IntroStyle: Training-Free Introspective Style Attribution using Diffusion Features

Conference: ICCV 2025 arXiv: 2412.14432 Code: GitHub Area: Image Generation / Style Attribution Keywords: Style attribution, diffusion model features, training-free, Wasserstein distance, copyright protection, style retrieval

TL;DR

This paper proposes IntroStyle, a training-free style attribution method that leverages only channel-wise mean and variance statistics from intermediate layers of a diffusion model's own denoising network, measuring inter-image style similarity via the 2-Wasserstein distance. IntroStyle substantially outperforms supervised state-of-the-art methods on WikiArt and DomainNet without any task-specific training.

Background & Motivation

• Core Problem: T2I diffusion models (e.g., Stable Diffusion, DALL-E 3) are trained on large-scale web data and may replicate copyrighted artistic styles. How can one determine whether a generated image imitates a specific artistic style without retraining the model or introducing external modules?
• Limitations of Prior Work:
  • Style unlearning requires expensive model retraining and cannot fully prevent indirect replication via alternative prompts.
  • Style cloaking degrades the authentic viewing experience and places the burden on creators.
  • Existing attribution methods (e.g., CSD, GDA) require training external models (fine-tuned CLIP/DINO), incurring high computational cost and deployment complexity.
  • Semantic entanglement: existing methods tend to retrieve semantically similar rather than stylistically similar images.
• Key Insight: Different layers of the diffusion UNet naturally disentangle structural, chromatic, and textural attributes; the internal features of the model are already sufficient for style attribution—no external modules or additional training are required.
• Dataset Gap: Existing evaluation benchmarks (e.g., WikiArt) lack fine-grained separation between style and semantics, making it difficult to assess whether a method truly distinguishes "style" from "content."

Method

Overall Architecture

IntroStyle treats the diffusion model's denoising network as an autoencoder, extracting style descriptors from intermediate feature layers and measuring style similarity via a probabilistic distance metric. The entire pipeline consists of forward inference passes with no training or fine-tuning required.

Three-step pipeline: 1. Image \(I\) is encoded into a latent \(z_0 = \mathcal{E}(I)\) via the VAE, then noised to timestep \(t\): \(z_t = \sqrt{\bar{\alpha_t}} z_0 + \sqrt{1-\bar{\alpha_t}} \epsilon_t\) 2. \(z_t\) is passed through the denoising network \(\epsilon_\theta\) (using a null text embedding \(c=\varnothing\)), and feature tensors \(F^{t,idx}\) are extracted from the specified upsampling block \(idx\). 3. Channel-wise statistics of the feature tensor are computed as the style descriptor.

Key Designs: IntroStyle Features

Motivated by the AdaIN formulation in StyleGAN—where style information is encoded primarily in the first-order (mean) and second-order (variance) statistics of features—the per-channel statistics of feature tensor \(F^{t,idx}\) are defined as:

\[\mu_c = \frac{1}{WH} \sum_{i,j} F^{t,idx}_{c,i,j}, \quad \sigma_c^2 = \frac{1}{WH} \sum_{i,j} (F^{t,idx}_{c,i,j} - \mu_c)^2\]

The style feature vector is: \(f^{t,idx}(I) = (\mu_1, \ldots, \mu_C, \sigma_1^2, \ldots, \sigma_C^2)^T\)

Here \(t\) (noise timestep) and \(idx\) (UNet layer index) are hyperparameters; upsampling block \(idx=1\) is used by default.

Similarity Metric: 2-Wasserstein Distance

Each image's IntroStyle representation is modeled as a \(C\)-dimensional multivariate Gaussian with diagonal covariance. The style distance between two images is measured by the \(W_2\) distance:

\[W_2^2 = \|\mu_1 - \mu_2\|_2^2 + \text{tr}(\Sigma_1 + \Sigma_2 - 2(\Sigma_1^{1/2}\Sigma_2\Sigma_1^{1/2})^{1/2})\]

The diagonal covariance assumption enables efficient computation. Compared against alternative metrics including L2, Gram matrices, and JSD, \(W_2\) yields the best performance.

