From Concept Erasure to Style Purification: Contrastive Eigenbases for Artist Style Protection

arXiv:2602.08059 · cs.CV, cs.AI · Submitted 2026-02-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "From Concept Erasure to Style Purification".

Tom: The DICE framework is a training-free, inference-time method designed for on-the-fly artist style erasure in diffusion models to combat style mimicry and protect intellectual property.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: This paper, "From Concept Erasure to Style Purification: Contrastive Eigenbases for Artist Style Protection," is basically proposing a training-free way to strip away an artist’s style from a generated image, focusing on preserving the user's original content instead of just replacing it with another style.

Jane: Exactly. Instead of trying to swap out styles, which often ruins the picture or reduces diversity, they are aiming for this purification process where they remove the artist's characteristics while keeping what was intended in the image itself.

Lu: The authors introduce a core idea centered around constructing contrastive triplets—Anchor, Positive, and Negative samples—to get the model to distinguish between style elements and content elements within its latent space. This is a clever mathematical way to compel the model toward disentanglement.

Meng: That sounds complex; how does this contrastive setup actually work in practice without needing any manual labeling or extensive training data specific to each artist?

Lalam: It suggests that we can build a system where any generated image can be analyzed and purified on the fly, which means every user's output could be protected from unauthorized style replication instantly.

The paper's summary: Tom: The main thrust of this work is that they solve the problem of style mimicry by using a contrastive approach to decompose the latent space into separate style and content subspaces, which allows them to perform on-the-fly style purification rather than needing explicit replacement styles.

Jane: So, instead of just guessing what a style is, they use mathematical relationships between different inputs to figure out exactly what part of the image is the artist's signature and what part is the content we want to keep.

Lu: They formalize this disentanglement using a generalized eigenvalue problem based on Canonical Correlation Analysis, aiming to find a projection direction that maximizes stylistic correlation between an anchor and a positive sample while minimizing content correlation with a negative sample.

Meng: So, if I understand correctly, they are using these mathematical projections to create two separate spaces—one for style and one for content—and then they use those spaces during inference to selectively edit the image.

Lalam: That separation is powerful because it means the system isn't just blindly suppressing everything that looks artistic; it's targeting the specific, mathematically defined style components.

The paper's improvements: Tom: The authors propose a couple of key improvements, first using orthogonal suppression on Key and Value matrices to strip out style features by projecting the token features onto the learned style subspace, and second, they use a different subspace for content enhancement during the Query matrix editing phase.

Jane: That dual strategy is smart; they aren't just removing style, which often hurts things, but they are simultaneously using another component to reinforce the content boundaries that might get weakened by the removal process.

Lu: The Adaptive Erasure Controller comes in here; it calculates a "style score" for each token based on its position in the style subspace and then combines these scores to create an adaptive erasure strength factor, which makes the removal uneven and localized.

Meng: I see how that addresses the issue of non-uniform style intensity across different parts of an image; if a patch has a lot of brushstrokes, it gets more aggressive erasure than a smooth area. But what about the limitations they mentioned?

Lalam: The authors pointed out that this method relies on pre-computed style and content subspaces derived from the contrastive setup, which means while it's training-free for deployment, generating those initial representations is still a computational step.

Conclusion: Tom: So we've seen how "From Concept Erasure to Style Purification: Contrastive Eigenbases for Artist Style Protection" uses contrastive eigenbases and adaptive controllers to move away from clumsy style editing toward a training-free purification method that preserves content integrity.

Jane: It really boils down to achieving that optimal balance between thoroughly removing an artist's signature and ensuring the core visual information remains intact, which they show they can do better than existing methods.

Lu: The potential here is huge; we could see AI tools becoming incredibly versatile for content repurposing, allowing creators to safely manipulate styles without infringing on others' intellectual property.

Meng: From an engineering perspective, it means deployment systems can handle style removal in real-time with much lower computational overhead than previous iterative optimization methods they mentioned.

Lalam: I think the biggest implication is cultural; imagine a future where artists can freely remix content without fear of having their signature style being automatically stolen by a model.

