Research map / Image generation / Text-to-image diffusion
Stylebreeder: Exploring and Democratizing Artistic Styles through Text-to-Image Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces STYLEBREEDER, a CC0 dataset of 6.8M user-generated images and 1.8M prompts from 95K Artbreeder users, and benchmarks style clustering, personalization methods, and style-based recommendation.
Problem
Existing AI-generated art datasets are small, short in time span, lack original prompts, or are limited in stylistic diversity, leaving user-generated artistic styles on text-to-image platforms undocumented.
Approach
Collects a large, long-duration dataset of user-generated images with original prompts and metadata from Artbreeder, clusters images by style embeddings (CSD) to discover styles, benchmarks personalization methods per style cluster, and releases style LoRAs via a Style Atlas platform.
Key results
- EDLoRA achieves the highest CLIP-I (0.6957) and DINO (0.4125) averages among the four personalization methods on the style benchmark.
- Clustering 6.8M images with CSD embeddings into 10,000 clusters maximizes silhouette score (0.110) among tested k values.
- Style-based SVD recommender reaches mean RMSE 0.1425 and MAE 0.082 across 5-fold cross-validation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EDLoRA instead of Textual Inversion for style personalization | Style personalization · 40 clusters, average | CLIP-I: 0.6869 -> 0.6957 | Helped |
| Use EDLoRA instead of Textual Inversion for style personalization | Style personalization · 40 clusters, average | DINO: 0.3801 -> 0.4125 | Helped |
| Use LoRA w/DreamBooth instead of EDLoRA for style personalization | Style personalization · 40 clusters, average | CLIP-T: 0.1822 -> 0.1896 | Helped |
| Increase the cluster count from 10000 to 20000 (ablation) | K-Means++ style clustering · Style clustering | Silhouette Score: 0.110 -> 0.111 | No clear effect |
Surprising findings
- Clusters are often dominated by a single artist or small groups of artists, and a substantial share of the dataset contains potentially NSFW content.
Limitations
- Potential over-reliance on AI in art and copyright or originality concerns when styles mimic existing artists.
- Dataset contains substantial NSFW and toxic content and risk of bias reinforcement; full 6.8M version with unsafe prompts is gated.
- Personalization benchmark uses a limited sample (40 clusters, 10 prompts) and the recommender uses a subset of 96K images from 1434 users.
Experimental setup
Datasets: STYLEBREEDER
Benchmarks: STYLEBREEDER
Metrics: CLIP-I CLIP-T DINO silhouette score RMSE MAE
Baselines: Textual Inversion LoRA w/DreamBooth Custom Diffusion
Builds on: DreamBooth Textual Inversion Custom Diffusion EDLoRA CSD
Closest papers by meaning
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.87)
- Social Reward: Evaluating and Enhancing Generative AI through Million-User Feedback from an Online Creative Community (ICLR 2024 · similarity 0.87)
- Gender Artifacts from Art History to Text-to-Image Generation (NeurIPS 2026 · similarity 0.87)
- Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative Models (ICLR 2025 · similarity 0.87)
- ArtBank: Artistic Style Transfer with Pre-trained Diffusion Model and Implicit Style Prompt Bank (AAAI 2024 · similarity 0.87)
- FiVA: Fine-grained Visual Attribute Dataset for Text-to-Image Diffusion Models (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.