Benchmark and task for trademark‑safe text‑to‑image generation. We introduce unbranding — fine‑grained removal of both explicit logos and implicit trade‑dress while preserving object and scene semantics — together with a dataset and a VLM‑QA evaluation protocol.

Dawid Malarz, Artur Kasymov, Filip Manjak, Maciej Zięba, Przemysław Spurek

Modern text‑to‑image diffusion models can faithfully reproduce trademarks. Prior unlearning works target general concepts (e.g., styles, celebrities) and miss brand‑specific identifiers. Brand recognition is multi‑dimensional, spanning explicit logos and distinctive structural features (e.g., a car’s front grille). We define unbranding as the fine‑grained removal of both trademarks and subtle trade‑dress while preserving semantic coherence. We introduce a benchmark dataset and a VLM‑based QA metric that probes for both explicit and implicit brand signals, going beyond classic logo detectors. Our analysis shows newer models tend to synthesize brands more readily, underscoring the urgency of unbranding. Results validated by our metric indicate that unbranding is a distinct, practically relevant problem requiring specialized techniques.

• Task: fine‑grained removal of brand identifiers while preserving object fidelity.
• Benchmark: dataset spanning explicit logos and trade‑dress cues.
• Metric: VLM‑based QA probing both explicit and implicit brand signals.
• Motivation: newer models (e.g., SDXL, FLUX) reproduce brands more readily than older ones.

Trade‑off between fidelity and removal. Baseline preserves structure but often fails to remove brands; ESD removes brands but alters semantics. Effective unbranding must achieve both.

• Python 3.13+
• CUDA-compatible GPU (for image generation and VLM inference)
• uv package manager

1. Clone the repository:

git clone https://github.com/gmum/UNBRANDING.git
cd UNBRANDING

2. Install dependencies using uv:

uv sync

This will create a virtual environment in .venv and install all required packages including:

• PyTorch 2.8.0 with CUDA support
• Diffusers 0.35.2 (Stable Diffusion, FLUX models)
• Transformers 4.57.1 (HuggingFace models)
• VLLM 0.11.0 (optimized VLM inference)
• Streamlit 1.51.0 (annotation UI)
• Ray 2.51.1 (distributed processing)

3. Activate the environment:

source .venv/bin/activate

Or run commands directly with uv:

uv run python <script.py>

# Using Stable Diffusion XL
uv run python generate_images.py --model sdxl --prompts configs/vlm_vss.json --output-dir output/sdxl --seed 42

# Using FLUX.1 Schnell (fast variant)
uv run python generate_images.py --model flux-schnell --prompts configs/vlm_bps.json --output-dir output/flux

# Using Stable Diffusion 3.5 Large
uv run python generate_images.py --model sd35 --prompts configs/vlm_vss.json --output-dir output/sd35

Supported Models:

• sd14 - Stable Diffusion v1.4
• sdxl - Stable Diffusion XL
• sd35 - Stable Diffusion 3.5 Large
• flux-schnell - FLUX.1 Schnell
• flux-dev - FLUX.1 Dev
• qwen-image - Qwen Image
• custom - Any HuggingFace model (use with --model-id)

Start VLLM server (in separate terminal):

uv run vllm serve llava-hf/llava-1.5-7b-hf --port 8000

Run VLM evaluation:

uv run python client.py \
  --gt_imgs_dir data/ground_truth \
  --gen_imgs_dir output/sdxl \
  --model-type llava \
  --output results/evaluation.json

Supported VLM Types:

• llava - LLaVA 1.5 7B
• nemotron - NVIDIA Nemotron Nano VL 8B
• gemma3 - Google Gemma-3 4B IT

Please cite our work if you find it useful:

@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},
}

The project page is adapted from the Academic Project Page Template (inspired by Nerfies). Images in this repository are for research/illustration.