EmoScene

EmoScene is a large-scale visual emotion dataset that fuses three sources of information on the same images: what an image depicts, how it looks, and how it feels. It contains approximately 1.2 million images across more than 300 scene categories, each paired with:

• Affective space: a categorical emotion label (eight emotions — amusement, anger, awe, contentment, disgust, excitement, fear, sadness — plus Neutral) and continuous valence–arousal–dominance (VAD) ratings on a 1–9 scale.

• Perceptual space: measurable appearance attributes computed from pixels, including color proportions, mean HSV brightness and saturation, hue warmth, texture entropy, edge density, and curvilinearity.

• Contextual semantics: scene labels and descriptions of the people, objects, and interactions in the image.

• 📄 Paper: EmoScene: Fusing Affective, Perceptual, and Contextual Information to Measure and Control Visual Emotion (under review). Preprint of an earlier version: arXiv:2604.00933

• 💻 Code: github.com/0having1/AffectCtrl

Release status. This repository is a placeholder for the dataset release. The annotations will be published here upon acceptance of the paper. The schema below describes the planned release format and may be refined before publication.

Dataset structure

The dataset is organized by scene. Each image identifier is paired with one JSON annotation record.

EmoScene/
  beach/
    <stem>.json   # annotation record
  forest/
    <stem>.json
  ...

An abridged example record:

{
  "scene": "beach",
  "emotion": "awe",
  "description_InternVL3-8B": "The image captures a serene sunset over the ocean, ...",
  "color_proportion": {"Black": 0.06, "White": 0.03, "Orange": 0.12, "Pink": 0.17, "Brown": 0.18, "Gray": 0.44, "...": "..."},
  "average_color": {"hue": 20, "saturation": 45, "value": 69},
  "people_count": 0,
  "curvilinearity": 0.0822,
  "complexity_entropy": 0.9402,
  "complexity_edge_density": 0.0183,
  "aesthetic_score": 6.5981,
  "liqe_score": 3.4802,
  "valence_internvl3_8B": 7, "arousal_internvl3_8B": 6, "dominance_internvl3_8B": 5, "emotion_internvl3_8B": "awe",
  "valence_qwen2_5_vl_7B": 7, "arousal_qwen2_5_vl_7B": 5, "dominance_qwen2_5_vl_7B": 6, "emotion_qwen2_5_vl_7B": "awe",
  "valence_llama3_2_11B": 8, "arousal_llama3_2_11B": 6, "dominance_llama3_2_11B": 5, "primary_emotion_llama3_2_11B": "awe"
}

Each record also carries the source platform and the original URL or identifier where available. EmoScene does not prescribe a single official train/test split; the task-specific splits used in the paper are described in its appendix.

Images and licensing

• Annotations in this repository are released under CC BY-NC 4.0 (attribution, non-commercial use).
• Images are not covered by this license. The images come from Unsplash, Pexels, Pixabay, Flickr, and The Metropolitan Museum of Art Open Access collection. Copyright remains with the original platforms and creators, and use of each image is governed by the terms of its source. Image files will be distributed only for subsets whose source licenses permit redistribution; other images can be retrieved from their original sources using the released identifiers.

Annotation process

1. Curation. Images were collected by keyword retrieval across scene categories, filtered for aesthetic quality and sharpness, and checked for scene consistency with Places365 predictions and CLIP image–text similarity.
2. Affective annotation. Three multimodal language models — InternVL3-8B, Qwen2.5-VL-7B-Instruct, and Llama 3.2 11B Vision — independently assess the categorical emotion and VAD of each image using a fixed scene-and-subject template. Model disagreements and apparent intensity mismatches are routed to human review.
3. Perceptual measurement. Appearance attributes are computed deterministically from pixels, separately from the affective labels.
4. Human verification. Reviewers accept or reject each machine-suggested field, and rejected fields are corrected and re-verified.

Per-model outputs are included in the records so that users can apply their own aggregation. Model outputs remain subject to the licenses of the respective models. Built with Llama.

Annotation reliability

Study	Images	Categorical	VAD (V/A/D)
Random audit	30,519	91.22% acceptance	MSE 0.070 / 0.050 / 0.033 (1–9 scale)
Challenge audit (scenes that contradict appearance heuristics)	297	85.00% acceptance	MSE 0.181 / 0.102 / 0.084
Three-rater verification	3,000	Fleiss' κ = 0.85	Mean pairwise inter-rater r = 0.960 / 0.984 / 0.901

Reviewers were shown the machine suggestions during verification. See the paper for the full protocol.

Intended use

EmoScene is intended for research on visual emotion analysis, the relationship between image appearance and affect, and affect-controllable image generation.

Out-of-scope uses. Affect labels describe interpretations of imagery. They are not identity labels or reliable inferences about the internal state of any depicted person. The dataset must not be used for re-identification, surveillance, inferring sensitive personal attributes, or generating manipulative or deceptive content, including targeted persuasion or harassment.

Limitations and bias

Affective judgments are subjective and may reflect the biases of the annotation models, the reviewers, and the source platforms. Emotion categories are not uniformly represented across the collection. The associations between appearance and affect reported in the paper are statistical tendencies, not deterministic rules.

Privacy and opt-out

The records contain only a coarse people count and non-identifying contextual attributes; no identity labels, biometric identifiers, or face recognition annotations are provided. Exact and near-duplicate images are removed before release. If you are the creator of an image or appear in one and would like it removed, please open an issue in the code repository with the image URL or identifier, and it will be removed from subsequent releases.

Citation

If you use EmoScene, please cite:

@article{he2026emoscene,
  title   = {EmoScene: A Dual-Space Dataset for Controllable Affective Image Generation},
  author  = {He, Li and Zhang, Longtai and Zhang, Wenqiang and Wang, Yan and Qi, Lizhe},
  journal = {arXiv preprint arXiv:2604.00933},
  year    = {2026}
}

This entry cites the arXiv preprint and will be updated to the journal version upon publication.