bytedance/UNO

Uses in-context generation with diffusion transformers to synthesize high-consistency multi-subject paired training data, enabling a progressive cross-modal alignment architecture with universal rotary position embeddings. Builds on FLUX.1-dev as the base diffusion model and supports both single and multi-image conditioning for subject-driven generation. Provides training and inference implementations with fp8 quantization support for consumer GPUs (~16GB VRAM), plus a Hugging Face dataset (UNO-1M) and pre-trained weights.

1,353 stars. No commits in the last 6 months.