From Unlearning to UNBRANDING: A Benchmark for Trademark‑Safe Text‑to‑Image Generation

Dawid Malarz1,2, Artur Kasymov1,3, Filip Manjak1, Maciej Zięba4,5, Przemysław Spurek1,2
Jagiellonian University; IDEAS Research Institute; Wrocław University of Science and Technology;
Preprint 2025
  1. Jagiellonian University, Faculty of Mathematics and Computer Science
  2. IDEAS Research Institute, Warsaw, Poland
  3. Jagiellonian University, Doctoral School of Exact and Natural Sciences, Krakow, Poland
  4. Wroclaw University of Science and Technology
  5. Tooploox, Wroclaw, Poland
Paper Code 🤗 Dataset
UNBRANDING teaser

Unbranding illustrated with the Coca‑Cola brand across varied settings.

Abstract

The rapid progress of text‑to‑image diffusion models raises significant concerns regarding the unauthorized reproduction of trademarked content. While prior work targets general concepts (e.g., styles, celebrities), it fails to address specific brand identifiers. Crucially, brand recognition is multi‑dimensional, extending beyond explicit logos to encompass distinctive structural features (e.g., a car’s front grille). We introduce unbranding, a novel task for the fine‑grained removal of both trademarks and subtle structural brand features, while preserving semantic coherence. To facilitate research, we construct a comprehensive benchmark dataset and present a VLM‑based QA evaluation probing both explicit and implicit brand signals. Our results confirm unbranding is a distinct, practically relevant problem requiring specialized techniques.

Brand Cues: Logos and Trade Dress

Coca‑Cola 1
Coca‑Cola 2
McDonald’s 1
McDonald’s 2
BMW 1
BMW 2
Nutella 1
Nutella 2

Brand identity is signaled by logos and distinctive shapes (e.g., BMW grille, Coca‑Cola bottle). Effective unbranding must suppress both.

Brand‑Safety Policy Tensions

Coca‑Cola policy
Pepsi policy
Tesla 1
Tesla 2
McDonald’s policy
McDonald’s alt
Apple 1
Apple 2

Text‑to‑image systems readily generate branded scenes in disallowed contexts. This underscores the need for dedicated unbranding techniques.

Fidelity vs. Removal Trade‑off

Trade‑off figure

The baseline preserves visual structure but fails at removal; ESD removes brands but alters semantics. Effective unbranding must achieve both.

BibTeX

@misc{malarz2025unlearningunbrandingbenchmarktrademarksafe,
  title={From Unlearning to UNBRANDING: A Benchmark for Trademark-Safe Text-to-Image Generation},
  author={Dawid Malarz and Artur Kasymov and Filip Manjak and Maciej Zięba and Przemysław Spurek},
  year={2025},
  eprint={2512.13953},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.13953},
}