💻 computer science

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

This paper introduces "unbranding," a novel task and benchmark for fine-grained text-to-image generation that removes both explicit trademarks and subtle structural brand features (trade dress) while preserving semantic coherence, addressing the limitations of existing detectors and highlighting the increased risk posed by higher-fidelity models.

Original authors: Dawid Malarz, Filip Manjak, Maciej Zięba, Przemysław Spurek, Artur Kasymov

Published 2026-03-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dawid Malarz, Filip Manjak, Maciej Zięba, Przemysław Spurek, Artur Kasymov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a magical paintbrush (an AI) that can draw anything you describe. If you say, "Draw a red soda bottle," it might draw a generic bottle. But if you say, "Draw a Coca-Cola bottle," the AI doesn't just draw a bottle; it draws the exact shape, the specific red label, and the famous script logo.

The problem? That's like the AI stealing the brand's identity. If a company doesn't want its product shown in a bad movie or a weird context, the AI might accidentally do it anyway.

This paper introduces a new challenge called "Unbranding." Here is the simple breakdown:

1. The Problem: The "Eraser" vs. The "Scalpel"

Previously, researchers tried to teach AI to "unlearn" things. Imagine you have a giant eraser. If you want to remove "Coca-Cola" from the AI's brain, you use the eraser.

The Old Way (Unlearning): You tell the AI, "Forget everything about Coca-Cola." The AI complies, but it gets confused. It might stop drawing any soda bottles, or it might draw a weird, melted blob that looks nothing like a bottle. It's like trying to remove a stain from a shirt by burning a hole in the fabric.
The New Goal (Unbranding): We want a scalpel. We want to surgically remove only the brand (the logo, the specific red color, the unique bottle shape) while keeping the rest of the object (the glass, the liquid, the bottle shape) perfectly intact.

2. The Challenge: It's Not Just About Logos

The paper points out that brands are tricky. They aren't just a logo on a shirt.

The "Coca-Cola" Bottle: Even if you paint over the logo, if the bottle has that specific curvy shape, everyone still knows it's a Coke.
The "BMW" Car: Even without the round blue-and-white logo on the hood, if the car has that specific "kidney-shaped" grille, people know it's a BMW.

The Analogy: Imagine a celebrity wearing a disguise.

Old Method: You tell the AI to "forget the celebrity." The AI draws a random stranger who looks nothing like the celebrity, but also nothing like a human.
Unbranding: You want the AI to draw the same person, but wearing a hat and sunglasses so no one recognizes them, while keeping their face and body exactly the same.

3. The New Tool: The "Detective AI"

How do you know if you successfully removed the brand? You can't just ask a human to look at thousands of images. The authors built a special Evaluation System using a "Detective AI" (a Vision Language Model).

The Detective's Job: This AI looks at the generated image and asks, "Is this a BMW?"
The Twist: The detective is smart. It doesn't just look for the logo. It looks at the whole picture. "Hmm, I don't see the logo, but that grille shape and those headlights scream 'BMW' to me."
The Score: If the detective says "Yes, this is a BMW," the unbranding failed. If it says "No, this is just a generic car," the unbranding succeeded.

4. The Big Discovery: The AI is Getting Too Good

The paper found something scary but important: Newer, smarter AI models are actually better at drawing brands than older ones.

Old AI: Tried to draw a Nike shoe but got the swoosh logo wrong or made the shoe look weird.
New AI: Draws a perfect Nike shoe with a perfect logo.
Why this matters: Because the new AI is so good at copying brands, the "Unbranding" task is now much harder. We need better tools to stop the AI from accidentally (or intentionally) infringing on trademarks.

5. The Result: A New Benchmark

The authors created a "Test Drive" (a dataset) with 1,700 prompts for 12 famous brands (like Apple, McDonald's, Adidas). They tested all the current "Unlearning" methods on this test drive.

The Verdict:

Current methods are failing. They are either too lazy (leaving the logo there) or too destructive (destroying the object entirely).
The Gap: There is no perfect tool yet that can remove the brand identity without ruining the picture.

Summary

Think of this paper as a warning label and a new rulebook for the future of AI art. It says: "Hey, AI is getting really good at copying famous brands. We need a new way to fix the AI so it can draw a 'red soda bottle' without accidentally drawing a 'Coca-Cola,' and we need a smart way to check if we did it right."

It's about teaching AI to respect intellectual property without losing its artistic touch.

1. Problem Definition

The rapid advancement of text-to-image diffusion models (e.g., SDXL, FLUX) has led to a significant increase in the unauthorized reproduction of trademarked content. While previous research has addressed "concept unlearning" (erasing entire categories like "celebrities" or "styles"), these methods are ill-suited for brand protection because:

Coarse Granularity: Existing unlearning methods often erase the entire object (e.g., removing the concept of a "car" entirely) rather than just the brand identifiers.
Multidimensional Brand Identity: Brands are not defined solely by explicit logos. They rely heavily on trade dress—distinctive structural features, shapes, and design languages (e.g., the Coca-Cola bottle contour, BMW's kidney grille, or Adidas' three stripes).
The Gap: Current methods fail to disentangle brand-specific visual cues from the underlying object semantics. Simply deleting a logo is insufficient if the object retains the brand's distinctive shape or design language.

