USO
ByteDance's Unified Style-Subject Optimized — consistent character generation
ByteDance's Unified Style-Subject Optimized controls style and subject appearance simultaneously — consistent character generation.
Image generation and computer vision — detection, editing, 3D.
Key features
- Unified framework for style-driven and subject-driven generation via disentangled learning
- Disentangles content from style: style-alignment training plus content-style
- Style reward-learning paradigm to further enhance output quality
- Natively supported in ComfyUI with official tutorial and workflow examples
- FP8 mode for consumer-grade GPUs; peak VRAM usage around 16GB
Good for
- Style transfer and stylized content generation
- Subject/identity-driven generation preserving person identity
Stack & tags
PythonPyTorchstyle-transfersubject-drivenimage-generationopen-weightscomfyuibytedanceidentity-preserving
Quick start
pip install -r requirements.txt
python inference.py --prompt "A cat sleeping on a chair." --image_paths "" "assets/gradio_examples/style1.webp" --width 1024 --height 1024From the project README on GitHub.
Similar open-source projects
FLUX.2 klein
★ 25,996Black Forest Labs' sub-second image generation — generate and edit in <0.5s
DiffSynth-Studio
★ 13,200The framework behind Qwen-Image-i2L — the first "Image→LoRA" visual generation system