💻 computer science

Erase but Preserve: Controllable Removal of Copyrighted Animation Characters via Optimized Semantic Anchors

This paper proposes a controllable method for erasing copyrighted animation characters from text-to-image diffusion models by optimizing structural and detailed semantic anchors to replace target embeddings, achieving state-of-the-art removal effectiveness while preserving image fidelity and enabling fine-grained control.

Original authors: Qiao Li, Xiaomeng Fu, Wangjia Yu, Runze He, Baisen Wang, Jiao Dai, Jizhong Han

Published 2026-08-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiao Li, Xiaomeng Fu, Wangjia Yu, Runze He, Baisen Wang, Jiao Dai, Jizhong Han

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 a world where you can type a few words into a computer, and it paints a picture for you. This isn't magic; it's a type of artificial intelligence called a "text-to-image diffusion model." Think of these models as incredibly talented, but slightly mischievous, digital artists. They have learned to draw by looking at millions of pictures and the words that describe them. If you ask them to draw a "dog," they can. If you ask for a "sunset," they can do that too. But because they learned from everything on the internet, they sometimes accidentally draw things they shouldn't, like famous cartoon characters that belong to big companies. This is a bit like a kid who loves drawing Mickey Mouse but gets in trouble because they don't own the rights to draw him.

The big question scientists are trying to solve is: How do we teach these digital artists to stop drawing specific, copyrighted characters without making them forget how to draw anything else? Imagine if you had to tell a painter, "Never draw a red hat," but you still wanted them to be able to draw a red apple, a red car, and a red sunset. If you just shout "No red hats!" too loudly, the painter might get confused and start drawing blue apples or forgetting how to draw hats at all. This is the tricky balance between erasing the bad stuff and keeping the good stuff.

This paper, titled "Erase but Preserve," introduces a clever new way to solve this problem for animated characters like Spider-Man, Minions, or Kung Fu Panda. The authors suggest that instead of trying to delete the character from the AI's memory entirely (which is hard and often breaks the AI), we should give the AI a "stand-in" or a "decoy."

Here is how their method works, using a simple analogy: Imagine the AI is a chef who knows exactly how to cook a specific, copyrighted dish (let's say, "The Minion Burger"). You want the chef to stop making that exact burger because it's illegal, but you still want them to make a burger that looks and tastes similar enough to be satisfying, just without the secret sauce.

The researchers' solution has three main steps:

Designing the Decoy: First, they create a special "anchor" concept. Think of this as a secret code word, like "Anchor*," that doesn't exist in the dictionary yet. They train this code word to be a perfect "look-alike" of the forbidden character in terms of its shape and outline (so the burger still looks like a burger), but completely different in its fine details (so it doesn't have the Minion's yellow skin or goggles). It's like molding a clay figure that has the same silhouette as the Minion but is made of gray clay instead of yellow.
The Smart Swap: When you ask the AI to draw a picture, it usually translates your words into a list of instructions (embeddings). If you say "Draw Spider-Man," the AI gets the instruction for "Spider-Man." The researchers' method intercepts this instruction. Instead of letting the AI draw Spider-Man, it quietly swaps the "Spider-Man" instruction with the "Anchor*" instruction. But here is the clever part: they don't just swap it instantly. They wait until the drawing process has settled into a stable shape (like when a pot of soup has stopped bubbling wildly and is just simmering). Then, they make the swap. This ensures the background and the overall scene stay perfect, only changing the specific character.
The Dimmer Switch: One of the coolest features is that you can control how much of the character gets erased. Because the "Anchor" is a mathematical code, you can mix it with the original "Spider-Man" code. You can slide a slider from 0% to 100%. At 0%, you get the original Spider-Man. At 100%, you get the gray, non-copyrighted stand-in. At 50%, you get a weird, half-Spider-Man, half-gray blob. This lets users decide exactly how much of the character they want to keep or remove, which is great for creative control.

The authors tested this on 80 different animation characters, including famous ones like SpongeBob, Bugs Bunny, and Peppa Pig. They compared their method to other ways people have tried to stop AI from drawing these characters. The results were impressive: their method was much better at actually removing the character (only about 4% to 6% of the time did the AI still accidentally draw the character, compared to much higher rates for other methods) while keeping the rest of the picture looking beautiful and natural.

They also showed that this "decoy" code word works on different types of AI models, not just the one they trained it on. It's like creating a universal remote control that works on different brands of TVs. Furthermore, they found that this "Anchor*" code word could be used to help other methods that try to stop AI from drawing bad things, making those other methods work better too.

In short, the paper suggests that by creating a smart, shape-matching "decoy" and swapping it in at just the right moment, we can stop AI from drawing copyrighted characters without ruining the rest of the picture. It's a way to respect copyright laws while still letting people enjoy the creative power of these amazing digital artists. The authors believe this approach could help platforms and creators use AI safely without accidentally breaking the law.

