FluxGym

FluxGym is a local web interface for training FLUX LoRAs on your own images, with support for GPUs with 12GB, 16GB or 20GB of VRAM. It's for people who want to customize an image model through a browser while retaining access to detailed training controls. It runs on Windows and Linux, with Docker support, and is open source under the MIT license.

Its Gradio interface builds on a fork of AI-Toolkit's WebUI, while Kohya sd-scripts handles training. The main screen keeps the basic workflow small: supply images and captions, describe the LoRA, and start training. An expandable Advanced tab exposes the full set of Kohya sd-scripts features for users who need more control.

FluxGym supports Flux1-dev and Flux1-dev2pro, and accepts custom base models. It also supports Flux1-schnell, though the project reports poor training results with that model. It downloads selected models automatically, so fetching model files requires an internet connection.

You can import existing text captions alongside images and generate sample images at chosen intervals during training. Samples help you compare the LoRA's output as training progresses. Controls include custom image resolution, prompts and fixed seeds for comparisons across training stages.

Training runs on your machine. Optional publishing sends a trained LoRA to Hugging Face and requires a Hugging Face token, which FluxGym stores locally. The MIT license covers FluxGym’s code. FLUX.1-dev weights use a separate noncommercial license, and their download requires accepting the model terms. Other base models and derived adapters have their own conditions.