A$^3$: Towards Advertising Aesthetic Assessment

Kaiyuan Ji, Yixuan Gao, Lu Sun, Yushuo Zheng, Zijian Chen, Jianbo Zhang, Xiangyang Zhu, Yuan Tian, Zicheng Zhang, Guangtao Zhai

TL;DR

The $A^3$ (Advertising Aesthetic Assessment) framework operationalizes cognitive marketing theory into an executable AI evaluation pipeline. By structuring ad evaluation into a three-stage cognitive hierarchy and training a specialized multimodal model ($A^3$-Align) using reinforcement learning, the framework significantly outperforms commercial giants like Gemini 2.5 Pro and GPT-4o in ad critique and quality selection.

What Problem Does This Paper Address?

In an era of digital "marketing overload," an advertisement's aesthetic quality is critical to cutting through the noise and driving brand recall. However, evaluating advertising images has traditionally been a bottleneck, relying on expensive, non-scalable, and highly subjective human scoring.

While Multimodal Large Language Models (MLLMs) show promise for visual tasks, standard models struggle with ad assessment. They typically perform one-step holistic scoring, which completely ignores the progressive nature of human cognition (how we look at an ad vs. how we process its message). Furthermore, standard MLLMs suffer from severe reasoning-to-output misalignment—frequently generating unstable rationales that don't match their final scores—and they struggle to precisely localize visual elements like promotional badges.

Methods

The authors bridge this gap by introducing a theory-driven, hierarchical paradigm called $A^3$-Law, inspired by the classic marketing AIDA (Attention, Interest, Desire, Action) model. It breaks evaluation down into three sequential stages:

  1. Perceptual Attention: Low-level physiological screening (Is the image clear? Is it free of rendering artifacts?).
  2. Formal Interest: Structural composition (Is the color palette harmonious? Is the layout adaptive and balanced?).
  3. Desire Impact: High-level semantic and affective evaluation (Is the copywriting tone appropriate? Does it evoke a desire to buy?).
Figure 3. A3-Align under the A3-Law. The top panel shows examples from the A3-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 A3-Law rules, structured output format, tool use, and Chain of Thought from the A3-Dataset with token-level cross-entropy. In the GRPO phase, the model is optimized with multi-signal rewards, ultimately leading to A3-Align, which produces rule-based judgments.

To build a model that respects this hierarchy, the authors developed $A^3$-Align (built on Qwen3-VL-8B-Instruct) using a two-phase training pipeline:

  • Supervised Fine-Tuning (SFT) Phase: Teaches the model the $A^3$-Law rules, structured output formats, and how to call lightweight analytical tools (like hue analysis and OCR) to ground its reasoning.
  • Group Relative Policy Optimization (GRPO) Phase: A reinforcement learning step utilizing a multi-source reward framework. Crucially, instead of using rigid binary rewards for subjective criteria (like aesthetic pleasure), they implement a Gaussian-shaped continuous reward function to softly reward predictions that closely align with fine-grained human ratings. It also uses IoU-based rewards to sharpen spatial localization.

Datasets

The paper introduces two major resources:

  • $A^3$-Dataset: A large-scale dataset containing 30K diverse advertising images (spanning e-commerce, social media, and web banners) annotated with 120K instruction-response pairs. It includes fine-grained multi-dimensional labels, object detection bounding boxes for promotional icons, and expert-validated Chain-of-Thought (CoT) rationales.
  • $A^3$-Bench: A comprehensive benchmark evaluating over 20 mainstream open- and closed-source models on objective classification, icon localization, and continuous subjective scoring.
Figure 2. A3-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 A3-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 A3-Dataset with 30K images and 120K instruction and response pairs of high quality.

Key Insights and Ideas

  • The Cognitive Hierarchy Stabilizes MLLMs: Forcing the model to evaluate an image sequentially—from low-level perception to high-level desire—drastically reduces reasoning instability and aligns the model's internal logic with its final output.
  • "Thinking" Helps Closed-Source, Hurts Open-Source: The benchmark revealed a fascinating divergence: enabling explicit "thinking" inference modes consistently improved closed-source models (like Gemini 2.5 Pro), but frequently degraded or destabilized open-source models on subjective rating tasks.
  • Localization is the Universal Blindspot: Prior to this work, even the strongest closed-source models could identify if a promotional icon existed, but completely failed at accurately bounding its coordinates (the best baseline mAP was a dismal 0.283).

Key Results

$A^3$-Align achieved state-of-the-art (SOTA) performance across every single evaluation metric in the benchmark:

  • Perceptual & Formal Subtasks: Achieved scores of 0.870 (Fidelity), 0.824 (Realism), and 0.917 (Polish), outperforming the strongest baselines by massive margins.
  • Promotional Icon Localization: Boosted spatial precision to an mAP of 0.701—more than doubling the performance of the best closed-source baseline.
  • Subjective Alignment: Attained a continuous score correlation of 0.880, marking a >0.108 improvement over prior models.
  • Downstream Utility: In real-world ad selection tests, filtering ads through $A^3$-Align yielded a +1.05 lift in human satisfaction and a +1.20 lift in action intent compared to unfiltered baselines.
Model performance comparison on prescriptive advertisement critique.

Limitations and Future Work

Despite its strong performance, the authors note two primary failure modes for $A^3$-Align:

  • Spatial Crowding: In highly dense visual layouts, the model still occasionally suffers from icon hallucinations or merges adjacent bounding boxes.
  • Minimalist Bias: The model tends to over-associate visual simplicity with low advertising effectiveness, causing it to occasionally underestimate premium, minimalist luxury designs.

Additionally, the current framework focuses on universal commercial prerequisites. The authors leave cultural calibration—adjusting the model to account for localized cultural symbols and regional aesthetic preferences—as an open direction for future work.

Summary updated

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.
BibTeX
@inproceedings{ji2026towards,
  author = {Kaiyuan Ji and Yixuan Gao and Lu Sun and Yushuo Zheng and Zijian Chen and Jianbo Zhang and Xiangyang Zhu and Yuan Tian and Zicheng Zhang and Guangtao Zhai},
  title = {{A$^3$: Towards Advertising Aesthetic Assessment}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=qW36ZWcaII},
}
Figure