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.

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

Construction Notes

• The subset was sampled from metadata.csv in the parent EmoVisDB-1M workspace.
• Sampling was class-balanced across the eight categorical emotions.
• Duplicate source images were removed so the final export contains 160,000 unique image files.
• The export keeps the original metadata fields and adds sample_id plus split.

Local Regeneration

The export can be regenerated locally with:

python create_aigiemo_160k.py
python copy_aigiemo_160k_images.py
python export_aigiemo_160k_hf.py

Intended Use

This dataset is intended for research on:

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

Please review the source dataset and your project requirements before public release or redistribution.