USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning

2025-08-26

AI Summary

Existing methods treat style-driven and subject-driven image generation as disjoint tasks, failing to jointly optimize style fidelity and subject consistency. To address this, we propose USO—the first unified framework that synergistically models content and style via disentangled learning. Our approach constructs a large-scale triplet dataset and introduces a three-stage training paradigm: style alignment, content-style disentanglement, and style reward learning (SRL), all built upon diffusion-based optimization. Furthermore, we release USO-Bench—the first comprehensive benchmark for joint evaluation of style fidelity and subject consistency. Extensive experiments demonstrate that USO achieves state-of-the-art performance among open-source models on both metrics, significantly outperforming prior methods. To foster reproducibility and community advancement, we fully open-source the code, pre-trained models, and dataset.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Learning to Search
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

Existing literature typically treats style-driven and subject-driven generation as two disjoint tasks: the former prioritizes stylistic similarity, whereas the latter insists on subject consistency, resulting in an apparent antagonism. We argue that both objectives can be unified under a single framework because they ultimately concern the disentanglement and re-composition of content and style, a long-standing theme in style-driven research. To this end, we present USO, a Unified Style-Subject Optimized customization model. First, we construct a large-scale triplet dataset consisting of content images, style images, and their corresponding stylized content images. Second, we introduce a disentangled learning scheme that simultaneously aligns style features and disentangles content from style through two complementary objectives, style-alignment training and content-style disentanglement training. Third, we incorporate a style reward-learning paradigm denoted as SRL to further enhance the model's performance. Finally, we release USO-Bench, the first benchmark that jointly evaluates style similarity and subject fidelity across multiple metrics. Extensive experiments demonstrate that USO achieves state-of-the-art performance among open-source models along both dimensions of subject consistency and style similarity. Code and model: https://github.com/bytedance/USO

Problem

• Unifying style-driven and subject-driven image generation tasks
• Disentangling content and style features for better composition
• Enhancing both style similarity and subject consistency simultaneously

Innovation

• Unified Style-Subject Optimized customization model
• Disentangled learning scheme with style-alignment and content-style objectives
• Style reward-learning paradigm to enhance performance