A$^3$: Towards Advertising Aesthetic Assessment

Kaiyuan Ji; Yixuan Gao; Lu Sun; Yushuo Zheng; Zijian Chen; Jianbo Zhang; Xiangyang Zhu; Yuan Tian

cs.CVMar 25, 2026

A $^3$ : Towards Advertising Aesthetic Assessment

Authors: Kaiyuan Ji, Yixuan Gao, Lu Sun, Yushuo Zheng, Zijian Chen, Jianbo Zhang, Xiangyang Zhu, Yuan Tian, +2 more

Organizations: Shanghai Artificial Intelligence Laboratory · School of Information and Electronic Engineering, East China Normal University · Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University · School of Computer Science and Technology, Xi’an Jiaotong University

Abstract

Advertising images significantly impact commercial conversion rates and brand equity, yet current evaluation methods rely on subjective judgments, lacking scalability, standardized criteria, and interpretability. To address these challenges, we present A^3 (Advertising Aesthetic Assessment), a comprehensive framework encompassing four components: a paradigm (A^3-Law), a dataset (A^3-Dataset), a multimodal large language model (A^3-Align), and a benchmark (A^3-Bench). Central to A^3 is a theory-driven paradigm, A^3-Law, comprising three hierarchical stages: (1) Perceptual Attention, evaluating perceptual image signals for their ability to attract attention; (2) Formal Interest, assessing formal composition of image color and spatial layout in evoking interest; and (3) Desire Impact, measuring desire evocation from images and their persuasive impact. Building on A^3-Law, we construct A^3-Dataset with 120K instruction-response pairs from 30K advertising images, each richly annotated with multi-dimensional labels and Chain-of-Thought (CoT) rationales. We further develop A^3-Align, trained under A^3-Law with CoT-guided learning on A^3-Dataset. Extensive experiments on A^3-Bench demonstrate that A^3-Align achieves superior alignment with A^3-Law compared to existing models, and this alignment generalizes well to quality advertisement selection and prescriptive advertisement critique, indicating its potential for broader deployment. Dataset, code, and models can be found at: https://github.com/euleryuan/A3-Align.

Figures & tables

Figure 1: Overview of A 3 : Advertising Aesthetic Assessment. A 3 centers on the A 3 -Law, a three stage paradigm with Perceptual Attention , Formal Interest , and Desire Impact . Built on this paradigm, A 3 -Dataset contains 30K images and 120K instruction-response pairs with Chain of Thought; A 3 -Align learns the rules of A 3 -Law; and A 3 -Bench evaluates MLLMs and two tasks such as quality selection and prescriptive critique.

Figure 2: A 3 -Dataset construction pipeline. The pipeline has two stages. In the Human Centric Phase we collect diverse advertising images, assign preliminary aesthetic and content tags under A 3 -Law, and remove noisy or inconsistent samples. In the Model Enhanced Phase multimodal LLMs generate Chain of Thought rationales that are refined and validated by experts. The result is the A 3 -Dataset with 30K images and 120K instruction and response pairs of high quality.

Figure 3: A 3 -Align under the A 3 -Law. The top panel shows examples from the A 3 -Dataset organized by the three stages Perceptual Attention , Formal Interest , and Desire Impact , with subcriteria and Chain of Thought summaries. The bottom panel presents a two-phase training pipeline. In the SFT phase the multimodal LLM learns A 3 -Law rules, structured output format, tool use, and Chain of Thought from the A 3 -Dataset with token-level cross-entropy. In the GRPO phase, the model is optimized with multi-signal rewards, ultimately leading to A 3 -Align, which produces rule-based judgments.

Model	Perceptual Attention			Formal Interest			Desire Impact
	Image Fidelity	Integration Realism	Professional Polish	Hue Adaptability	Color Harmonization	Layout Adaptability	Copywriting Tone	Promotional Iconography		Aesthetic Attribute		Advertising Attribute
	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	[email protected] $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$
Open-source Models
Qwen3-VL-8B-Instruct [ 83 ]	0.454	0.491	0.463	0.491	0.444	0.472	0.611	0.157	0.011	0.564	0.562	0.533	0.528
Gemma-3-27B-it [ 82 ]	0.648	0.574	0.722	0.639	0.583	0.694	0.667	0.333	0.001	0.677	0.657	0.660	0.660
Qwen3-VL-32B-Instruct [ 83 ]	0.602	0.583	0.518	0.694	0.574	0.482	0.880	0.259	0.172	0.640	0.617	0.728	0.701

Table 1: Performance across models under A 3 -Law. The table reports results for the three stages Perceptual Attention , Formal Interest , and Desire Impact with their subcriteria. For classification rules we show Acc . For icon grounding we show [email protected] . For continuous subjective prediction on Aesthetic Attribute and Advertising Attribute we show SRCC and PLCC . Models are grouped by type, including open source, open source with thinking, closed source, and closed source with thinking, with A 3 -Align listed in the final row.

Method	Perceptual Attention			Formal Interest			Desire Impact
	Image Fidelity	Integration Realism	Professional Polish	Hue Adaptability	Color Harmonization	Layout Adaptability	Copywriting Tone	Promotional Iconography		Aesthetic Attribute		Advertising Attribute
	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	Acc $\uparrow$	[email protected] $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$	SRCC $\uparrow$	PLCC $\uparrow$
w/o CoT	0.848	0.796	0.880	0.769	0.833	0.806	0.962	0.889	0.678	0.805	0.811	0.792	0.788
w/o Accuracy Reward	0.833	0.788	0.870	0.778	0.824	0.787	0.935	0.833	0.692	0.862	0.859	0.824	0.817
w/o Tool Reward	0.861	0.824	0.898	0.751	0.824	0.833	0.962	0.917	0.681	0.850	0.843	0.827	0.824
w/o IoU Reward	0.861	0.815	0.907	0.806	0.870	0.870	0.981	0.861	0.624	0.870	0.873	0.822	0.825

Table 2: Ablation analysis under A 3 -Law. Accuracy is reported for classification rules, [email protected] for icon grounding, and SRCC/PLCC for continuous subjective prediction.

Figure 4: Satisfaction and Action Intent Gains Through A 3 -Law Screening. The figure shows the cumulative score gains in Satisfaction (left) and Action Intent (right) across the three stages of A 3 -Law screening: Perceptual Attention , Formal Interest , and Desire Impact .

Figure 5: Evaluation of Problem Identification Accuracy, Depth of CoT, and Overall Clarity. The figure presents the scores of various models on three evaluation dimensions: Problem Identification Accuracy , Depth of Chain of Thought (CoT) , and Overall Clarity .

Explore similar work

Sense it with your eyes: Sensation Generation and Understanding for Advertisements
Jul 28, 2026Aysan Aghazadeh, Sina Malakouti, Adriana KovashkaOnline Advertising
University of Pittsburgh
Can Vision Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective
Mar 1, 2026Arctanx An, Shizhao Sun, Danqing Huang +5VLM Evaluation
Peking University · Microsoft Research Asia · Microsoft +1
Evaluating and Pricing Advertisements in AI-Generated Responses
Jul 30, 2026John L. Turner-Smith, Zimeng Huang, Yuhan Fu +2Mechanism Design Online Advertising
Carnegie Mellon University, Pittsburgh, PA, USA · College of AI, Tsinghua University, Beijing, China

Jul 28, 2026J/K move · Enter open · S save

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

Aysan Aghazadeh, Sina Malakouti, Adriana Kovashka

University of Pittsburgh

Online Advertising

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

Can Vision Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective

Evaluating and Pricing Advertisements in AI-Generated Responses