ArtSplit Synthetic Dataset

To rigorously evaluate style–semantics disentanglement, the paper introduces the Artistic Style Split (ArtSplit) dataset: - For each real painting, two types of prompts are designed: "semantic prompts" (removing style cues) and "style prompts" (removing semantic content). - A diffusion model is used to generate synthetic images covering all "style × semantics" combinations. - This enables precise quantification of a method's sensitivity to style versus semantics.

Key Experimental Results

Main Results: Style Retrieval Performance

Method	WikiArt mAP@1	WikiArt mAP@10	WikiArt Recall@100	DomainNet mAP@1
VGG-Net Gram	0.259	0.194	0.804	-
CSD (Prev. SOTA)	requires training	requires training	requires training	requires training
IntroStyle	substantial lead	substantial lead	substantial lead	substantial lead

• IntroStyle substantially outperforms all baselines (CSD, GDA, etc.) on both WikiArt and DomainNet without any training.
• Qualitative results show that CSD/GDA retrieval is heavily biased by semantic similarity, whereas IntroStyle accurately focuses on style.

Ablation Study

Ablation	Finding
UNet layer selection	Upsampling block 1 (\(idx=1\)) performs best; encoder layers are biased toward semantics.
Timestep \(t\)	Moderate noise levels are optimal; performance degrades at both extremes.
Distance metric	\(W_2\) > L2 > Gram > JSD
Text conditioning	Null text (\(c=\varnothing\)) outperforms conditioning on actual prompts.

Key Findings

• IntroStyle demonstrates strong style-focusing ability on the ArtSplit style-vs.-semantics disambiguation benchmark.
• The upsampling layers of diffusion models naturally separate style from content, validating the feasibility of introspective attribution.
• The method can be directly applied to style-based rejection sampling to prevent generation of specific artistic styles.

Highlights & Insights

1. Zero training overhead: A fully training-free approach requiring no additional datasets, fine-tuning, or external models—only the diffusion model's own internal features are used.
2. Strong style–semantics disentanglement: By selecting the appropriate UNet layer and channel-wise statistics, the method naturally avoids semantic entanglement—a persistent challenge for existing trained approaches.
3. Theoretical elegance: The AdaIN-inspired style representation (mean + variance as style) is naturally combined with a probabilistic distribution distance (\(W_2\)), forming a coherent theoretical framework.
4. ArtSplit dataset: Provides the first fully crossed "style × semantics" benchmark for precise evaluation of style attribution methods.
5. Practical utility: Directly applicable to copyright protection—detecting whether generated images imitate a specific artist's style.

Limitations & Future Work

• Relies on a specific diffusion model architecture (Stable Diffusion UNet); generalization to newer architectures such as DiT has not been verified.
• The ArtSplit dataset is based on synthetically generated images, which may introduce distribution bias.
• The definition of "style" is inherently subjective—social-constructive definitions (by artist or genre) do not necessarily capture all dimensions of visual style.
• The noise timestep \(t\) and layer index \(idx\) are hyperparameters requiring manual selection.
• Applicability to non-Western artistic traditions or digital art styles remains to be validated.

Related Work & Insights

• Style transfer: Gram matrices (Gatys et al.), AdaIN (Huang and Belongie), AdaIN layer-based style control in StyleGAN.
• Style-aware T2I models: Textual Inversion, DreamBooth, and related personalized generation methods.
• Data attribution: CSD (contrastive style descriptors), GDA (generative data attribution), unlearning-based attribution.
• Diffusion feature exploitation: DDAE (latent representation classification), REPA (representation alignment), zero-shot correspondence and segmentation.

Rating

Dimension	Score (1–5)
Novelty	4
Theoretical Depth	3.5
Experimental Thoroughness	4
Writing Quality	4
Value	4.5
Overall	4.0

The core contribution lies in the discovery that channel-wise statistics of diffusion model intermediate features serve as effective style descriptors—a minimal yet highly effective approach. The ArtSplit dataset fills a gap in style attribution evaluation. The main limitations are the method's dependency on the UNet architecture and the insufficient theoretical explanation of the "introspective" capability.

Related Papers

• [ICCV 2025] LaRender: Training-Free Occlusion Control in Image Generation via Latent Rendering
• [AAAI 2026] AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction
• [CVPR 2026] HAM: A Training-Free Style Transfer Approach via Heterogeneous Attention Modulation for Diffusion Models
• [ICCV 2025] OmniVTON: Training-Free Universal Virtual Try-On
• [ICCV 2025] Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing