dongSHE/AIGIEmo

---
pretty_name: AIGIEmo-160K
license: cc-by-4.0
task_categories:
- image-classification
language:
- en
tags:
- emotion
- image-emotion
- aigc
- synthetic-images
- classification
size_categories:
- 100K<n<1M
dataset_info:
features:
- name: image
dtype: image
- name: sample_id
dtype: string
- name: split
dtype: string
- name: img_name
dtype: string
- name: Caption
dtype: string
- name: ANP
dtype: string
- name: Emotion_Categorical
dtype: string
- name: Style
dtype: string
- name: Scene_Type
dtype: string
- name: Object_Types
dtype: string
- name: Facial_Expression
dtype: string
- name: Human_Action
dtype: string
- name: Reason_for_Emotion
dtype: string
- name: Average_Color
dtype: string
- name: Primary_Color
dtype: string
- name: Secondary_Color
dtype: string
- name: file_name
dtype: string
- name: Valence
dtype: float32
- name: Arousal
dtype: float32
- name: Brightness
dtype: float32
- name: Contrast
dtype: float32
- name: colorfulness
dtype: float32
splits:
- name: train
num_examples: 128000
- name: validation
num_examples: 16000
- name: test
num_examples: 16000
---

# AIGIEmo

AIGIEmo is a large-scale affective image dataset built for understanding emotions in AI-generated images. It provides rich annotations including categorical emotions, valence-arousal scores, captions, visual attributes, scene and object information, and other emotion-related metadata. The dataset is designed to support research on affective computing, vision-language models, emotion understanding, and AIGC image analysis.
The project page is available at https://aigiemo.github.io/.

## Dataset Summary

- Total samples: 160,000
- Total unique images: 160,000
- Splits:
- Train: 128,000
- Validation: 16,000
- Test: 16,000
- Emotion labels:
- Amusement
- Anger
- Awe
- Contentment
- Disgust
- Excitement
- Fear
- Sad

Each emotion class contains 20,000 images in total:

- Train: 16,000 per class
- Validation: 2,000 per class
- Test: 2,000 per class

## Files

- `train.csv`, `val.csv`, `test.csv`: split-specific metadata files
- `metadata.csv`: merged metadata with an added `split` column
- `imgs/`: flattened image directory used by the CSV files
- `hf_parquet/`: Hugging Face-ready Parquet shards with embedded image bytes

## Main Columns

- `sample_id`: unique sample identifier derived from the exported image filename
- `split`: one of `train`, `val`, or `test`
- `img_name`: original image identifier from the source dataset
- `Caption`: image caption / prompt-style description
- `ANP`: adjective-noun pair summary
- `Emotion_Categorical`: emotion label
- `Valence`, `Arousal`: continuous affective attributes
- `Style`, `Scene_Type`, `Object_Types`: semantic and scene descriptors
- `Facial_Expression`, `Human_Action`: optional human-centric attributes
- `Reason_for_Emotion`: free-text explanation for the assigned label
- `Average_Color`, `Primary_Color`, `Secondary_Color`: color descriptors
- `Brightness`, `Contrast`, `colorfulness`: low-level visual attributes
- `file_name`: relative path to the image inside this export
- `image`: only in the Parquet export; HF-compatible image column with embedded bytes

## Intended Use

This dataset is intended for research on:

- image emotion classification
- multimodal affective computing
- evaluation of AIGC-generated emotional content