AIGIEmo

Abstract

AI-generated images (AIGI) have emerged as a dominant form of visual media; however, the emotional responses they elicit in human observers remain largely underexplored. Existing affective image datasets mainly focus on natural images and do not provide sufficient support for studying emotion in AIGI. To address this gap, we present AIGIEmo, an emotion-annotated dataset for AI-generated images. The dataset contains 160,000 images, balanced across eight emotion categories, and includes both categorical emotion labels and continuous valence--arousal annotations. In addition, AIGIEmo provides rich multimodal metadata, including image captions, reasons for emotion, visual style, scene type, object information, facial expressions, human actions, and low-level visual attributes such as color and brightness. To improve annotation quality, we use human annotation for emotion labels and machine-assisted pre-annotation for non-emotion fields. We further analyze the emotional distribution, affective consistency, and visual-textual properties of the dataset, showing that AIGIEmo offers class-balanced affective structure and diverse multimodal annotations. AIGIEmo provides a useful resource for studying emotional responses to AIGI and supports future research on affective understanding, analysis, and reasoning of AI-generated images.

Download

Open on Hugging Face

• License: The AIGIEmo dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
• Access: AIGIEmo is publicly available on Hugging Face and can be downloaded directly without any access request or approval.

Data Preview

Supplementary Materials

Open Supplementary PDF

‹ ›

Annotation Tool

Annotation interface for efficient multi-field labeling and human verification.

Data Analysis Results

Left: style-family distribution. Middle: attribute-emotion association. Right: semantic cue proportions in captions and emotion reasoning.

Middle-panel metrics

Continuous: ε² = (H - k + 1) / (n - k)

Categorical: V = sqrt(χ² / (n · min(r - 1, c - 1)))

High-level semantic cues are stronger than low-level visual attributes.

Baseline Visual Model

Backbone and shared neck with emotion and V-A branches, plus gated auxiliary attribute cues.

V-A Space Scatter

Emotion categories form a clear structure in valence-arousal space, with positive and negative groups well separated.

V-A Marginal Violins

Most overlap appears between nearby affective categories rather than across opposite regions.

Style to Emotion

Different style families show distinct emotion preferences, while several major styles remain broadly balanced.

Emotion to Style

Each emotion category covers multiple styles, but their proportions vary in interpretable ways.

Brightness and Colorfulness

Positive emotions are slightly brighter and more colorful, but overlap remains substantial.

Effect-size metrics

Continuous: ε² = (H - k + 1) / (n - k)

Categorical: V = sqrt(χ² / (n · min(r - 1, c - 1)))

Hue Composition

Warm hues dominate across categories, while cool and warm shares vary by emotion.

Facial Expression Enrichment

Facial expressions show the clearest emotion-specific enrichment patterns.

Semantic enrichment

Lift(a, e) = P(a | e) / P(a)

Lift above 1 indicates that an attribute is over-represented in that emotion.

Human Action Enrichment

Action categories provide useful behavior-level cues for affective understanding.

Object Enrichment

Objects provide diverse semantic grounding, with different categories enriched for different emotions.

Scene Enrichment

Scene cues are broader than human-centered attributes, but still show coherent emotion-related trends.

Baseline Benchmark and Evaluation

We summarize benchmark results for both visual baselines and representative VLMs on AIGIEmo, keeping the tables compact while preserving the full evaluation metrics.

Additional Evidence of AIGIEmo Value

We further evaluate whether AIGIEmo provides useful supervision beyond existing VEA data, both under controlled data mixing and cross-dataset transfer.

Data Mixing on AIGIEmo

Starting from EmoSet as the base training set, adding more AIGIEmo data steadily improves ResNet50, ConvNeXt-T, and Qwen3-VL on the AIGIEmo test split.

Transfer of Data Mixing to AIGI-VC

The same mixing strategy also improves performance on AIGI-VC, showing that AIGIEmo contributes useful supervision beyond its own in-domain benchmark.

Together, these results show that AIGIEmo is not only a strong in-domain dataset, but also a practically useful source of transferable supervision for emotion recognition on AI-generated images.

Potential Applications

Contact

Zhanpeng Jin
zjin@scut.edu.cn

Yang Gao
gaoyang2025@scut.edu.cn

Dong She
ftdshe@mail.scut.edu.cn

Cong Liu
202420163862@mail.scut.edu.cn

Nuo Chen
202230310052@mail.scut.edu.cn