acceptodds
Withdrawn from ICLR 2027

Sense it with your eyes: Sensation Generation and Understanding for Advertisements

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sensory advertising evokes human senses through visual cues, enabling audiences to mentally simulate experiences and increasing persuasive impact. Despite the recent increase in using AI in generating and understanding creative and persuasive content, how advertisements visually evoke sensations remains largely unexplored. In this work, we introduce the first study of evaluating and generating sensory ads. We introduce the Sensory Ad dataset, and propose SenseScore, an automated evaluation metric for sensation evocation achieving strong agreement with human judgments. Further, we introduce the Sensory Ad Generation (SenseGen) task and propose SAGA, a multi-agent framework that improves message–image alignment, sensory evocation, and persuasion. Our work establishes a foundation for sensory-aware visual persuasion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-to-Image Generation, Persuasive Images, Image Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_f5kJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies sensation understanding and generation in advertisements. It introduces hierarchical sensation annotations and trains SenseScore using CPO with a hierarchy constraint to improve agreement with human judgments. It also proposes SAGA, a Planner–Prompt Refiner–Evaluator framework for iteratively generating and editing advertisements to convey intended messages and sensations.

Critical strengths
  • The paper offers a valuable perspective on advertisement generation by examining sensory evocation alongside message alignment.
  • The dataset provides detailed hierarchical sensation annotations for 1,100 advertisements, approximately 700 of which are used for training.
  • Table 2 provides useful evidence for the scoring method: overall Cohen’s κ increases from 0.33 with SFT to 0.57 with CPO and 0.69 with CPO plus the hierarchy constraint.
Critical weaknesses
  • The contribution of individual SAGA components is unclear. Table 3 compares the complete system against generation baselines, leaving the benefits of planning, prompt refinement, editing, and evaluator feedback unresolved. The authors should clarify whether SenseScore itself is used in the generation loop and provide ablations without evaluator feedback and without its sensation-specific feedback.
  • Stronger prompting baselines are needed. The shared action–reason and sensation template is a reasonable task specification, but a baseline that expands it into a detailed visual prompt would help establish the additional value of SAGA. Comparisons using the same generator and editor under comparable computational budgets would further strengthen the evidence.
  • Iteration behavior is insufficiently documented. Please report the maximum and typical numbers of iterations, stopping criteria, and computational cost, together with changes in SenseScore, message alignment, and persuasion across iterations.
Questions for authors
  • Could the evaluation include aesthetic quality and text legibility alongside its existing consistency and message-alignment criteria?
  • Text is already acknowledged in the paper, but its contribution to sensory evocation remains unclear. Separating textual and nontextual contributions would strengthen the analysis.
  • Why are Qwen-Image results included in Table 3 but omitted from Figure 6?
  • Could SenseScore support reward-based optimization of image generators? This is an interesting future direction, provided message alignment and visual quality are also considered.
2 comments

We thank Reviewer f5kJ for their constructive feedback. We are glad they found our study to provide valuable perspective, our dataset to provide detailed annotations and our evidence on evaluation metric useful. We address each concern and question inline.

[W1] SAGA component ablations. The contribution of individual SAGA components is unclear ... clarify whether SenseScore is used in the generation loop and provide ablations without evaluator feedback and sensation-specific feedback.

We first clarify that SenseScore is not used in SAGA's generation loop; it is an independent metric for evaluating sensation evocation. Instead, SAGA uses an MLLM-based evaluator to provide feedback on visual consistency, message alignment, and sensation evocation (Sec. 3.3, L293–301; Appendix F).

To further evaluate SAGA's design choices, we conducted additional ablations on 270 images, using the same FLUX generator, FluxKontext editor, and refinement stopping criteria (Sec. 3.3, L293–301; Fig. 4). Removing evaluator priority increases SenseScore (0.85 → 0.93) but decreases AIM (0.69 → 0.67), showing that stronger sensation evocation can come at the expense of message alignment. Removing the text refiner also decreases both metrics. Here is the table including the the ablations and the strongest baselines.

MethodSenseScore ↑AIM ↑
SAGA (Full)0.850.69
SAGA w/o Priority0.930.67
SAGA w/o Text Refiner0.830.67
SAGA w/o Critic0.830.65
DALLE3 (Baseline)0.790.67

We further clarify that our primary contribution is introducing the sensory advertisement generation task, dataset, and evaluation metric, rather than proposing a fundamentally novel agentic method. SAGA provides a task-specific iterative refinement baseline, and these ablations support the benefit of its design choices in balancing sensation evocation and message communication. We will include the results in the revised paper (Table X, highlighted in blue).

