Approaches to Artistic Style Suppression: An Evaluation Framework for Copyright Compliance in Generative Artificial Intelligence

Text-to-image diffusion models raise copyright concerns due to their ability to reproduce distinctive artistic styles learned from training data. Parameter-based machine unlearning methods address these concerns by modifying model weights to reduce learned associations, but their computational requirements make them impractical in resource-constrained environments. This thesis investigates inference-time negative prompting as an accessible alternative for style suppression, evaluating its effectiveness and limitations. The research develops a method-agnostic evaluation framework that employs paired-comparison methodology, multidimensional assessment, and geometric validation in embedding space. The framework is demonstrated through a systematic evaluation of inference-time negative prompting across five artists representing diverse traditions: James Jean and Esao Andrews (contemporary illustration), Claude Monet (Impressionism), Pablo Picasso (Cubism), and Leonardo da Vinci (Renaissance). Artistic style similarity is quantified with contrastive style descriptors, a pretrained metric validated through portfolio discrimination analysis before its application to intervention assessment. The evaluation establishes four findings. First, inference-time negative prompting achieves statistically significant similarity reduction (13.8% mean decrease, p < 0.001), repositioning generated images from same-artist similarity ranges towards different-artist similarity ranges. Second, effectiveness varies substantially across artists (8.1% to 19.1%) and prompts (1.28% to 29.96%), indicating context-dependent performance. Third, style suppression and content preservation prove statistically independent (r = -0.10, p = 0.44), demonstrating that the reduction of stylistic similarity does not systematically compromise semantic content fidelity. Fourth, computational style discrimination metrics show only weak alignment with human authenticity judgements (r = -0.14, p = 0.28), reflecting differences in construct between technical style discrimination and human aesthetic assessment. The research contributes a reusable evaluation framework for diverse interventions. It also offers an empirical characterisation of accessible methods given resource constraints. Finally, it shows that thorough copyright assessment requires both computational screening and human expert evaluation, rather than relying only on automated approaches.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Cumming, Aedan

Advisors dc:contributor.advisor

• Burger, L. E.
• Jansen, G.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135706

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135706