Research map / Image generation / Text-to-image diffusion
Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative 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.
Proposes ArtSavant, a set-level artist-style classifier (DeepMatch neural signatures and interpretable TagMatch tag signatures) over a 372-artist WikiArt reference corpus, and uses it to audit style copying in three text-to-image models.
Problem
Whether text-to-image models copy artists' styles (rather than individual artworks) and how to detect and argue style infringement in an interpretable, legally relevant way.
Approach
Reformulate artistic style infringement as set-level classification of artwork portfolios to artists, with a black-box CLIP-MLP detector (DeepMatch) and a novel zero-shot CLIP tagging plus tag-composition matcher (TagMatch) that yields interpretable, attributable evidence.
Key results
- DeepMatch recognizes 89.3% of 372 artists on held-out real art, with 72.8% per-image accuracy.
- TagMatch achieves 61.6% top-1, 82.5% top-5 and 88.4% top-10 accuracy on held-out real art.
- DeepMatch average match rate on generated images is 20.2%, so about 20% of artists appear at risk of style copying across three models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeepMatch with majority-vote aggregation on held-out WikiArt portfolios | DeepMatch · WikiArt held-out test split | artist recognition rate: 89.3% | Helped |
| Use interpretable tag signatures for artist matching instead of DeepMatch neural signatures | TagMatch versus DeepMatch · WikiArt held-out test split | top-1 accuracy: 89.3% -> 61.6% | Hurt |
| Prompt text-to-image models with artwork titles and artist names, then classify generated portfolios | DeepMatch · Generated art prompted with artist names, averaged across 3 models | artist match rate: 20.2% | Mixed |
| Use TagMatch to identify artists in generated art | TagMatch · Generated art prompted with artist names | top-1 match rate: 12.95% | Hurt |
Surprising findings
- Even for famous artists, generated images rarely near-duplicate originals (most DINO similarities below 0.75), yet style is still recognized.
- Only three held-out artists were flagged under the stricter unprecedented-similarity criterion, and in those cases the held-out artist was not flagged as strongly.
Limitations
- Style is qualitative; the work proxies the space of artists with a 372-artist reference set that may miss styles.
- CLIP-based atomic tagging is imperfect with only 17% majority-disagreement rate and low inter-annotator agreement (51% all agree).
- Only off-the-shelf models were studied; fine-tuned style-mimicry scenarios were not analyzed.
Experimental setup
Datasets: WikiArt
Benchmarks: WikiArt held-out test split
Metrics: top-1 accuracy top-5 accuracy top-10 accuracy match rate match confidence
Baselines: DeepMatch TagMatch
Builds on: CLIP Vicuna WikiArt
Closest papers by meaning
- Stylebreeder: Exploring and Democratizing Artistic Styles through Text-to-Image Models (NeurIPS 2024 · similarity 0.87)
- Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI (ICLR 2025 · similarity 0.86)
- Disguised Copyright Infringement of Latent Diffusion Models (ICML 2024 · similarity 0.86)
- Gender Artifacts from Art History to Text-to-Image Generation (NeurIPS 2026 · similarity 0.84)
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.83)
- ArtBank: Artistic Style Transfer with Pre-trained Diffusion Model and Implicit Style Prompt Bank (AAAI 2024 · similarity 0.83)
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.