Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative Models
Authors: Mazda Moayeri, Sriram Balasubramanian, Samyadeep Basu, Priyatham Kattakinda, Atoosa Chegini, Robert Brauneis, Soheil Feizi
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we build on prior legal scholarship to develop an automatic and interpretable framework to quantitatively assess style infringement. Our methods hinge on a simple logical argument: if an artist s works can consistently be recognized as their own, then they have a unique style. Based on this argument, we introduce Art Savant, a practical (i.e., efficient and easy to understand) tool to (i) determine the unique style of an artist by comparing it to a reference corpus of works from hundreds of artists, and (ii) recognize if the identified style reappears in generated images. We then apply Art Savant in an empirical study to quantify the prevalence of artistic style copying across 3 popular text-to-image generative models, finding that under simple prompting, 20% of 372 prolific artists studied appear to have their styles be at risk of copying by today s generative models. |
| Researcher Affiliation | Academia | 1University of Maryland, 2 George Washington University |
| Pseudocode | Yes | Algorithm 1 Iterative Algorithm to Obtain Tag Composition Per Artist a A |
| Open Source Code | No | We note that all code will be released upon acceptance. |
| Open Datasets | No | To this end, we curate a dataset D consisting of artworks from Wiki Art 3 (like others (Tan et al., 2017; Karayev et al., 2014)) to serve as (i) a reference set of artistic styles, (ii) a validation set of real art to show (most) artists have unique styles and our methods can recognize them on held-out sets of their works, and (iii) a test-bed to explore if text-to-image models replicate the styles of the artists in our dataset in their generated images. We include 91k artworks from 372 artists A spanning diverse eras and art movements, including any artist with at least 100 works on Wiki Art. Each work is labeled with its genre (e.g., landscape) and style (e.g., Impressionism), though we primarily use the artist and title labels. We provide an easy-to-execute script to enable others to scrape newer versions of this dataset if desired. 3https://www.wikiart.org/; note that we only include Public domain or fair use images. |
| Dataset Splits | Yes | For our classifier, we train a two layer MLP on top of embeddings from a frozen CLIP Vi T-B\16 vision encoder (Radford et al., 2021), using a train split containing 80% of our dataset. |
| Hardware Specification | Yes | Since we utilize frozen embeddings, training takes only a few minutes on one RTX2080 GPU. |
| Software Dependencies | No | The paper mentions models and tools such as 'CLIP Vi T-B\16 vision encoder', 'Vicuna-13b', 'Chat GPT', and 'DINOv2', but does not specify version numbers for any programming languages, libraries, or other software dependencies used in their implementation. |
| Experiment Setup | Yes | For our classifier, we train a two layer MLP on top of embeddings from a frozen CLIP Vi T-B\16 vision encoder (Radford et al., 2021), using a train split containing 80% of our dataset. We employ weighted sampling to account for class imbalance. |