Download
USO is ByteDance’s “Unified Style and Subject-Driven Generation” framework, open-sourced to allow customization in generative modeling by disentangling style and subject representation and using reward learning to guide generation. The system is designed such that users can control both “what” is generated (the subject: e.g. a person, object, scene) and “how” it is generated (the style: artistic style, color palette, aesthetic) separately, giving much more flexibility than conventional monolithic generative models. By decoupling style and subject, USO enables reuse of learned style/style-embeddings across different subjects, or vice versa, which makes generation more modular and controllable. The project provides tooling (in Python) including inference and workflow scripts, example configurations, and support for generation pipelines; and as of 2025, USO is also natively supported in some mainstream generative-art UI pipelines (e.g. ComfyUI) to ease adoption.
Features
- Disentangled representation of “subject” and “style” allowing independent control over both
- Reward-learning–based generation to better align output with user-specified constraints (style + subject)
- Open-source Python implementation with example workflows and inference scripts included
- Integration support for existing generative-art UIs (e.g. ComfyUI) for easier adoption
- Flexibility to reuse style embeddings across different subjects (or vice versa) to mix and match generation parameters
- Ability to support future extensions (new styles, new subject classes) thanks to modular, extensible design
Project Samples
