kohya-ss GUI

Maintained by GitHub user bmaltais, this project puts a browser interface over kohya-ss's sd-scripts, the widely used command-line trainers for diffusion models. Each tab exposes training parameters as form fields, then assembles and runs the matching accelerate command. Methods include LoRA with its LoHa and LoKr variants, DreamBooth, full fine-tuning, textual inversion and LECO concept erasure, plus masked loss, preview images rendered while training and GPU selection. Base models range from SD 1.5 and 2.x, SDXL and SD3 to FLUX.1, HunyuanImage 2.1 and Lumina Image 2.0. A config.toml prefills model, dataset and output paths, and a headless flag makes the app safe on remote GPU boxes over SSH. Windows is the primary target, installed with uv or pip, Python 3.11 and the CUDA 12.8 toolkit; Linux is kept working by community contributors and macOS may not cooperate. An NVIDIA GPU is required, memory needs depend on model and method, and Docker, Runpod and Colab routes exist for anyone without one. Apache-2.0 licensed. Aimed at hobbyists and small teams training custom styles or characters without memorizing CLI flags.