FluxGym

FluxGym is a browser interface for training FLUX LoRAs on your own images. It's for people who want to customize image models locally, with support for GPUs with 12GB, 16GB or 20GB of VRAM.

• Images and captions: Train with your own image collection and import matching text caption files alongside the images.
• Training controls: The optional Advanced tab exposes Kohya sd-scripts features, while the main interface keeps basic training tasks together.
• Progress samples: Generate sample images at chosen training intervals to compare outputs as training proceeds; sample generation is off by default.
• Base models: Use Flux1-dev or Flux1-dev2pro, or add custom base models. Flux1-schnell is supported, though the project reports poor training results with it.
• Hugging Face publishing: Optionally upload trained LoRAs to Hugging Face with an account token, which the app stores locally.

FluxGym runs on Windows and Linux and supports Docker. Training takes place on your machine. Downloading selected models requires an internet connection, and optional publishing sends the trained LoRA to Hugging Face.

The interface builds on a fork of AI-Toolkit's Gradio WebUI, with Kohya sd-scripts handling training. FluxGym's code is open source under the MIT license.