Rethinking Artistic Copyright Infringements in the Era of Text-to-Image Generative Models

TL;DR

The paper tackles the problem of artistic style copyright infringement in the era of text-to-image generation by reframing style copying as a classification task over image portfolios. It introduces ArtSavant, a practical tool combining a neural detector (DeepMatch) and an interpretable tag-based detector (TagMatch) to identify unique artist signatures from a WikiArt reference set of $372$ artists. Large-scale experiments show that only about $20.2\%$ of artists exhibit detectable style copying in generated images prompted by contemporary models, with DeepMatch achieving high accuracy on real art but lower, more nuanced performance on generated content. The work provides actionable, transparent methods for artists, lawyers, and judges to assess copying risk and understand the stylistic elements involved through concrete tag signatures and example attributions.

Abstract

Recent text-to-image generative models such as Stable Diffusion are extremely adept at mimicking and generating copyrighted content, raising concerns amongst artists that their unique styles may be improperly copied. Understanding how generative models copy "artistic style" is more complex than duplicating a single image, as style is comprised by a set of elements (or signature) that frequently co-occurs across a body of work, where each individual work may vary significantly. In our paper, we first reformulate the problem of "artistic copyright infringement" to a classification problem over image sets, instead of probing image-wise similarities. We then introduce ArtSavant, 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 dataset of works from 372 artists curated from WikiArt, and (ii) recognize if the identified style reappears in generated images. We leverage two complementary methods to perform artistic style classification over image sets, includingTagMatch, which is a novel inherently interpretable and attributable method, making it more suitable for broader use by non-technical stake holders (artists, lawyers, judges, etc). Leveraging ArtSavant, we then perform a large-scale empirical study to provide quantitative insight on the prevalence of artistic style copying across 3 popular text-to-image generative models. Namely, amongst a dataset of prolific artists (including many famous ones), only 20% of them appear to have their styles be at a risk of copying via simple prompting of today's popular text-to-image generative models.

Figures (17)

• Figure 1: We define artistic style as a set of elements (or signature) that appear frequently over a body of work, and reduce the problem of style copy detection to classification of sets of images to artists. ( left) We propose two ways to recognize artistic styles over a set of images, including a novel inherently interpretable and attributable tag-based method. ( right) In an empirical study of $372$ prolific artists, we find generative models potentially copy artistic styles for $20.2\%$ of these artists.
• Figure 2: Example generations from Stable Diffusion 2 when prompted to produce specific paintings by Vincent Van Gogh, along with the histogram of similarities between the generated image and corresponding real image. Even for a famous artist like Vincent Van Gogh, generative models rarely produce near-exact duplicates. However, Van Gogh's style appears consistently, even when similarity is low.
• Figure 3: DeepMatch on held-out real art: $89.3\%$ of artists can be recognized. The remaining $10.7\%$ of artists have very similar styles to other artists: e.g., Palma Il Giovane's work differs marginally from other Italian renaissance painters.
• Figure 4: Example atomic tags assigned via our proposed CLIP-based zero-shot method. We perform selective multilabel classification along various aspects of art (e.g. medium, colors, shapes, etc), so that atomic tags span diverse categories. Details in section \ref{['subsec:tagging']}.
• Figure 5: Composing atomic tags results in more unique tags, towards artistic signatures. We propose an efficient algorithm to count composed tags; see Algorithm \ref{['alg:cap']}.