rule34nexus/image-aesthetic-scores

Rule34.nexus  ·  Licence: Rule34.nexus Derived Dataset Licence 1.0

Rule34.nexus Image Aesthetic Scores

1. Overview

This dataset contains per-image aesthetic predictions for images in the Rule34.nexus corpus. Predictions were generated using discus0434/aesthetic-predictor-v2-5. Source images are not included in this dataset — only opaque post identifiers, the source image's SHA-256 hash, the post's content type, and the predicted score.

2. Dataset Description

Each row is one Rule34.nexus post that has been scored by the aesthetic predictor. See Inclusion Criteria below for what's excluded; beyond that, the full score distribution is published unfiltered — filter client-side, e.g. dataset.filter(lambda x: x["aesthetic_score"] >= 5.5). doujin posts (see Glossary) often contain embedded dialogue/caption text that can lower their score for reasons unrelated to image quality — consider dataset.filter(lambda x: x["media_type"] != "doujin") if that matters for your use case.

• Rows: 2,612,431
• Dataset version: 2026.10.03
• Source corpus version: r34nexus-2026.10.03
• Schema version: 1.1

3. Data Fields

Field	Type	Description
post_uuid	string	Unique Rule34.nexus post identifier
short_id	string	Short, unique, public post identifier used in post URLs; see Glossary. Nullable.
sha256	string	SHA-256 hash of the source image
aesthetic_score	float32	Predicted aesthetic score
media_type	string	"image", "video", or "doujin" — see Glossary
aesthetic_model	string	Model identifier
aesthetic_model_version	string	Version of the model used
dataset_version	string	Version of the Rule34.nexus dataset

4. Glossary

Terms and values that aren't self-explanatory from the field name alone:

• short_id — Rule34.nexus's short (9-character), public post identifier, used in canonical post URLs, e.g. https://rule34.nexus/post/{short_id}.
• media_type — One of "image", "video", or "doujin". doujin posts are auto-synced from Rule34.nexus's doujin catalogue rather than user-uploaded, and often contain embedded dialogue/caption text baked into the image, which can reduce their aesthetic score independent of visual quality.

5. Data Generation

Rule34.nexus corpus
        ↓
image retrieval
        ↓
aesthetic predictor
        ↓
score validation
        ↓
Parquet generation
        ↓
Hugging Face dataset

The exact implementation is not disclosed; see the schema and methodology below for what is published.

6. Inclusion Criteria

Every row-level exclusion applied while building this dataset:

• Only active, non-deleted posts are considered; draft/pending/flagged/deleted posts are excluded.
• Of 2,640,953 active, non-deleted posts exported, 28,522 had no aesthetic score yet (excluded rather than backfilled or published with a null score); 2,612,431 rows remain in this dataset.
• No minimum-quality threshold is applied on top of the above — this dataset is the complete score distribution of scored, visible posts.

7. Methodology

• Aesthetic model: discus0434/aesthetic-predictor-v2-5, version 2024.12.18.1.
• Score range is approximately 1-10 (SigLIP-based predictor); 5.5+ is considered a strong aesthetic score.
• Observed range in this release: 0.600 - 7.812, mean 5.004, median 5.062.

8. Data Distribution

Distribution of aesthetic_score across all 2,612,431 published posts.

9. Limitations

• Aesthetic scores are model predictions, not human ratings.
• The predictor may exhibit domain-specific biases (e.g. style, subject matter, color palette) not representative of aesthetic quality in general.
• Scores from different aesthetic_model_version values should not be directly compared; always check both aesthetic_model and aesthetic_model_version.
• A high or low aesthetic score is not a definitive measure of artistic quality.

10. Intended Uses

• Training and evaluating aesthetic-quality models.
• Filtering image-generation datasets by predicted quality.
• Selecting high-quality training images.
• Analysing aesthetic score distributions.
• Benchmarking image-quality models.
• Research into aesthetic prediction.

11. Licensing

This dataset is released under the Rule34.nexus Derived Dataset Licence 1.0: https://rule34.nexus/licenses/derived-dataset-1.0

Copyright in the original images, artwork, characters, and other third-party material from which this dataset was derived is not claimed by Rule34.nexus and is not licensed under this agreement. This licence covers only the original statistical/metadata contributions in this dataset.

12. Provenance

Source: Rule34.nexus
Dataset: rule34nexus/image-aesthetic-scores
Source corpus: r34nexus-2026.10.03
Generated: 2026-10-03
Images included: No

https://rule34.nexus

This dataset was generated from the Rule34.nexus image corpus and is published by Rule34.nexus for research and machine-learning purposes. The underlying source images are not included in this dataset.