- Category
- AI Tools
- Rank
- No. 1081Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- bytedance
- GitHub
- 1.2k stars
- Date
About
Unified style- and subject-driven image generation via disentangled and reward learning (CVPR 2026).
What it does
USO creates images from a text prompt plus optional reference images. The first reference supplies subject or layout content, while later references supply one or more visual styles. It supports subject preservation, style transfer, combined subject-and-style generation, layout-preserving edits, and prompt-only generation.
Why it's ranked here
USO deserves attention because one inference stack covers several reference-image workflows that normally require separate setups. It also offers both scriptable generation and a visual interface. The cost is substantial setup and memory demand, so this is a capable research release rather than a lightweight creative utility.
What's good
The input model is unusually flexible. Users can preserve a subject, borrow a style, combine both, blend multiple style references, or retain layout with an empty prompt. Batch generation accepts JSON task lists and supports multiple processes. Native ComfyUI support also makes the model composable with ControlNet and LoRA workflows.
Tradeoffs
Installation requires a compatible Python environment, pinned machine-learning dependencies, several external checkpoints, environment variables, and potentially a Hugging Face token. Even the reduced-memory FP8 mode reports roughly 16 GB for one reference and 18 GB for multiple references. Multi-style input is labeled beta, and base-model licensing remains the user's responsibility.
How to use it well
It suits researchers and technical image creators who already manage local diffusion models and want one workflow for identity, layout, and style references. Start with the supplied examples, match portrait framing to the requested composition, and use ComfyUI when combining other controls. It does not replace a hosted, low-resource image service or a documented training toolkit.
Technical notes+
inference.py provides dataclass-driven CLI inference, JSON evaluation batches, Accelerate process sharding, deterministic seeds, and PNG plus JSON output. app.py builds a Gradio interface around USOPipeline. In uso/flux/pipeline.py, content references pass through the autoencoder while style references become SigLIP inputs. uso/flux/model.py projects SigLIP hidden features into the text-conditioning sequence and concatenates encoded content references with image tokens. uso/flux/util.py configures FLUX variants, LoRA processors, local environment paths, and optional Hugging Face loading. weights/downloader.py fetches USO, FLUX, T5, CLIP, and SigLIP assets.
Observed
- License
- Apache License 2.0; users must also follow the selected base model's licensing terms.
- Primary language
- Python
- Installation surface
- Python 3.10 through 3.12 environment with dependencies installed from requirements.txt; checkpoints are downloaded separately.
- Interfaces
- Command-line inference, local Gradio interface, and native ComfyUI workflows.
- Model assets
- Uses separate USO, FLUX, T5, CLIP, SigLIP, autoencoder, LoRA, and projection checkpoints.
- Device choices
- The application exposes CUDA, CPU, and Apple MPS device selections.
- Low-memory mode
- FP8 with model offloading is documented at roughly 16 GB for one reference and 18 GB for multiple references.
Read from README.md, requirements.txt, app.py, inference.py, weights/downloader.py, uso/flux/math.py, uso/flux/util.py, uso/flux/model.py, uso/flux/pipeline.py, uso/flux/sampling.py, uso/flux/modules/layers.py, uso/flux/modules/autoencoder.py.
What it can do
Generate images with specific artistic styles
Style reference and content prompt → Stylized image
Generate images with specific subject matter
Subject description and parameters → Subject-focused image
Separate style and subject elements in image generation
Combined style and subject specifications → Image with disentangled style and subject control
Apply reward-based learning to improve image quality
Training data and reward signals → Optimized image generation model
Control image generation with unified style-subject parameters
Unified style and subject control parameters → Precisely controlled generated image
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.