Pith. sign in

REVIEW 4 major objections 6 minor 45 references

Sensory advertising becomes computable: a metric matches human judgment, and a generator produces more persuasive ads.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:47 UTC pith:PHI45N3M

load-bearing objection First real benchmark for sensation evocation in ads, worth refereeing, but the headline Kappa is an inflated one-vs-all ranking and should be reframed. the 4 major comments →

arxiv 2607.25314 v1 pith:PHI45N3M submitted 2026-07-28 cs.CV

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

Aysan Aghazadeh , Sina Malakouti , Adriana Kovashka This is my paper
classification cs.CV
keywords sensory advertisingsensation evocationtext-to-image generationevaluation metricadvertisement understandingmulti-agent generationhierarchical taxonomyvision-language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to make sensory advertising—ads that evoke a specific feeling like cold, taste, or pressure through imagery—computationally tractable. To do so it introduces a dataset of ads annotated with a 96-label hierarchy of sensations, two classification benchmarks, an automatic scorer (SenseScore), and a generator (SAGA). Its central claim is that a fine-tuned language model, fed an MLLM-generated description, scores sensation intensity with strong human agreement (κ=0.80 on real ads; 40% relative improvement over the best baseline). A second claim is that a multi-agent editing loop, unlike one-shot text-to-image generation, balances message alignment, sensation evocation, and persuasion. If these claims hold, visual persuasion gains a measurable sensory dimension and a way to generate it.

Core claim

The paper's central discovery is that sensation evocation in advertisements is measurable and generable. On the measurement side, a two-stage pipeline—an image-describing MLLM, then an LLM fine-tuned with a hierarchy-aware contrastive objective—yields SenseScore, which assigns an intensity to each of 96 sensations. The objective (L_CPO+Hierarchy) forces the model to prefer the more strongly evoked sensation while keeping a parent sensation's probability above its child's, encoding the taxonomy's structure. Fine-tuned on preference pairs built from 100 annotated ads, SenseScore reaches κ=0.80 on 100 held-out real ads and κ=0.68 on AI-generated ads, versus κ=0.57 for the best baseline. On the

What carries the argument

The load-bearing mechanism is a hierarchy-aware contrastive preference objective applied to a fine-tuned LLM. Given an MLLM-generated description of an ad image, the LLM emits log-probabilities for each of 96 sensations; the training objective L_CPO+Hierarchy makes the model prefer the more strongly evoked sensation over the weaker one while keeping the parent sensation's probability above the child's. This encodes both relative strength and the taxonomy's parent–child structure in a single scalar per sensation, which is SenseScore's intensity value. On the generation side, the central mechanism is SAGA's iterative agent loop—an editing planner outputs structured actions, a prompt refiner tu

Load-bearing premise

SenseScore's κ=0.80 rests on fine-tuning a language model on preference pairs derived from 100 hand-annotated ads (judged by 12 US-based annotators) and on a Kappa computation that drops human-tied pairs; if that sample or that tie-handling is unrepresentative, the agreement will not generalize.