[W2] Stronger prompting and controlled baselines. Stronger prompting baselines...

Thanks for the suggestion. We clarify that our evaluation already includes DALLE3, which automatically rewrites and expands prompts before generation, providing a strong prompting baseline beyond direct generation with the original template. SAGA achieves higher SenseScore (0.83 vs. 0.79) and significantly higher AIM (0.66 vs. 0.63, p < 1e-5) than DALLE3 (Table 3). Further, our new additional ablation (W1) use the same FLUX generator, FluxKontext editor, and refinement stopping criteria, showing that removing evaluator priority in critic agent, the text refiner, or critic agent reduces AIM (0.69 → 0.67). These results support the benefit of SAGA's task-specific refinement beyond stronger initial prompting and generic agentic pieline. We further clarify we see SAGA as a strong baseline for the newly introduced SenseGen task, rather than an optimal or state-of-the-art agentic generation method. We hope that this work inspire future works to develop stronger and more efficient methods.

[W3] Iteration behavior and computational cost. Report maximum and typical numbers of iterations, stopping criteria, computational cost, and changes in evaluation metrics across iterations.

We clarify that the stopping criteria are already described in Sec. 3.3 (L293–301) and Fig. 4: SAGA terminates when the Evaluator returns "No Issue" or the maximum conversation budget is reached. We will additionally report the average number of refinement iterations 3, maximum number of messages 20 and the average number of messages 11 per image (revised the paper in C4 (L937-941) and L302), and provide a qualitative example illustrating how evaluator feedback guides successive image refinements.

[Q1-Q2]Aesthetic quality, text legibility, and textual contributions.

Thanks for the suggestions. We clarify that our work is the first to study sensory advertisement generation, focusing on two objectives: (1) conveying the intended advertisement message and (2) evoking a target sensation. While aesthetics, text legibility, and disentangling textual from nontextual contributions are interesting directions, they are beyond the scope of this work. We hope our task, dataset, and evaluation metric will encourage future research on these aspects, including the contributions of individual advertisement components to sensory evocation. We will clarify this scope and discuss these directions as future work.

[Q3] Qwen-Image in Fig. 6.

Thanks, we will include a full set including both Qwen-Image images and ablation results in appendix.

[Q4] SenseScore for reward-based optimization.

Thanks for the interesting suggestion. SenseScore could potentially serve as a reward signal for optimizing image generators toward stronger sensation evocation. However, as shown in Fig. 7, improving sensation intensity alone does not necessarily improve message alignment or persuasion. This suggests that reward-based generation should jointly optimize sensation evocation, message alignment, and visual quality, rather than maximizing SenseScore alone. We will discuss this potential direction in the revised paper.

Reviewer_Xv9jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the recognition, evaluation, and generation of visually evoked sensations in advertisements. It introduces the Sensory Ad dataset, comprising 1,100 real advertisements annotated using a hierarchy of 96 sensation labels, with evocation strengths and descriptions of relevant visual elements. SenseScore uses an MLLM-generated image description and an LLM trained with preference and hierarchy losses to score target sensations. The paper also introduces the SenseGen task and SAGA, a framework that iteratively plans, refines, and evaluates image edits. Experiments report improved agreement with human sensation judgments, higher automatic scores for generated advertisements, and human preferences favoring SAGA over DALLE3 for message communication and sensation evocation.

Critical strengths
  • S1. A useful task and annotated resource. The paper gives sensory evocation a concrete formulation beyond literal image-text alignment. The hierarchical labels, intensity annotations, and visual-cue descriptions provide a useful basis for studying this problem (Section 3.1 and Appendix B). This task formulation and resource are meaningful contributions even independently of the proposed generation system.
  • S2. Promising empirical evidence. SenseScore improves the reported real-ad Cohen's κ from VQA-score's 0.53 to 0.69, and its ablation improves κ from 0.57 with CPO alone to 0.69 with the hierarchy loss (Tables 1 and 2). The human comparison also favors SAGA over DALLE3, with reported preferences of 76% versus 10% for message communication and 48% versus 24% for sensation evocation (Section 4.2, lines 429-431). These results support the potential value of the approach.
Critical weaknesses

W1. SenseScore's validation does not adequately establish the intensity comparisons used downstream

  • Evidence: Section 4.1, lines 308-317, evaluates preferences between different sensation labels within the same image. At most nine of 96 labels receive positive scores, and all unselected labels receive zero. Consequently, under the stated all-pairs protocol, selected-versus-unselected comparisons must dominate the pair count. This is a consequence of the sampling design; without results separated by pair type, it is unclear how much the reported κ establishes fine-grained intensity discrimination among positively annotated sensations. Moreover, Table 3 and Figure 7 use SenseScore to compare different images, whereas correctly ranking sensations within an image does not establish correct ranking of images for a fixed sensation. Figure 5 provides illustrative cross-image examples, but I could not find an aggregate metric-validation result for this setting. The generated-image validation is also small, with Section 4.1 reporting 50 images, and Table 1 reports no uncertainty estimates for κ.
  • Impact on recommendation: SenseScore is both a central contribution and a principal basis for evaluating the generated advertisements. Its reported agreement may largely reflect sensation recognition, leaving its reliability for comparing evocation strength across images insufficiently established. Validation on matched-sensation image pairs, together with results restricted to pairs of positively annotated sensations, would directly address this concern.

W2. The generation experiments do not isolate SAGA's methodological contribution

  • Evidence: SAGA combines FLUX generation with iterative planning, evaluation, and image editing; Figure 12 identifies FluxKontext as the editor. Tables 3 and 5 compare the resulting pipeline with direct-generation baselines, without a controlled comparison using the same generator, editor, and inference budget. I could not find a SAGA ablation testing the benefit of separating the agents or prioritizing message alignment before sensation; Table 2 instead ablates SenseScore. The favorable human comparison with DALLE3 supports the complete pipeline, but does not distinguish these design choices from the benefit of additional refinement. Furthermore, that human study assesses message communication and sensation evocation, while the persuasion claim relies on Pcomp, an automatic score defined in Section 3.4 of CAP (Aghazadeh and Kovashka, 2025). The reported Pcomp improvement over DALLE3 is 0.62 versus 0.61, without uncertainty estimates (Table 3).
  • Impact on recommendation: The results show promise for the complete system, but leave the benefit of the proposed agent design and the claimed improvement in persuasion insufficiently supported. A matched-budget generic refinement baseline and a focused ablation of the evaluation priority would substantially strengthen the methodological case. The persuasion conclusion should be limited to the automatic proxy unless supported by a corresponding human assessment.
Questions for authors
  1. Metric validation, addressing W1. How well does SenseScore agree with humans when comparing different images for the same target sensation, and when evaluation is restricted to pairs of positively annotated sensations? Please report the number of independent images and uncertainty computed at the image level. Also reconcile the 15 images from each of five generators described in Section 3.1 with the 50 generated images evaluated in Section 4.1. Strong results on these evaluations would materially increase my confidence in the metric.
  2. Controlled generation comparison, addressing W2. Does SAGA retain its advantage over a simpler refinement loop using the same FLUX generator, FluxKontext editor, and comparable model-call budget? What happens when the evaluator's message-before-sensation priority is removed? Please specify the agent backbones and maximum/average refinement counts so that the comparison is interpretable. These results would determine how much of the gain can be attributed to the proposed design.
1 comment

We thank Reviewer Xv9j for their detailed and constructive feedback. We are glad they found our task formulation and annotated dataset valuable, and our empirical results promising, particularly the improvements of SenseScore over existing metrics and the human evaluation favoring SAGA.

We address each concern and question inline.

[W1/Q1] SenseScore Validation. To further validate our metric for our image - image comparisons, we conduct a human evaluation on 700 synthetic images asking annotators to choose the winner image in each pair on three criteria of sensation evocation, message-image alignment, and perssuasion. Following table shows the agreement of human annotations with SenseScore and our strongest baseline (VQAScore). We will include the full table in the paper.

MetricAgreement (95% CI)pCohen's κ (95% CI)p (κ)
VQAScore66.3% (58.8–73.0)3×10⁻⁵0.31 (0.16–0.46)7×10⁻⁵
SenseScore77.0% (70.0–82.7)2×10⁻¹²0.55 (0.42–0.66)2×10⁻¹³

Regarding the size of human evaluation, firstly, we would like to clarify that the size of validation is 350 images including 300 real and 50 generated images. However, to further address this the above image-image comparison is on 450 pairs (700 unique images) of synthetic images, expanding the number of synthetic images in our human evaluation by 15x.

