StyleComposer: Training-Free Multi-Reference Style Composition

2026-08-05

AI Summary

Existing approaches treat artistic style as a monolithic signal, making it difficult to independently control the source and intensity of distinct stylistic attributes such as color, texture, and structure. This work proposes a diffusion-based multi-path style routing mechanism that dynamically fuses stylistic features from multiple reference images during the denoising process via a time-coordinated strategy. For the first time, the method enables fine-grained, disentangled adjustment of the three stylistic attributes—without requiring additional training or inversion—and supports independent pairing with different reference images. Experiments demonstrate that the approach more accurately aligns generated outputs with both textual prompts and the three distinct reference styles while offering intuitive, per-attribute intensity control.

Abstract

The style of a painting is not monolithic: color, texture, and structure may come from different sources. Existing reference-guided methods transfer them as one style signal, leaving each attribute's source and strength outside the user's control. We ask where in a diffusion model one attribute can change while the others hold, and find that no single representation isolates all three. The proposed StyleComposer therefore routes each style attribute through the representation where it separates best and coordinates the routes over denoising time. Without training or inversion, it satisfies three references and the prompt jointly more closely than prior methods, and exposes one strength slider per attribute. Project page: https://lexxsh.github.io/StyleComposer

Problem

• style transfer
• multi-reference
• attribute control
• diffusion models
• style composition

Innovation

• style composition
• diffusion models
• multi-reference
• attribute disentanglement
• training-free