What would settle it

Run SenseScore on a fresh set of, say, 200 ads sampled from a different region or medium and compare with fresh annotators, keeping all tied pairs in the Kappa calculation; if agreement falls below 0.6 or the tie-keeping changes the ranking of metrics, the claim of strong human agreement is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Advertisers and platforms can filter or flag ads that evoke inappropriate sensations (e.g., pain or alcohol-related cues) for sensitive audiences, since the classifier and scorer are automated.
  • SenseScore can serve as a cheap automatic reward for text-to-image models, letting designers optimize for sensation without running human studies.
  • The inverted-U relation between sensation intensity and message alignment (Fig. 7a) suggests generators should target a moderate sensory intensity; SAGA's advantage is partly that it finds that sweet spot.
  • The demonstrated transfer to AI-generated ads (κ=0.68) means the metric can be embedded in generative pipelines as a feedback signal, not just an offline evaluator.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because SenseScore's training pairs all come from 100 images, the metric's reliability per-sensation likely varies with annotation frequency; rare sensations (e.g., certain sounds) may be scored less accurately than common ones (e.g., cold or sweetness). A stratified evaluation by leaf sensation would expose this.
  • The reported agreement ignores human-tied pairs; recomputing κ with ties scored as half-agreement, or using rank correlation on raw intensities, would give a more conservative estimate of how close SenseScore really is to humans.
  • The finding that non-ad sensory images score lower on SenseScore than sensory ads suggests the metric is partly encoding rhetorical, ad-specific cues; a direct test would be to apply SenseScore to everyday photographs and see whether it flags them as sensation-weak, which would limit its use beyond advertising.
  • If the trend of SAGA improving AIM and Pcomp holds, the framework could be repurposed as a general 'creativity critic' for aligning any abstract prompt with an image, not just sensation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces the Sensory Ad dataset (670 PittAds images annotated with a 96-label hierarchical sensation taxonomy, evocation-strength scores, and contributing visual elements), two Sensation Classification (SenseClass) tasks, an automatic sensation-evocation metric (SenseScore) that combines an MLLM image description stage with a fine-tuned LLM trained via a hierarchy-aware contrastive preference objective, and a multi-agent generation framework (SAGA) for the new Sensory Ad Generation (SenseGen) task. The central claims are that SenseScore 'achieves strong agreement with human judgments' (abstract, κ=0.80 in Table 2), and that SAGA 'improves message–image alignment, sensory evocation, and persuasion' (abstract, Table 3). The paper also analyzes the trade-off between sensation intensity, message alignment, and persuasiveness in generated ads.

Significance. If the claims hold, the paper would provide the first computational benchmark for a previously unstudied dimension of visual persuasion, a reusable evaluation metric, a dataset with a structured taxonomy, and an agentic generation framework. The authors ship code, release the dataset, and include prompts and detailed supplement materials, which are concrete strengths. However, the central metric-validation protocol is nonstandard and appears to inflate agreement, and the generation results are evaluated with the proposed metric alongside metrics from the authors' prior work. The significance is therefore conditional on a re-validation of SenseScore with standard agreement/ranking metrics and on a clearer account of its generalization beyond the 100 training images.

major comments (4)
  1. [§4.2, Supplement D.1, Table 5] The headline Kappa is not a standard agreement measure. Supplement D.1 states: 'We ignore the sensation pairs where the human annotators assign the same score to both sensations.' Since annotators select at most 3 of 96 sensations, nearly all ignored pairs are (0,0) ties; the surviving items are overwhelmingly one-vs-all comparisons of a selected sensation against an unselected one. This turns the evaluation into a pairwise ranking task, not a full-label agreement task, and mirrors the CPO training objective in Eq. (2). The Supplement's own Pearson r=0.38 vs κ=0.86 (Table 5) quantifies the gap. The claim of 'near-perfect agreement with human' (κ=0.85) is therefore not supported in the standard sense. Please report Cohen's kappa on the binarized 96-label vectors (including ties), Pearson/Spearman correlation on the intensity scores, and a ROC-AUC or similar one-vs-all ranking metric, and
  2. [§4.2, Tables 2, 5, 6] The reported Kappa values are inconsistent across evaluation sets and protocols: Table 2 gives κ=0.80 on 100 real images, the text in §4.2 states 'κ=0.85 on real ads', Table 5 reports κ=0.86 on '5000 real' pairs, and Table 6 reports κ=0.78 on 400 real images. The sample sizes also need reconciliation: Table 2 says 100 images/≈10,000 pairs, Table 5 says 5,000 real pairs (which would be ≈52 images at 96 pairs per image), and Table 6 uses 400 images. Please provide a single, clearly defined evaluation protocol with the exact number of images, pairs, and annotation sources for every table, or explain why the numbers differ (e.g., different subsets of sensations or images).
  3. [§3.3, Supplement C, Table 2] The zero-shot SenseScore has κ≈0 for both LLAMA3 and QwenLM, and the fine-tuned model is trained on 40,000 pairs derived from only 100 images. The reported 0.80 is therefore essentially a measure of how well the model fits those 100 images' labels, not a demonstration of a general property of sensation evocation. Table 6 shows robustness to 400 images from the same distribution and annotation pool, which helps, but this is still a single, small human-preference set. Please add image-level cross-validation (e.g., train on 80 images, test on the remaining 20, repeated) and, ideally, an independent external sample of ads, to show that the fine-tuned ranking transfers beyond the exact training images. Until then, 'strong agreement with human judgments' should be qualified as agreement on the specific annotation sample.
  4. [§4.3, Table 3] The SAGA generation improvements are small in absolute terms and are reported without uncertainty: SenseScore increases from 0.97 to 0.99 for Flux, AIM from 0.43 to 0.49, Pcomp from 0.54 to 0.61. Given that SenseScore's validity is the subject of Major Comment 1, and AIM/Pcomp come from the authors' prior work, the claim that SAGA 'improves ... persuasion' should be backed by significance tests or confidence intervals over the 600 images, and by showing that the SenseScore differences are not within the metric's own noise level. The non-monotonic relationship in Fig. 7a is used to argue that SAGA finds an 'optimal level'; please provide quantitative evidence (e.g., the location of the peak and how SAGA's mean intensity differs from the baselines' means).
minor comments (6)
  1. [Conclusion] Redundant 'significantly' in 'significantly outperforming existing baselines significantly' (p. 14, last paragraph).
  2. [§3.3, Eq. (2)] The abbreviation CPO is used in the introduction and Eq. (2) but not expanded at first use; define 'Contrastive Preference Optimization' where it appears.
  3. [Table 2] The 'Taste' column for VQA-score shows 'no'—please clarify whether this is a missing value, zero, or not applicable.
  4. [§4.2, Tables 2 and 6] The 95% confidence intervals (e.g., [0.80, 0.81]) appear to treat the 100×96 pairs as independent, but pairs from the same image are highly dependent. Please compute image-level bootstrapped CIs or otherwise account for clustering.
  5. [Supplement D.1, Table 5] The discrepancy between Table 2 (κ=0.80) and Table 5 (κ=0.86) for 'SenseScore (LLAMA3 + D_InternVL)' is confusing; the main text should cite a single consistent value and explain the source of the difference.
  6. [References] Reference [4] (DALL-E 3) has a malformed URL in the bibliography ('https://cdn. openai. com/papers/dall-e-3. pdf2(3), 8 (2023)').