Tom: That's a powerful thought, Lalam, and it really frames the impact of this work on how we deploy generative models responsibly.

Jane: It certainly offers a new direction for ensuring that AI tools serve as creative partners rather than just mimics of existing human work.

Tong Zhang, Ru Zhang, *Jianyi Liu

Beijing University of Posts and Telecommunications

cs.CV, cs.AI

Submitted: 2026-02-08

Updated: 2026-09-29

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: The DICE framework is a training-free, inference-time method designed for on-the-fly artist style erasure in diffusion models to combat style mimicry and protect intellectual property.

Key concepts

DICE framework
A training-free, inference-time method designed for on-the-fly artist style erasure in diffusion models to combat style mimicry and protect intellectual property.
Contrastive triplets
A core idea where the authors introduce Anchor, Positive, and Negative samples to compel the model to distinguish between style elements and content elements within its latent space.
Style purification
The goal of removing an artist's characteristics while keeping what was intended in the image itself, rather than just replacing styles which can ruin the picture or reduce diversity.

Terminology

Summary

The DICE framework is a training-free, inference-time method designed for on-the-fly artist style erasure in diffusion models to combat style mimicry and protect intellectual property. It addresses the limitations of existing methods that require explicit replacement styles or cause content distortion by performing style purification, removing the artist’s characteristics while preserving user-intended content. This approach is crucial for providing a practical, efficient deployment-side safety measure against unauthorized style replication.

Core Insight and Problem Formulation

The paper posits that a model cannot truly comprehend the artist style from a single text or image alone. To overcome this, the framework abandons the traditional paradigm of identifying style from isolated samples. Instead, it constructs contrastive triplets to compel the model to distinguish between style and non-style features in the latent space. The triplet comprises:

  1. Anchor (PA): Contains the artistic style to be erased and content to be preserved (e.g., A flower in the style of Van Gogh).

  2. Positive (PP): Contains the target artist’s style but with different content (e.g., A road in the style of Van Gogh).

  3. Negative (PN): Contains different artistic styles but the same content to be preserved (e.g., A flower in the style of Monet).

Style and Content Disentanglement via Generalized Eigenvalue Problem

The problem of disentangling style and content within the latent space is formalized as a solvable generalized eigenvalue problem, based on Canonical Correlation Analysis (CCA). The objective is to find a projection direction that maximizes stylistic correlation between the Anchor and Positive samples while minimizing content correlation between the Anchor and Negative samples. This is formulated as maximizing the Rayleigh Quotient:

(8) arg u ρ(u) = u T (ΣAP'ΣP'A)u / u T (ΣAA + λΣAN'ΣN'A)u

By solving this problem, the framework derives a subspace that captures the core style features. The artistic style is then represented by selecting the eigenvectors corresponding to the top 'r' largest eigenvalues, forming the style subspace:

(14) Ustyle = [u1, u2,..., ur] ∈ R D×r

Guided Erasure and Preservation via Attention Decoupling Editing

During inference, the pre-computed style and content subspaces are used for precise editing through two complementary strategies:

  1. Orthogonal Suppression on K and V matrices: The style information is stripped by computing the coordinate vector of each token feature in the style subspace, projecting it onto it, and subtracting this projection from the original Key (K) and Value (V) matrices to yield edited matrices K' and V'. This removes style-related features such as texture and brushstrokes.

  2. Content Enhancement on Q matrix: To compensate for content loss during style removal, a different subspace, Ucontent, is derived by swapping the reward and penalty terms in the optimization problem (Equation 19). The Query (Q) matrix is then enhanced by projecting it onto Ucontent and applying a weighted addition of this content component to recover weakened content boundaries:

(22) Q' = Q + γq((QUcontent)U T content)

Adaptive Control Mechanisms

The framework incorporates dynamic mechanisms to handle non-uniform style intensity across image patches:

  1. Adaptive Erasure Controller (AEC): This component calculates the style score for each token feature vector based on its coordinate vector in the style subspace Ustyle, quantifying local style concentration. It then combines these scores (from Q, K, and V components) using weights to obtain an overall patch style score (Si). This score is transformed into a smooth modulation factor mi via a sigmoid function to determine the adaptive erasure strength γ:

(28) γ = αmin + (αmax − αmin) · mi

  1. Attention Decoupling Editing Ablation: Ablations showed that suppressing style in only K and V causes substantial content damage, while the full approach—using Q for content enhancement and K/V for style suppression—achieves the highest holistic index Ho = Cstyle−Ccontent.

