StyleDecoupler: Generalizable Artistic Style Disentanglement

TL;DR

The paper tackles the challenge of disentangling artistic style from semantic content in visual representations by leveraging the complementary properties of multimodal and unimodal vision models. It introduces StyleDecoupler, an information-theoretic, plug-and-play module that minimizes content-related information to extract a purer style signal from frozen VLM embeddings, without fine-tuning. A key contribution is the WeART benchmark, a large-scale dataset with 280k artworks across 152 styles and 1,556 artists, designed to enable robust evaluation of style representations and generative-model fidelity. Empirically, StyleDecoupler achieves state-of-the-art style retrieval on both WikiArt and WeART, demonstrates meaningful style manifolds, and provides reliable metrics for downstream tasks, including generative evaluation, while maintaining strong generalization to unseen styles in a zero-shot setting.

Abstract

Representing artistic style is challenging due to its deep entanglement with semantic content. We propose StyleDecoupler, an information-theoretic framework that leverages a key insight: multi-modal vision models encode both style and content, while uni-modal models suppress style to focus on content-invariant features. By using uni-modal representations as content-only references, we isolate pure style features from multi-modal embeddings through mutual information minimization. StyleDecoupler operates as a plug-and-play module on frozen Vision-Language Models without fine-tuning. We also introduce WeART, a large-scale benchmark of 280K artworks across 152 styles and 1,556 artists. Experiments show state-of-the-art performance on style retrieval across WeART and WikiART, while enabling applications like style relationship mapping and generative model evaluation. We release our method and dataset at this url.

Figures (3)

• Figure 1: Overview of our information-theoretic style disentanglement framework. (a) Feature Space Alignment: We align DINO features with the CLIP embedding space using knowledge distillation. (b) Style Disentanglement: Guided by GPT-4o generated descriptions, we extract and separate style vectors from content vectors.
• Figure 2: Our method is motivated by the observation that VLMs, unlike unimodal models, perform robustly on both natural and artistic retrieval. We leverage this by introducing an information-theoretic framework that explicitly disentangles the VLM's native style and content features.
• Figure 3: The t-SNE visualization of representation models.