Circularity Check

0 steps flagged

No significant circularity; SenseScore's agreement is an empirical supervised fit, and the only self-citation (AIM/Pcomp) is secondary.

full rationale

I walked the paper's derivation chain and found no step where a prediction is equivalent to its input by construction. SenseScore is a supervised metric: zero-shot agreement with humans is near zero (Table 2: κ≈-0.03 to -0.06), and after fine-tuning on human-annotated preference pairs it reaches κ≈0.80 on held-out images. The paper explicitly states that fine-tuning is essential (§4.2), so the reported agreement is an empirical fit, not a disguised derivation. The evaluation uses image sets separate from those used for training (Supplement C; §4.2), so the improvement is not statistically forced on the training data. The main self-citation is the use of AIM and Pcomp from the authors' prior work [1] to evaluate SAGA; these are published metrics and their use is secondary to the paper's core sensory contribution. The κ protocol in Supplement D.1 discards tied sensation pairs, which may inflate the agreement number, but this is a disclosed methodological limitation, not a circular reduction: the surviving pairs are still on held-out images and the model could fail them. No equation in the paper reduces a predicted quantity to a fitted parameter or to the evaluation label set.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claims rest on the author-defined taxonomy and human annotations; no new physical entities are postulated. The main fitted component is the SenseScore LLM itself, and the main non-fitted assumptions are the representativeness of the dataset, the validity of the taxonomy, and the faithfulness of automated image descriptions.

free parameters (4)
  • SenseScore fine-tuned LLM weights (LoRA) = LoRA on LLAMA3-8B or Qwen2.5-7B; 40,000 preference pairs from 100 human-annotated images; lr 5e-5
    The central metric's agreement with humans is produced by these fitted weights, not by a closed-form derivation. The paper's own zero-shot baseline shows near-zero agreement, so essentially all predictive power is imported from the fitted model.
  • CPO temperature β in Eq. (2) = not reported
    The hierarchy-aware contrastive loss uses β to scale log-probability differences. The paper does not report its value or sensitivity, and a different β could change preference rankings.
  • Maximum number of SAGA dialogue iterations = not reported
    The agent loop repeats 'until reaching a maximum number of messages or No Issue' (§3.4). The maximum is not specified, and generation quality may depend on the stopping criterion.
  • Maximum of 3 sensations per image = 3
    The annotation protocol and classification tasks cap predictions at three sensations (§3.1). This truncates the label space and shapes the SenseScore training pairs.
axioms (5)
  • domain assumption The 96-label sensation hierarchy (5 root senses with hand-defined children) is a valid and complete representation of the sensations an ad can visually evoke.
    Defined by the authors in §3.1; no external validation that the taxonomy is exhaustive, balanced, or that the parent-child relations hold perceptually. All downstream labels and metrics inherit this structure.
  • domain assumption Human annotation scores from 12 Prolific workers are ground truth for sensation intensity.
    Inter-annotator κ=0.83 on ~10% of images supports reliability but does not establish validity. SenseScore is trained and evaluated against these scores; if the annotators are not representative, the metric inherits their biases.
  • domain assumption MLLM-generated image descriptions are faithful and free of interpretation.
    SenseScore and the LLM classification pipelines depend entirely on textual descriptions from InternVL/QwenVL/Gemma (Supplement C). If descriptions omit or hallucinate sensory cues, the metric's inputs are corrupted. The paper argues zero-shot baselines rule out information leakage, but description quality itself is not directly measured.
  • domain assumption The 670-image PittAds subset is representative of real advertisements.
    Sampled from one existing dataset [17]; no cross-domain validation is provided. SenseScore is trained on only 100 of these images and evaluated on held-out images from the same source.
  • domain assumption By definition, evoking a child sensation implies evoking its parent.
    Used for hierarchical consistency and Parent Recall in §3.2. This is a definitional modeling choice, not an empirically established fact about perception.

pith-pipeline@v1.3.0-alltime-deepseek · 18830 in / 12864 out tokens · 137154 ms · 2026-08-01T02:47:15.092439+00:00 · methodology

0 comments
read the original 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 understanding, evaluating, and generating sensory ads. We introduce the Sensory Ad dataset, and define sensation classification tasks (SenseClass) to benchmark LLMs and MLLMs. We further propose SenseScore, an automated evaluation metric for sensation evocation achieving strong agreement with human judgments. Finally, 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.

Figures

Figures reproduced from arXiv: 2607.25314 by Adriana Kovashka, Aysan Aghazadeh, Sina Malakouti.

Figure 1
Figure 1. Figure 1: Sensory Ad Generation: We show two rows containing intended ad messages and possible images for these messages. The message text is shown in a box (“I should... because...”). The first column of images (a, c) are images generated by a T2I model using the message as prompt. The images in the second column (b, d) are real human￾designed ads. The third column (e, f) are ads designed by a T2I model which also … view at source ↗
Figure 2
Figure 2. Figure 2: Left: Sensation hierarchy (only categorizing the Touch sensation): Green box represents first level sensation, blue represents second, and purple rep￾resents third. Right: Annotation example from our Sensory Ad dataset. affects their perception, judgment and behavior.” Subconscious sensory triggers may make the ad’s message more compelling than explicit messaging, including causing viewers to perceive spec… view at source ↗
Figure 3
Figure 3. Figure 3: SenseScore evaluation method. Left: training of LLM with two different sets of sensations for one image. Green border shows the winner sensation, blue border represents the parent of the winner (used in hierarchy loss), and red border denotes the loser in the pair. Each pair is derived from a triplet of annotations, where A is preferred over B, and B over C. Right: score computation in inference with the f… view at source ↗
Figure 4
Figure 4. Figure 4: SAGA framework. Generation and editing process in multi-agent framework. Blue box represents the conversation loop among the agents, which repeats until reach￾ing a maximum number of messages or ‘No Issue’ from the Evaluator Agent. Initial experiments using zero-shot LLMs show low agreement with human annotations, both in retrieving correct sensations and estimating their intensity. To address this, we fin… view at source ↗
Figure 5
Figure 5. Figure 5: Examples on human agreement with SenseScore and VQA-score [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensory Ad examples. Three examples of real ads and ads generated by Flux [5], SD3 [11], AuraFlow [12], PixArt [6], DALLE3 [4] and SAGA (ours) given the action-reason message and sensation annotation for the real advertisement. Green border represents the sensation used in the prompt of T2I models. sage and sensation (AR + Sensation). In supp, we also benchmark generating images that evoke sensations witho… view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of the relation between sensation intensity, text-image alignment, and persuasion in generated and real advertisements. the intensity. Importantly, it also outperforms other methods in terms of alignment with the intended message (AIM) and persuasion (Pcomp) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples of exaggeration in sensation evocation. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Image Distribution. Left: percentage of images per sensation category (sensation diversity). Right: distribution over the 10 most frequent topics (topic diversity). C Experimental Setup & Implementation Details In this section we explain the experimental setup. Hugging Face implementation of models are utilized, and code will be released upon the acceptance. Sensation Classification In sensation classifica… view at source ↗
Figure 10
Figure 10. Figure 10: Sensation Hierarchy. First level, represents the main five human sensations, and each sensation is categorized into different set of sensations [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Image Description Examples. Two examples of descriptions generated by InternVL, Gemma, and QwenVL. Both images are real advertisements from PittAd [17] dataset. evaluation of SenseScore performance was on a subset of the images not selected for training [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Conversation loop in SAGA. An example of two steps of conversation loop among agents with FluxKontext image editing. The input of the conversation loop is the advertisement message, target sensation, and the image generated by Flux for conveying the advertisement messages and targeting the sensation [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of human scores and metric scores for each sensation’s intensity evoked by the the corresponding image (right). and the baselines) choose different scores for each sensation. For computing κ agreement, we use the sensation intensity as the criteria for choosing the winner sensation for the image for each pair of sensations. We ignore the sensation pairs where the human annotators assign the sam… view at source ↗
Figure 14
Figure 14. Figure 14: Sensory Ad examples. Four examples of real advertisement and generated advertisements by Flux [5], SD3 [11], AuraFlow [12], and PixArt [6] given the action￾reason interpretation and sensation annotation for the real advertisement. Green border represents the sensation used in the prompt of T2I models [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Sensation Heatmap. Average SenseScore score for images generated by each model for each sensation. Each model generates ten images evoking each sensation [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 4 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Aghazadeh, A., Kovashka, A.: Cap: Evaluation of persuasive and creative image generation. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 16970–16980 (2025)

