StyleComposer lets AI artists mix color, texture, structure from 3 references

Style transfer has long treated a painting's style as a single, monolithic signal—but real artistic style is a combination of distinct attributes. Color palettes, brushstroke textures, and structural composition often come from different sources, and existing reference-guided methods force them through one representation, leaving users unable to control each attribute's origin or intensity. A new paper from Sanghyeok Lee, Jihye Kang, and Namhyuk Ahn tackles this with StyleComposer, a training-free framework that breaks style into color, texture, and structure, then independently routes each attribute through the diffusion model representation where it separates best. This routing is dynamically coordinated across denoising timesteps, allowing the model to combine multiple references without fine-tuning or inversion.

StyleComposer's key innovation is its decomposition strategy. The authors systematically investigated where in a diffusion model each style attribute can be manipulated independently while the others remain fixed, discovering that no single representation isolates all three. Instead, each attribute has a preferred latent layer or attention map. StyleComposer leverages this by assigning each attribute its optimal route and then syncing those routes over the generation timeline to maintain coherence. The result is a system that satisfies up to three reference images and a text prompt jointly, outperforming prior methods that blend references into one style signal. Crucially, it also exposes a per-attribute strength slider, letting users dial color intensity, texture roughness, or structural influence independently. This granular control, achieved entirely without training or inversion, makes StyleComposer a practical upgrade for AI art pipelines where artists want to remix styles from multiple sources with precision and speed.