[W2/Q2] ... SAGA ablation testing the benefit of separating the agents or prioritizing message alignment before sensation

We conducted additional ablations on 270 images using the same FLUX generator, FluxKontext editor, and refinement stopping criteria (Sec. 3.3, L293–301; Fig. 4)to evaluate SAGA's task-specific design choices -- will include the table in the updated version.

Removing evaluator priority increases SenseScore (0.85 → 0.93) but decreases AIM (0.69 → 0.67), showing that stronger sensation evocation can come at the expense of message alignment. Removing the text refiner also decreases both metrics. We'll include these results in the revised paper.

MethodSenseScore ↑AIM ↑
SAGA (Full)0.850.69
SAGA w/o Priority0.930.67
SAGA w/o Text Refiner0.830.67
SAGA w/o Critic0.830.65
DALLE3 (Baseline)0.790.67

We further clarify that our main contribution is introducing and studying sensory advertisement generation. To this aim, SAGA is developed to establish a strong baseline for this task anddemonstrate the limitations of direct T2I generation, rather than proposing a optimal agentic method. Hence, comparisons against existing general-purpose agentic refinement methods are not central to our contribution. We hope our task, dataset, metric, and baseline will encourage future work to develop more effective generation methods for sensory advertising.

... while the persuasion claim relies on Pcomp ...CAP (Aghazadeh and Kovashka, 2025).

Thanks, we address this by conducting new human evaluation on 450 images pairs to include persuasion, alongside sensation evocation and message alignment, comparing SAGA against both FLUX and DALLE3.For persuasion, SAGA is preferred in 75% and 63% of human persuasion comparisons against FLUX and DALLE3, respectively, providing human evidence beyond the automatic Pcomp metric. See new section "EXTENDED HUMAN EVALUATION OF SAGA" in Appendix in the updated paper for details on all metircs.

Reviewer_V2E3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates how advertisements evoke senses in the recipients through visual means. The authors develop a hierarchical sensation taxonomy and use it to collect annotations on 1100 ad images sourced from an existing dataset of advertisements. The authors propose SenseScore, an evaluation method to measure how effectively an image evokes a target sensation. They propose a framework to generate ad image descriptions using an MLLM and fine-tuning an LLM using CPO and hierarchical loss to ensure a child does not get a higher probability than its parent as per the taxonomy. The authors introduce SenseGen, the task of generating advertisement images that evoke a specified sensation and convey the ad message and introduce SAGA, a multi-agent image generation framework for SenseGen. The authors have evaluated multiple different metrics for scoring how effectively an image evokes a target sensation using their collected data by measuring the agreement between metric and human annotations. They showcase how SenseScore can capture the task much better than existing metrics. Further, the authors showcase how SAGA generates better ad images with specified sensation than existing T2I methods such as DallE, Flux etc. The authors conduct human evaluation to strengthen SAGA’s improvement over the best baseline.

Critical strengths

The authors investigate a relevant and underexplored problem: understanding how visual advertisements evoke sensory experiences in viewers. They demonstrate the limitations of existing evaluation metrics in capturing the extent to which an image evokes a target sensation and introduce SenseScore as a more effective alternative. The sound training methodology and stronger agreement with human judgments provide evidence supporting the effectiveness of the proposed metric. Furthermore, the annotated Sensory Ad dataset is a valuable contribution that can facilitate future research on sensory understanding and generation for visual advertising.

Critical weaknesses
  1. The SAGA architecture does not appear to introduce substantial methodological novelty, as it primarily combines multiple agents to iteratively refine and evaluate generated advertisements. The authors could better clarify how this design differs from existing agent-based image-generation and self-refinement frameworks, and what specific benefits its architectural choices provide.

  2. The evaluation of SAGA would benefit from reporting confidence intervals, such as 95% bootstrap confidence intervals, for the key metrics and pairwise differences against baseline methods. Without these estimates, it is difficult to assess the robustness of the reported improvements or determine whether the observed differences are statistically meaningful.

  3. The study examining the relationship between sensation evocation, message alignment, and persuasiveness suggests that these dimensions do not necessarily improve together. However, Line 456 claims that there is an optimal level of sensation evocation and that SAGA successfully identifies it based on its higher scores. This conclusion appears stronger than the evidence presented. The authors should provide further validation, ideally through controlled comparisons or human evaluations, to establish what constitutes an optimal level of sensation evocation and whether SAGA consistently achieves it.

