cs.CVNov 7, 2026

Disentangled Stylized Image Generation via IP-Cross-Attention Rearrangement

Authors: Beibei Dong, Yueming Lyu, Yue Jiang, Jing Dong

Organizations: Institute of Automation, Chinese Academy Of Sciences, China

Abstract

With a solid theoretical foundation and pretraining on large-scale datasets, diffusion-based methods have achieved controllable generation leveraging text and image conditions, and generating excellent stylized images is possible using diffusion models. However, previous methods face significant challenges. First, because diffusion generation is iterative, the common approach of learning implicit style representations via cycle-consistency strategies is too costly and not well-suited to diffusion, even for LoRA-based methods. Second, decoupling style is difficult in the latent space of diffusion models. Representations of attributes are coupled in the latent space of diffusion models; thus, extracting style attributes with diffusion models usually suffers from the issue of content leakage. Furthermore, because style is hard to represent in natural language, capturing style using text prompts or textual inversion is deficient. In this paper, we propose a method for achieving disentangled stylized image generation through IP-cross-attention rearrangement, based on our experimental observations. We also propose negative content guidance, inspired by classifier-free guidance, to reduce content leakage in stylized image generation.

Figures & tables

Explore similar work

Feb 24, 2026

CleanStyle: Plug-and-Play Style Conditioning Purification for Text-to-Image Stylization

Style transfer in diffusion models enables controllable visual generation by injecting the style of a reference image. However, recent encoder-based methods, while efficient and tuning-free, often suffer from content leakage, where semantic elements from the style image undesirably appear in the output, impairing prompt fidelity and stylistic consistency. In this work, we introduce CleanStyle, a plug-and-play framework that filters out content-related noise from the style embedding without retraining. Motivated by empirical analysis, we observe that such leakage predominantly stems from the tail components of the style embedding, which are isolated via Singular Value Decomposition (SVD). To address this, we propose CleanStyleSVD (CS-SVD), which dynamically suppresses tail components using a time-aware exponential schedule, providing clean, style-preserving conditional embeddings throughout the denoising process. Furthermore, we present Style-Specific Classifier-Free Guidance (SS-CFG), which reuses the suppressed tail components to construct style-aware unconditional inputs. Unlike conventional methods that use generic negative embeddings (e.g., zero vectors), SS-CFG introduces targeted negative signals that reflect style-specific but prompt-irrelevant visual elements. This enables the model to effectively suppress these distracting patterns during generation, thereby improving prompt fidelity and enhancing the overall visual quality of stylized outputs. Our approach is lightweight, interpretable, and can be seamlessly integrated into existing encoder-based diffusion models without retraining. Extensive experiments demonstrate that CleanStyle substantially reduces content leakage, improves stylization quality and improves prompt alignment across a wide range of style references and prompts.
Apr 20, 2026

UniCSG: Unified High-Fidelity Content-Constrained Style-Driven Generation via Staged Semantic and Frequency Disentanglement

Style transfer must match a target style while preserving content semantics. DiT-based diffusion models often suffer from content-style entanglement, leading to reference-content leakage and unstable generation. We present UniCSG, a unified framework for content-constrained, style-driven generation in both text-guided and reference-guided settings. UniCSG employs staged training: (i) a latent-space semantic disentanglement stage that combines low-frequency preprocessing with conditioning corruption to encourage content-style separation, and (ii) a latent-space frequency-aware detail reconstruction stage that refines details via multi-scale frequency supervision. We further incorporate pixel-space reward learning to align latent objectives with perceptual quality after decoding. Experiments demonstrate improved content faithfulness, style alignment, and robustness in both settings.
Apr 14, 2026

MAST: Mask-Guided Attention Control for Training-Free Regional-Multi Style Transfer

Style transfer applies the appearance of a reference image to a content image while preserving its spatial structure. Recent diffusion-based methods achieve strong stylization but typically assume a single global style. We instead consider regional-multi style transfer, which assigns multiple references to user-specified regions of a content image. Extending them to this setting reveals two coupled shared-attention issues: ambiguous mass allocation among content and style partitions and degraded selectivity as more styles are jointly normalized, while style aggregation at the attention output further suppresses fine details. We propose MAST (Mask-Guided Attention Control for Training-Free Regional-Multi Style Transfer), a unified attention-control framework for frozen diffusion models. Logit-level Attention Mass Allocation enforces mask-derived partition masses, Sharpness-aware Temperature Scaling adaptively restores selectivity, and Discrepancy-aware Detail Injection recovers high-frequency content. MAST jointly processes all style--mask pairs in a single denoising pass without training, optimization, or post-hoc composition. Across two to five styles, MAST achieves the best average ArtFID, FID, and R-FID among all baselines, demonstrating regional style fidelity, content preservation, and scalability.