    2025

  2. [2]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Akula, A.R., Driscoll, B., Narayana, P., Changpinyo, S., Jia, Z., Damle, S., Pruthi, G., Basu, S., Guibas, L., Freeman, W.T., et al.: Metaclue: Towards comprehen- sive visual metaphors research. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 23201–23211 (2023)

    2023

  3. [3]

    Advances in Neural Information Processing Systems36, 7351–7360 (2023)

    Bender, T., Sørensen, S., Kashani, A., Eldjarn Hjorleifsson, K., Hyldig, G., Hauberg, S., Belongie, S., Warburg, F.: Learning to taste: A multimodal wine dataset. Advances in Neural Information Processing Systems36, 7351–7360 (2023)

    2023

  4. [4]

    https://cdn

    Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee,J.,Guo,Y.,etal.:Improvingimagegenerationwithbettercaptions.Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf2(3), 8 (2023)

    2023

  5. [5]

    Black Forest Labs: Flux.https://github.com/black-forest-labs/flux(2024)

    2024

  6. [6]

    In: ICLR (2024)

    Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024)

    2024

  7. [7]

    Journal of marketing research51(2) (2014)

    Cian, L., Krishna, A., Elder, R.S.: This logo moves me: Dynamic imagery from static images. Journal of marketing research51(2) (2014)

    2014

  8. [8]

    Educational and psycho- logical measurement20(1), 37–46 (1960)

    Cohen, J.: A coefficient of agreement for nominal scales. Educational and psycho- logical measurement20(1), 37–46 (1960)

    1960

  9. [9]

    arXiv preprint arXiv:2501.05710 (2025)

    Dang, S., He, Y., Ling, L., Qian, Z., Zhao, N., Cao, N.: Emoticrafter: Text- to-emotional-image generation based on valence-arousal model. arXiv preprint arXiv:2501.05710 (2025)

    arXiv 2025

  10. [10]

    Journal of Consumer Psychology32(2), 293–315 (2022)

    Elder, R.S., Krishna, A.: A review of sensory imagery for consumer psychology. Journal of Consumer Psychology32(2), 293–315 (2022)

    2022

  11. [11]

    In: Forty-first international conference on machine learning (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024)

    2024

  12. [12]

    Fal: Introducing auraflow v0.1, an open exploration of large rectified flow models (2024), available at:https://blog.fal.ai/auraflow/[Accessed: 2024-11-07]

    2024

  13. [13]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Ghosh, S., Kumar, S., Seth, A., Evuru, C.K.R., Tyagi, U., Sakshi, S., Nieto, O., Duraiswami, R., Manocha, D.: Gama: A large audio-language model with advanced audio understanding and complex reasoning abilities. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 6288–6313 (2024)

    2024

  14. [14]

    In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

    Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference- free evaluation metric for image captioning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514–7528 (2021)

    2021

  15. [15]

    In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf9

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf9

    2022

  16. [16]

    Routledge (2015) 16 Aysan Aghazadeh, Sina Malakouti, and Adriana Kovashka

    Hultén, B.: Sensory marketing: Theoretical and empirical grounds. Routledge (2015) 16 Aysan Aghazadeh, Sina Malakouti, and Adriana Kovashka

    2015

  17. [17]

    In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition

    Hussain, Z., Zhang, M., Zhang, X., Ye, K., Thomas, C., Agha, Z., Ong, N., Ko- vashka, A.: Automatic understanding of image and video advertisements. In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition. pp. 1705–1715 (2017)

    2017

  18. [18]

    In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Kim, J., Han, J., Choi, D., Yoon, J., Lee, E.J., Jo, Y.: Pvp: An image dataset for personalized visual persuasion with persuasion strategies, viewer characteristics, and persuasiveness ratings. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 19209– 19237 (2025)

    2025

  19. [19]

    Advances in neural information processing systems36, 36652–36663 (2023)

    Kirstain,Y.,Polyak,A.,Singer,U.,Matiana,S.,Penna,J.,Levy,O.:Pick-a-pic:An open dataset of user preferences for text-to-image generation. Advances in neural information processing systems36, 36652–36663 (2023)

    2023

  20. [20]

    Journal of consumer psychology22(3), 332–351 (2012)

    Krishna, A.: An integrative review of sensory marketing: Engaging the senses to affect perception, judgment and behavior. Journal of consumer psychology22(3), 332–351 (2012)

    2012

  21. [21]

    Current Opinion in Psychology10, 142–147 (2016)

    Krishna, A., Cian, L., Sokolova, T.: The power of sensory marketing in advertising. Current Opinion in Psychology10, 142–147 (2016)

    2016

  22. [22]

    Journal of consumer psychology24(2), 159–168 (2014)

    Krishna, A., Schwarz, N.: Sensory marketing, embodiment, and grounded cogni- tion: A review and introduction. Journal of consumer psychology24(2), 159–168 (2014)

    2014

  23. [23]

    In: Proceedings of the AAAI conference on artificial intelligence

    Kumar, Y., Jha, R., Gupta, A., Aggarwal, M., Garg, A., Malyan, T., Bhardwaj, A., Shah, R.R., Krishnamurthy, B., Chen, C.: Persuasion strategies in advertisements. In: Proceedings of the AAAI conference on artificial intelligence. vol. 37, pp. 57–66 (2023)

    2023

  24. [24]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Liao, J., Chen, X., Fu, Q., Du, L., He, X., Wang, X., Han, S., Zhang, D.: Text-to- image generation for abstract concepts. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3360–3368 (2024)

    2024

  25. [25]

    In: ECCV (2024)

    Lin, Z., Pathak, D., Li, B., Li, J., Xia, X., Neubig, G., Zhang, P., Ramanan, D.: Evaluating text-to-visual generation with image-to-text generation. In: ECCV (2024)

    2024

  26. [26]

    Strategic Direction22(2) (2006)

    Lindstrom, M.: Brand sense: How to build powerful brands through touch, taste, smell, sight and sound. Strategic Direction22(2) (2006)

    2006

  27. [27]

    arXiv preprint arXiv:2403.12356 (2024)

    Menon, S., Wang, S., Chilton, L.: Moodsmith: Enabling mood-consistent multime- dia for ai-generated advocacy campaigns. arXiv preprint arXiv:2403.12356 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    The American Journal of Occupational Therapy69(1), 6901290010p1–6901290010p8 (2015)

    Nau, A.C., Pintar, C., Arnoldussen, A., Fisher, C.: Acquisition of visual perception in blind adults using the brainport artificial vision device. The American Journal of Occupational Therapy69(1), 6901290010p1–6901290010p8 (2015)

    2015

  29. [29]

    arXiv preprint arXiv:2511.20544 (2025)

    Ozguroglu, E., Liang, J., Liu, R., Chiquier, M., DeTienne, M., Qian, W.W., Horowitz, A., Owens, A., Vondrick, C.: New york smells: A large multimodal dataset for olfaction. arXiv preprint arXiv:2511.20544 (2025)

    arXiv 2025

  30. [30]

    In: Proceedings of the Asian Conference on Computer Vision (2020)

    Park, C., Lee, I.K.: Emotional landscape image generation using generative ad- versarial networks. In: Proceedings of the Asian Conference on Computer Vision (2020)

    2020

  31. [31]

    Journal of Interactive Market- ing45(1), 42–61 (2019)

    Petit, O., Velasco, C., Spence, C.: Digital sensory marketing: Integrating new technologies into multisensory online experience. Journal of Interactive Market- ing45(1), 42–61 (2019)

    2019

  32. [32]

    Journal of NeuroEngineering and Rehabilitation 22(1), 21 (2025) 17

    Prahm, C., Eckstein, K., Bressler, M., Wang, Z., Li, X., Suzuki, T., Daigeler, A., Kolbenschlag, J., Kuzuoka, H.: Phantomar: gamified mixed reality system for alleviating phantom limb pain in upper limb amputees—design, implementation, and clinical usability evaluation. Journal of NeuroEngineering and Rehabilitation 22(1), 21 (2025) 17

    2025

  33. [33]

    arXiv preprint arXiv:2510.22768 (2025)

    Qiu, H., Zhou, Y., Venkit, P.N., Huang, K.H., Zhang, J., Peng, N., Wu, C.S.: Mm- persuade: A dataset and evaluation framework for multimodal persuasion. arXiv preprint arXiv:2510.22768 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    arXiv preprint arXiv:2405.00942 (2024)

    Singh, S., SI, H., Singla, Y.K., Baths, V., Shah, R.R., Chen, C., Krishnamurthy, B.: Teaching human behavior improves content understanding abilities of llms. arXiv preprint arXiv:2405.00942 (2024)

    Pith/arXiv arXiv 2024

  35. [35]

    In: ICLR (2025)

    Singh, S.K., Singla, Y.K., Krishnamurthy, B., et al.: Measuring and improving persuasiveness of large language models. In: ICLR (2025)

    2025

  36. [36]

    Cognition89(3), 237–261 (2003)

    Ward, J., Simner, J.: Lexical-gustatory synaesthesia: linguistic and conceptual fac- tors. Cognition89(3), 237–261 (2003)

    2003

  37. [37]

    arXiv preprint arXiv:2508.02324 (2025)

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    Pith/arXiv arXiv 2025

  38. [38]

    In: International Conference on Machine Learning

    Xu, H., Sharaf, A., Chen, Y., Tan, W., Shen, L., Van Durme, B., Murray, K., Kim, Y.J.: Contrastive preference optimization: Pushing the boundaries of llm perfor- mance in machine translation. In: International Conference on Machine Learning. pp. 55204–55224. PMLR (2024)

    2024

  39. [39]

    Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    2023

  40. [40]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, F., Feng, C., Chen, Z., Park, H., Wang, D., Dou, Y., Zeng, Z., Chen, X., Gangopadhyay, R., Owens, A., et al.: Binding touch to everything: Learning unified multimodal tactile representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26340–26353 (2024)

    2024

  41. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, J., Feng, J., Huang, H.: Emogen: Emotional image content generation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6358–6368 (2024)

    2024

  42. [42]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhai, A.J., Shen, Y., Chen, E.Y., Wang, G.X., Wang, X., Wang, S., Guan, K., Wang, S.: Physical property understanding from language-embedded feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28296–28305 (2024)

    2024

  43. [43]

    Generate an image evoking sensation

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023) A Outline This supplement provides additional discussion of the dataset, experimental setup and implementation details, and extended results. It also discusses pot...

    2023

  44. [44]

    Convey the intended advertisement message, and

  45. [45]

    type_of_action

    Evoke the specified sensation (e.g., refreshment, heat, softness, luxury). When you receive an issue from the critic, you MUST focus your edits on addressing that SPECIFIC issue: Image-Message Alignment: The image does not clearly convey the adver- tisement message. Focus on making the product or brand more prominent, ensuring the image directly relates t...