Questions for authors
  1. For SenseScore, rather than using the MLLM for generating the description of the image then finetuning an LLM, did the authors conduct any ablation where MLLM was finetuned directly?

  2. In table 3, authors report the evaluation numbers of different image generation methods for SenseGen task comparing SAGA with baselines. Could the authors also report 95% confidence intervals (e.g., via bootstrap resampling) and, where applicable, statistical significance tests for pairwise comparisons? This would help assess the robustness of the reported differences and determine whether the observed improvements over the baseline are statistically meaningful. Same would be helpful for the numbers reported in Table 2 for SenseScore ablations.

  3. In figure 8, which method is used to generate the left image?

  4. Could the authors explain how Persuasion and SenseScore are related? Fig 7b indicates two methods have persuasion staying almost the same as SenseScore increases while it first stays constant then decreases then increases sharply for AuraFlow. How does this relate to discussion on line 458?

  5. Minor suggestions:

In line 186 and 373, having the link to the corresponding appendix section would improve the readability of the draft.

Sensation Intensity, SenseScore and EvoSense are used interchangeably in the text and in Fig. 5. A quick note in the beginning clarifying this would be helpful.

In table 4 single-level classification, could the authors clarify line 1055 “…develop a better understanding of the abstract, semantic relationships between sensations in the hierarchy”

3 comments
AuthorsOpenReview ↗

Response (Part 1/2)

We thank Reviewer V2E3 for their thoughtful feedback, we are glad they found the problem relevant and underexplored, our evaluation metric and experiments/training sound and effective, and our dataset a valuable contribution.

We address each concern and question inline.

[W1] SAGA Novelty. The SAGA ... methodological novelty... clarify how this design differs from existing agent-based image-generation and self-refinement frameworks, ...

Thanks! We would like to clarify that the main contribution of our work is introducing and systematically studying the new task of sensory advertisement generation. To this end, we introduce a new dataset and taxonomy, SenseGen task, and SenseScore, a novel evaluation metric that achieves strong agreement with human judgment. Developing a fundamentally novel agentic method was not the goal of this paper, we see SAGA as a task-specific framework to demonstrate the limitations of existing T2I models in simultaneously conveying an intended message and evoking a target sensation, and to establish a strong baseline for this new task. We hope our task, dataset, metric, and baseline will encourage future work to develop more effective generation methods for sensory advertising. We will further clarify this in the revised paper.

To further evaluate SAGA's design choices, we conducted additional ablations on 270 images:

MethodSenseScore ↑AIM ↑
SAGA (Full)0.850.69
SAGA w/o Priority0.930.67
SAGA w/o Text Refiner0.830.67
SAGA w/o Critic0.830.65
DALLE3 (Baseline)0.790.67

Interestingly, removing evaluator priority increases SenseScore but reduces message alignment, demonstrating that maximizing sensation evocation alone can compromise the intended message. Removing the text refiner also decreases both metrics. These results support the motivation behind SAGA's task-specific design, which prioritizes message communication while encouraging sensation evocation. We will include these ablations and example outputs in the revised paper.

[W2/Q2] SAGA Evaluation...report 95% confidence intervals...statistical significance tests for pairwise comparisons. ....

Thanks for the constructive feedback. We've updated Table1/Table2/Table3, including CI along with paired significance test comparing against SAGA (see updated table in the revised paper). In Table 2, Cohen's κ increases from 0.33 [0.30, 0.37] (SFT) to 0.57 [0.56, 0.61] (CPO), and 0.69 [0.66, 0.71] (CPO + hierarchy). In Table 3, The results confirm that SAGA significantly improves message alignment over all baselines (AIM, p < 1e-5). For SenseScore and Pcomp, SAGA also significantly outperforms five of the six baselines. Here we show the parts of Table 1 and Table 3 that represents our best design choice and best baselines. Please refer to paper for full tables:

SenseScore agreement with human

ModelAgreement95% CIp-value
VQAScore0.53[0.51, 0.56]< 0.0001
MLLM as a Judge - InternVL0.46[0.43, 0.50]< 0.0001
Evosense-LLAMA3-InternVL0.69[0.66, 0.71]< 0.0001

SensoryAd Evaluation

(SenseScore)

