cocktailpeanut

fluxgym: Train FLUX LoRAs with a Web UI

FluxGym provides a web interface for training FLUX LoRAs using Kohya sd-scripts, with options designed for 12GB, 16GB, and 20GB VRAM systems. It suits creators who want a guided workflow without giving up Kohya’s advanced training options.

View on GitHub Compare
Stars
3.3k
Forks
371
Stars, last 30 days
+5

Our take

Health score

36/100

How it is scored
Activity
33
Community
42
Issues
21
Pull requests
46

+5 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

3,263

Forks

372

Change since Nov 17, 2026: 0 stars·0 forks

Overview

FluxGym is a local web UI for training FLUX LoRAs. It combines a Gradio interface based on AI-Toolkit’s UI with Kohya sd-scripts as the training backend, aiming to make training more approachable while retaining access to Kohya’s options.

It is intended for users whose GPUs have less VRAM than the 24GB requirement cited for AI-Toolkit in the README. The project advertises support for 12GB, 16GB, and 20GB VRAM configurations. It can also download selected base models automatically and supports extending the model list through models.yaml.

Key Features

  • Train FLUX LoRAs through a browser-based interface.
  • Use Kohya sd-scripts features through an advanced settings tab.
  • Upload training images and provide captions, including matching .txt caption files.
  • Configure sample image prompts and generate samples periodically during training.
  • Use sample-generation flags for options such as seed, dimensions, CFG scale, and steps.
  • Select supported base models, which are downloaded automatically when training starts.
  • Publish trained LoRAs to Hugging Face from the UI.
  • Run locally using a manual setup, Pinokio launcher, or Docker.

Use Cases

  • FLUX users with a 12GB, 16GB, or 20GB VRAM GPU who want to try LoRA training through a guided interface.
  • Creators who prefer uploading images and captions in a UI rather than assembling training commands by hand.
  • Experienced Kohya users who want a web interface while retaining access to advanced script options.
  • Users who want to inspect training progress by generating sample images at specified intervals.
  • LoRA authors who want to publish a completed model to Hugging Face from the application.

Getting Started

For a manual setup, clone FluxGym and its Kohya sd-scripts dependency, install the dependencies in a virtual environment, then start the app:

git clone https://github.com/cocktailpeanut/fluxgym
cd fluxgym
git clone -b sd3 https://github.com/kohya-ss/sd-scripts

After following the dependency installation steps in the README, launch from the FluxGym directory with the environment active:

python app.py

Considerations

  • Training requires a suitable GPU and a compatible PyTorch/CUDA setup. The README describes 12GB, 16GB, and 20GB VRAM support, but actual performance and fit depend on the training configuration.
  • Manual installation includes a separate Kohya sd-scripts checkout and nightly PyTorch installation. NVIDIA RTX 50-series users are directed to a different CUDA wheel and an updated bitsandbytes package.
  • The README describes Flux1-schnell as not recommended for high-quality results, while Flux1-dev and Flux1-dev2pro are also listed as supported.
  • The project relies on Kohya sd-scripts, so changes in that dependency and its launch options may affect setup and training behavior.
  • Publishing to Hugging Face requires a token. The README says the token is saved locally in an HF_TOKEN file.

See the repository README for complete setup and configuration instructions.

Found this useful?

Share it with someone who would like fluxgym.