← 研踪 PaperTrailarXiv:2608.05213 · v2
StyleComposer: Training-Free Multi-Reference Style Composition
首次提交 2026-08-05 17:17
版本更新 2026-11-17 15:27
雷达观测 2026-11-18 19:43
摘要
The color, texture, and structure of an image can come from different references, but reference-guided generators transfer each reference as a single style signal. We find that in a diffusion transformer, every pathway that brings in a reference transfers multiple attributes, yet each attribute has one representation in which it is most directly controllable. Guided by this finding, we design StyleComposer to route color through the chromatic coordinates of the latent, texture through the attention keys and values, and structure through the queries, and to limit what else each route admits from its reference. Without training or inversion, StyleComposer follows the references and the prompt together more closely than stylization methods and multi-image generators, and it sets the strength of each attribute separately.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 15:27 · 最新观测版本
StyleComposer: Training-Free Multi-Reference Style Composition
The color, texture, and structure of an image can come from different references, but reference-guided generators transfer each reference as a single style signal. We find that in a diffusion transformer, every pathway that brings in a reference transfers multiple attributes, yet each attribute has one representation in which it is most directly controllable. Guided by this finding, we design StyleComposer to route color through the chromatic coordinates of the latent, texture through the attention keys and values, and structure through the queries, and to limit what else each route admits from its reference. Without training or inversion, StyleComposer follows the references and the prompt together more closely than stylization methods and multi-image generators, and it sets the strength of each attribute separately.
查看此版本解读 →查看此版本原文 ↗