Experimental Validation

Extensive experiments validate DICE's performance across multiple artist styles. The evaluation uses two primary metrics:

  1. Clip Score (csstyle and cscontent): A refined criterion where style prompts and content prompts are used separately to accurately assess erasure completeness versus content retention completeness, respectively.

  2. Differential-LPIPS (DLPIPS): This metric measures perceptual differences across three dimensions: Erasure Generalization, Erasure Effectiveness (Cstyle), and Content Preservation (Ccontent). The paper demonstrates that DICE achieves the optimal balance between the thoroughness of style erasure and the integrity of content preservation, achieving superior scores on both metrics compared to baselines.

Improvements for AI systems

Based on the DICE framework described in this paper, here are the specific improvements that can be made to existing image generation models, and what these improved systems will be capable of doing:

I. Improving Model Deployment Safety and Intellectual Property Protection

The primary improvement is shifting from costly, explicit style editing (which requires knowing the target style) to a training-free, on-the-fly style purification process.

  1. ​Extract and Purify Artist Style from Generated Content

  2. ​The improved system can take any generated image (or prompt) and autonomously identify the unique artistic signature of the artist's latent space representation (the style subspace). It then applies a differential editing strategy to remove this style component while preserving the core content (e.g., subject, layout, objects).

  3. ​This allows platforms to provide a style purification service that removes an artist's specific aesthetic without needing prior knowledge of their training data or explicit style parameters.

II. Enhancing Robustness Against Style Mimicry Attacks

The framework directly counters malicious style mimicry attempts where users use prompts like in the style of [target style].

  1. ​The system can actively detect and neutralize these stylistic injections during inference by comparing the input latent features against a learned, disentangled representation of the target artist's style.

  2. ​This provides a strong defense mechanism against adversarial attacks (like those discussed in Section 2.3) designed to hide an artist's signature within content or prompt embeddings.

III. Achieving High Fidelity in Style Erasure and Content Preservation Simultaneously

The paper introduces complex, adaptive mechanisms to ensure that the removal of style features does not degrade the essential visual information of the content.

  1. ​Through the use of a generalized eigenvalue problem derived from contrastive triplets, the system can precisely separate style directions from content directions in latent space, leading to highly targeted erasure rather than broad suppression.

  2. ​The Adaptive Erasure Controller ensures that erasure strength is applied dynamically based on the style concentration of individual image patches (using scores derived from Q, K, and V vectors). This means high-intensity brushstrokes in a sky background are erased thoroughly, while subtle texture details on a foreground object are left intact.

  3. ​The Attention Decoupling Editing strategy further refines this by suppressing style features from the Key (K) and Value (V) matrices while using the Query (Q) matrix to selectively enhance structural content components, leading to a purified output that retains clean edges and natural forms.

IV. Enabling Fine-Grained Control Over Style Intensity

The sensitivity analysis reveals that style intensity is non-uniform across an image, which is crucial for high-quality results.

  1. ​The system can be tuned via the spatial dimension of the style subspace (r). By selecting an optimal range (e.g., 12–21 eigenvectors), it can suppress complex artistic features like swirling brushstrokes and color palettes comprehensively without introducing excessive noise artifacts or residual textures in areas where they are not present.

  2. ​The content enhancement coefficient (γq) allows for fine-tuning the trade-off: increasing this coefficient can boost content structure preservation, ensuring that even aggressive style removal leaves the primary subject clear and morphologically accurate (e.g., preserving an apple's contour).

In summary, the improved AI system will be a Style Purification Engine capable of taking any generated image and deterministically stripping away its artistic style while guaranteeing that the underlying content remains structurally sound, resulting in high-quality, de-stylized images with superior content fidelity compared to current methods.

Sources

Related papers