UNBRANDING v1

This is the official dataset for the paper From Unlearning to UNBRANDING: A Benchmark for Trademark-Safe Text-to-Image Generation.

Project Page | GitHub | Paper

This dataset contains prompt-based image generation inputs for multiple well-known brands, organized into two prompt styles (“prompt sets”). Each record pairs a brand label with a text prompt and a random seed to support reproducible image generation.

Dataset Summary

• Version: v1
• Rows: 1,759
• Columns: 5
• Number of brands: 12
• Prompt sets: 2 (biased, directed)

Brands

The dataset includes the following brands:

• adidas
• apple
• audi
• bmw
• coca-cola
• emirates
• mcdonalds
• mercedes
• monster
• nike
• puma
• singapore_airlines

Prompt Sets

Each example belongs to one of two prompt sets:

• biased: prompts that describe brand-associated aesthetics or attributes without explicitly naming the brand.
• directed: prompts that explicitly reference the brand (e.g., asking for a photo of a brand-specific object).

Data Schema

Each row corresponds to a single generation example with the following fields:

• brand (string)
  The target brand label associated with the example.

• seed (int)
  The random seed used for generation to support reproducibility.

• prompt (string)
  The text prompt to be used with an image generation model.

• prompt_set (string)
  The prompt style category: biased or directed.

• filename (string)
  A unique identifier for the example, formatted as brand_seed_id (e.g., bmw_1531_123).

Example

{
  "brand": "bmw",
  "seed": 1531,
  "prompt": "…",
  "prompt_set": "biased",
  "filename": "bmw_1531_123"
}

Intended Use

This dataset is suitable for experiments involving brand representation in generative prompts, prompt style comparisons, and reproducible image generation pipelines.

Citation

@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},
}

Notes

• filename can be used as a stable example identifier.
• seed is provided to ensure the same prompt can be regenerated consistently with compatible generation settings.