule verification, hard gates, and critical-error penalties, together with holistic visual preference feedback. Although we adopt standard GRPO, the construction of effective feedback for this structured visual generation task constitutes a central methodological contribution. Moreover, our controlled experiments reveal that individual feedback sources become less effective in more complex design scenarios, whereas their combination provides substantially greater improvements. This finding offers empirical insights into reinforcement learning for visual design tasks involving both explicit structural constraints and subjective visual preferences.
  2. Baseline selection and task-specific evaluation The reviewers raised common concerns regarding comparisons with specialized design generation systems and image generation models. As detailed in our responses, these approaches differ from CommerceVibe in task constraints, output representations, optimization objectives, or inference procedures. Our unified one-pass HTML/CSS generation protocol enables controlled comparisons while addressing practical requirements such as product asset preservation, editability, verifiability, long-form creative generation, and production costs. Comparisons with adapted specialized methods would provide valuable complementary evidence. Similarly, existing public benchmarks for poster, graphic design, and product webpage generation do not jointly cover the specific requirements of e-commerce creatives. This motivates our task-specific benchmark of 1,300 cases, with product-level separation from the training data.
  3. Evaluation reliability and empirical evidence Our evaluation framework is grounded in practical e-commerce design requirements. Individual rule checks were manually reviewed on a large internal dataset during development, and their weights reflect the relative importance of different design requirements in actual production. Furthermore, our automatic evaluation and blind assessments by five e-commerce design experts consistently support the effectiveness of CommerceVibe. The combined-feedback approach improves the automatic score from 87.3 to 94.0 and the expert score from 81.6 to 90.0, with an inter-rater agreement of ICC(A,5) = 0.811. These results provide complementary evidence of improved constraint satisfaction and visual quality. We also recognize the value of further validation through alternative benchmarks, evaluation criteria, and sensitivity analyses.
  4. Manuscript improvements In response to the reviewers' feedback, we will further clarify our research contributions, distinctions from related work, empirical insights, data construction process, and the rationale behind our verification and evaluation framework. Overall, we believe CommerceVibe contributes not only a practical solution for large-scale e-commerce creative generation but also a meaningful investigation into how structured verification and visual preference feedback can jointly optimize multimodal generation. We hope that our responses help clarify the scientific contributions and practical significance of our work. Thank you for your time and consideration. The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.