ModelMean [95% CI]SAGA − model (p-value)
DALLE3+0.79 [+0.71, +0.86]+0.04 (0.19)
QwenImage+0.74 [+0.67, +0.81]+0.09 (0.015)
SAGA+0.83 [+0.77, +0.90]n/a

(AIM)

ModelMean [95% CI]SAGA − model (p-value)
DALLE30.63 [0.63, 0.64]+0.02 (< 1e-5)
QwenImage0.61 [0.60, 0.62]+0.04 (< 1e-5)
SAGA0.66 [0.65, 0.66]n/a

Persuasion

ModelMean [95% CI]SAGA − model (p-value)
DALLE30.61 [0.61, 0.63]+0.01 (0.14)
QwenImage0.57 [0.56, 0.58]+0.05 (< 1e-5)
SAGA0.62 [0.62, 0.63]n/a
AuthorsOpenReview ↗

Response (Part 2/2)

[W3] Optimal sensation evocation. The study ... suggests that these dimensions do not necessarily improve together. However, ... claims that ... an optimal level of sensation evocation and that SAGA successfully identifies ... appears stronger than the evidence presented....

Thanks for the constructive feedback. We clarify that SAGA better balances sensation evocation and message alignment, rather than identifying an optimal sensation intensity. We revised L461–463 to clarify this point and avoid confusion: SAGA achieves stronger sensation evocation while preserving message alignment, demonstrating the benefit of jointly considering these objectives rather than maximizing sensory intensity alone.

Specifically, Figs. 7a–b show that stronger sensation evocation does not necessarily improve message alignment or persuasion, with some methods initially improving and then declining as sensation intensity increases. In contrast, SAGA achieves the highest SenseScore and AIM (Table 3), supporting its ability to balance both objectives. We further conducted an additional human evaluation comparing SAGA and FLUX on both objectives. Interestingly, When annotators preferred SAGA for sensation evocation, they also preferred it for message alignment in 96% of cases, compared with 66% when FLUX was preferred for sensation evocation. Moreover, the new Fig. 7c shows that when FLUX is preferred for sensation, agreement with message alignment decreases at higher sensation intensities (87% to 58%), due to the exaggeration issue. These findings provide additional evidence that higher sensory evocation does not always translate to higher message alignment, and effectiveness of SAGA by jointly considering both objectives.

[Q1] Direct MLLM fine-tuning. ...ablation where the MLLM was fine-tuned directly?

Thanks for constructive feedback. Our main focus on design of SenseScore metric is on optimization of the method. We conducted an additional experiment directly fine-tuning InternVL (the main MLLM backbone in SenseScore). The directly fine-tuned MLLM achieves Cohen's κ = 0.592 [0.569, 0.615] compared with κ = 0.69 [0.66, 0.71] for SenseScore. This result further supports our two-stage design choice for main metric, where an MLLM describes the visual content and a fine-tuned LLM evaluates sensation evocation. We will include this ablation in the revised paper.

[Q3] In figure 8, which method is used to generate the left image?

All left images in Fig.8 were produced by FLUX.1-dev. The corresponding right-side images are real ads. We will clarify this in the figure caption.

[Q4] Could the authors explain how Persuasion and SenseScore are related? ....

Prior marketing research (cited in paper) show that evoking sensations (e.g., refreshing coolness) can improve persuasion by helping viewers mentally imagine themselves in the situation pictured in the advertisement image. However, sensory evocation is only one of several factors contributing to persuasion, and other factors can influence the overall score. Since we cannot control these factors independently, we focus on the overall trend across models rather than expecting a monotonic increase for each individual model. Fig.7b shows and overall tendency for persuasion to increase with sensory evocation. We will revise L458-463 to clarify this point in the paper.

[Q5] minor

Thanks, we have fixed accordingly (highlighted in blue)

Thank you for the rebuttal and performing the direct MLLM fine-tuning ablation. The authors have also shared the confidence intervals as suggested in the review. However, I remain unconvinced regarding the SAGA framework:

Against DALL-E 3, SAGA’s improvements are not statistically significant for SenseScore and Persuasion. Failing to meaningfully outperform a standard prompt-expansion baseline like DALL-E 3 undermines its practical value. Incremental Architecture: As the authors note, SAGA functions as a standard multi-agent wrapper rather than introducing novel methodology. While the dataset and SenseScore metric are solid contributions, the SAGA generation framework lacks a statistically robust edge over existing strong baselines. My rating remains a weak reject.

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

Related papers

Discussion (0)

Sign in to comment.