← Ye Wang · Homepage

OmniStyle

Filtering High Quality Style Transfer Data at Scale

Ye Wang1, Ruiqi Liu1, Jiang Lin2, Fei Liu3, Zili Yi2, Yilin Wang4*, Rui Ma1,5*

1 Jilin University · 2 Nanjing University · 3 ByteDance · 4 Adobe

5 Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MOE, China

* Corresponding authors

📄 Paper 💻 Code 🤗 Model 🗂️ Dataset

Overview

OmniStyle is an end-to-end style transfer framework based on a Diffusion Transformer (DiT), targeting high-quality 1K stylization with both instruction-guided and reference-guided settings.

OmniStyle-1M is a million-scale triplet dataset (content, style, stylized image) spanning 1,000 style categories. OmniStyle-150K is its high-quality filtered subset for training.

OmniStyle teaser examples

Diverse Stylization Results

OmniStyle result sample 1 OmniStyle result sample 2 OmniStyle result sample 3 OmniStyle result sample 4 OmniStyle result sample 5 OmniStyle result sample 6 OmniStyle result sample 7 OmniStyle result sample 8 OmniStyle result sample 9 OmniStyle result sample 10

Method

OmniStyle dataset generation and filtering pipeline
Large-scale generation pipeline with OmniFilter for quality control.

Dataset

OmniStyle dataset overview
Overview of OmniStyle-1M style/content distribution and triplet examples.

OmniStyle Model

OmniStyle model architecture
The architecture of OmniStyle.

More Results

More OmniStyle results
Instruction-guided and reference-guided style transfer across diverse styles.

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China (No. 62202199, 62406134), the Suzhou Key Technologies Project (No. SYG2024136), and the Fundamental Research Funds for the Central Universities.