Technical Summary: Erase but Preserve

Problem Statement

The widespread adoption of text-to-image diffusion models has introduced significant copyright risks, particularly the unauthorized generation of copyrighted animation characters. While existing concept erasure techniques exist, they are ill-suited for the specific challenges posed by animation characters:

Anchor Selection Difficulty: Model modification methods typically require mapping a target concept to a "benign" anchor. While generic anchors (e.g., "ground" for "nudity") work for broad categories, they fail for highly distinctive, unique animation characters where suitable semantic synonyms or taxonomic relations are difficult or impossible to define manually.
Lack of Fine-Grained Control: Prompt-based steering methods often rely on discrete text adjustments, offering only coarse control. They struggle to balance the removal of specific infringing characters with the preservation of the user's original creative intent (composition, style, background), often resulting in incomplete erasure or degraded image fidelity.

Methodology

The authors propose a controllable erasure method that operates on the model's continuous textual representation during the generation process. The pipeline consists of three main stages:

1. Anchor Concept Construction

Instead of using pre-defined generic anchors, the method constructs a custom "Anchor*" concept optimized to serve as a benign surrogate for a specific target character. This is achieved by minimizing a combined loss function:

Structural Constraint ( $L_S$ ): Ensures the anchor shares a similar structural outline and global layout with the target to prevent background distortion. This is achieved by maximizing the low-frequency structural similarity between reconstructed samples of the target and the anchor.
Detailed Differentiation ( $L_D$ ): Ensures the anchor's detailed features are distinct from the target to prevent copyright infringement. This is achieved by maximizing the noise prediction error when the anchor prompt is used to reconstruct fine details of the target.

2. Anchor Embedding Optimization

The constructed concept is optimized as a learnable word vector ( $v^*$ ) in the continuous textual embedding space of the text encoder (e.g., CLIP).

Initialization: The vector is initialized using the token "Anchor*".
Optimization: The vector is updated via backpropagation to minimize the combined structural and detailed loss functions defined above.
Output: The process yields an optimized anchor embedding ( $e^*_{anchor}$ ) and corresponding contextualized embeddings for special tokens (SOT, EOT, Paddings).

3. Structure-Aware Adaptive Replacement (Inference)

During image generation, the method replaces target-related embeddings with the optimized anchor embeddings.

Targeted Replacement: Only embeddings corresponding to the target character (and related special tokens like EOT and Paddings) are replaced or fused with the optimized anchor embeddings, leaving unrelated prompt elements untouched.
Adaptive Timing: To preserve scene coherence, the replacement is not applied immediately at the start of denoising. The method monitors the low-frequency structural signals of the denoising process. Replacement is triggered adaptively once the structural layout stabilizes (i.e., when the $L_2$ distance between consecutive low-frequency components ceases to increase significantly). This ensures the global structure is established before the specific character is swapped.

Key Contributions

Novel Erasure Mechanism: A method to erase one or more copyrighted animation characters directly during the generation process by operating on continuous textual representations rather than modifying model weights or relying solely on discrete prompts.
Principled Anchor Optimization: A continuous anchor optimization approach that leverages visual features of the target to identify a controllable, semantically proximate anchor, solving the difficulty of selecting suitable surrogates for unique characters.
Structure-Aware Strategy: A replacement strategy that jointly achieves precise character removal and high-fidelity preservation by dynamically timing the embedding swap based on structural stabilization.
Plug-and-Play Capability: The optimized anchors can be integrated into existing model modification baselines as a benign surrogate, improving their performance without requiring re-engineering of the baseline methods.

Experimental Results

The method was evaluated on a dataset of 80 animation characters using Stable Diffusion v1.4 and various other architectures (SD v1.5, v2, v2.1, SDXL, and the DiT-based Z-Image).

Erasure Effectiveness: The method achieved the lowest identification accuracy for target characters by vision-language models (LLaVA-1.5: 6.0%, BLIP-3: 4.0%), significantly outperforming baselines like SLD, SAFREE, and TraSCE.
Image Fidelity: The method maintained superior image fidelity compared to baselines, achieving the highest SSIM (0.467) and lowest LPIPS (0.505) between original and erased images. It also preserved high aesthetic scores.
Generalization: The method demonstrated transferability across different model architectures (UNet and DiT) and text encoder dimensions without requiring model fine-tuning.
Control and Versatility:
- Fine-Grained Control: By interpolating between the target and anchor embeddings, the method allows users to control the degree of erasure continuously.
- Multi-Target Removal: The method successfully erases multiple characters simultaneously while maintaining visual coherence.
- Baseline Improvement: Integrating the optimized anchors into existing model modification methods (UCE, ESD-u, AC, MACE) consistently improved their erasure accuracy and fidelity preservation compared to using general anchors.

Significance and Claims

The paper claims that this work addresses a critical gap in copyright protection for generative AI. By enabling the removal of specific, highly distinctive animation characters without destroying the overall image quality or user intent, the method offers a practical solution for platform governance and legal compliance.

The authors emphasize that their approach:

Maximizes User Intent: Unlike coarse blocking methods, it preserves the composition, style, and background of the generated image.
Offers Controllability: It provides a mechanism for adjustable erasure degrees, allowing for nuanced enforcement of copyright policies.
Facilitates Deployment: The plug-and-play nature of the optimized anchors and the lack of requirement for model fine-tuning make the method suitable for real-world deployment across diverse diffusion models.

The work is positioned as a step toward trustworthy, user-centric AI systems that can operate within legal frameworks while maintaining creative utility.