StyleComposer

Training-Free Multi-Reference Style Composition

Sanghyeok Lee · Jihye Kang · Namhyuk Ahn

Given separate references, StyleComposer composes their color, texture, and spatial structure in one generation. The method requires no training or inversion and provides one control per attribute.

Paper Code

StyleComposer teaser showing outputs as color, texture, and structure references are replaced independently

TL;DR

Color, texture, and structure are routed separately.

Instead of representing style as a single signal, StyleComposer handles each attribute in a different representation of the frozen diffusion model.

Colorpalette and tone Texturemedium and brushwork Structurelayout and scale

Analysis

Representation analysis

Image embedding

Content-dependent embedding directions

The same edit produces different CLIP directions across scenes, limiting whole-image embedding arithmetic.

Content-dependent edit directions in CLIP image embedding space

Attention

Asymmetric but coupled attention routes

Texture follows K–V correspondence, color follows global V statistics, and Q is biased toward layout.

K and V interventions reveal different roles for texture and palette transfer

VAE latent

A compact latent color representation

Color edits concentrate in a three-dimensional chromatic subspace, providing an explicit color coordinate system.

Method

Attribute-specific routing

Three coordinated routes operate on a frozen FLUX model.

Color

Latent alignment

Align chromatic latent coordinates with the color-reference distribution.

Texture

K/V routing

Route reference K/V features while limiting positional content transfer.

Structure

Early Q guidance

Guide spatial layout with reference queries during early denoising.

The pretrained model remains frozen, and the three routes are coordinated within the same sampling process without inversion.

Results

Three-reference composition

We evaluate all three references together with prompt alignment.

A clock tower

Additional compositions

Browse compositions using one, two, or three independently assigned references.

A field of flowers

Quantitative evaluation

Three-reference benchmark

All methods use one fixed configuration across 250 cases. Higher values are better.

Method	Composition	Selectivity	Human pref.
FLUX	0.347	0.434	6.3%
SDXL	0.407	0.430	5.3%
InstantStyle	0.243	0.324	10.3%
StyleAligned	0.274	0.440	9.5%
IP-Adapter	0.303	0.370	13.7%
SADis	0.432	0.446	20.0%
B-LoRA	0.301	0.391	1.5%
StyleComposer	0.621	0.745	33.3%

Interactive control

Attribute strength

Move each control while keeping the prompt, seed, and references fixed.

Color0.0

→

Generated output with adjustable color strength — Output

lesscolor strengthmore

Texture1.0

→

Generated output with adjustable texture strength — Output

lesstexture strengthmore

Structure0

→

Generated output with adjustable structure strength — Output

lessstructure strengthmore

Pairwise control

Pairwise attribute balance

Move the slider between two attribute emphases while the prompt, references, and seed remain fixed.

Color priorityTexture priority

Citation

BibTeX

arXiv:2608.05213, 2026.

@article{lee2026stylecomposer,
  title     = {StyleComposer: Training-Free Multi-Reference Style Composition},
  author    = {Lee, Sanghyeok and Kang, Jihye and Ahn, Namhyuk},
  journal   = {arXiv preprint arXiv:2608.05213},
  year      = {2026}
}