The authors define Unbranding as a novel task: the fine-grained removal of both explicit trademarks (logos, text) and implicit structural trade dress features, while strictly preserving the semantic integrity and visual coherence of the underlying object.

2. Methodology

A. Benchmark Dataset Construction

The authors created the first comprehensive benchmark for unbranding, comprising 1,759 prompts targeting 12 major brands across six domains (Airlines, Automotive, Beverages, Fast Food, Sportswear, Technology).

Prompt Strategy: Prompts are divided into Explicit (directly mentioning the brand, e.g., "BMW car") and Implicit/Biased (using generic descriptors that strongly imply the brand, e.g., "German luxury car with a kidney grille").
Validation: Baseline generations were verified using Vision Language Models (VLMs) and human annotation to ensure they consistently exhibit identifiable brand elements before being used as ground truth.

B. Evaluation Framework

The paper introduces a novel evaluation protocol that moves beyond simple logo detection:

Brand Presence Score ( $B$ ):
- Uses a Vision Language Model (VLM), specifically Qwen3-VL-8B-Thinking, to perform multi-class classification.
- The VLM is prompted to identify the brand based on both explicit signals (logos) and implicit signals (trade dress, silhouette, color schemes).
- It outputs a hard label and a confidence score. The metric measures the frequency with which the target brand is still correctly identified in the "unbranded" image.
Visual Similarity Score ( $S$ ):
- Measures the preservation of the object's semantic structure.
- Two-stage process: First, a CLIP model filters for gross semantic hallucinations. Second, a VLM performs a side-by-side comparison, instructed to ignore logo removal and stylistic shifts, focusing only on geometry, object identity, and layout.
Unbranding Quality Score ( $U$ ):
- A weighted combination: $U = \alpha \cdot S + \beta \cdot (1 - B)$ .
- This metric balances the need to suppress the brand ( $1-B$ ) with the need to preserve the object ( $S$ ).

C. Baseline Comparisons

The authors evaluated existing state-of-the-art unlearning methods (e.g., ESD, MACE, FMN, RECE, UCE) across different model architectures (Stable Diffusion 1.4, SDXL, FLUX.1-dev) to test their efficacy on the unbranding task.

3. Key Contributions

Formalization of Unbranding: The paper defines unbranding as a distinct, fine-grained task requiring the disentanglement of brand identifiers from object semantics, distinguishing it from general concept unlearning.
Comprehensive Benchmark: Introduction of a dataset covering 12 brands with both explicit and implicit prompts, specifically designed to test the removal of trade dress, not just logos.
Novel Evaluation Metrics:
- Development of a VLM-based metric that outperforms traditional object detectors (like YOLO) in recognizing abstract trade dress.
- Demonstration that Chain-of-Thought (CoT) reasoning in VLMs significantly reduces hallucination rates and improves the detection of subtle brand cues compared to standard classifiers.
Empirical Analysis of Model Fidelity: The study reveals that newer, high-fidelity models (FLUX, SD3.5) reproduce trademarks with significantly higher accuracy than older models, making the unbranding challenge more urgent.

4. Results

Failure of Current Methods: Existing unlearning techniques fail to solve the unbranding task. They typically exhibit a severe trade-off:
- Ineffective Removal: Methods like Prompt Negation preserve image quality but fail to remove the brand (high Brand Residual, e.g., 73% for FLUX).
- Destructive Degradation: Methods like ESD or SDD successfully remove the brand but destroy the object's structure (e.g., turning a car into a blob), resulting in near-zero Visual Similarity scores.
Performance Ceiling: No evaluated method achieved an Unbranding Quality Score ( $U$ ) above 67% (with UCE on SDXL being the highest at 66.95%). Most methods hovered around or below 55%.
VLM Superiority: The Qwen3-VL-8B-Thinking model achieved a 90.8% accuracy and 21.28% hallucination rate in detecting brands, significantly outperforming custom-trained YOLO detectors (54.34% accuracy) which struggled with abstract trade dress and overfitted to color/shape heuristics.
Trade Dress Recognition: Ablation studies confirmed that the VLM can identify brands even when explicit logos are digitally removed, relying solely on structural features (e.g., identifying a BMW by its grille shape alone).

5. Significance and Conclusion

The paper establishes that unbranding is an open and critical challenge for the safe deployment of generative AI.

Legal & Commercial Necessity: As models become more capable of reproducing high-fidelity trademarks, the risk of intellectual property infringement and brand dilution increases. Current "brand-safe" filters are insufficient.
Technical Gap: The field lacks methods that can selectively erase brand identity without destroying the utility of the generated image.
Future Direction: The authors conclude that specialized techniques are required to disentangle brand signals from object semantics. The provided benchmark and evaluation framework serve as a foundational tool for developing these next-generation unlearning algorithms.

In summary, the paper argues that simply "unlearning" a concept is too blunt an instrument for brand protection; the future of generative AI safety lies in fine-grained unbranding that respects both legal IP boundaries and the visual integrity of the generated content.