cocktailpeanut/fluxgym

Dead simple FLUX LoRA training UI with LOW VRAM support

Flux Gym – Simple Web UI for training FLUX LoRA models on modest GPUs

What it is – A lightweight Gradio‑based web interface that wraps the powerful Kohya‑ss training scripts, letting you fine‑tune FLUX diffusion models (LoRAs) with as little as 12 GB of VRAM. The UI is a fork of the AI‑Toolkit front‑end, but unlike the original it works on 12‑, 16‑ and 20‑GB cards.

Why it matters – Training LoRAs for the new FLUX family normally requires a 24 GB GPU or a command‑line workflow. Flux Gym gives the same flexibility (full Kohya‑ss flag set) through a point‑and‑click browser UI, plus a few niceties such as automatic model download, sample‑image generation during training, and one‑click publishing to Hugging Face.

Core features (as described in the README)

  • Low‑VRAM support – works on 12 GB/16 GB/20 GB GPUs.
  • Full Kohya‑ss compatibility – the Advanced tab exposes every training flag supported by the latest sd‑scripts release.
  • Automatic model handling – base FLUX models (dev, dev2pro, schnell) are downloaded on‑demand; additional models can be added via models.yaml.
  • Sample‑image generation – optional generation of preview images every N steps, with full control over prompts, seeds, size, CFG, etc.
  • Hugging Face publishing – store a personal HF token locally and push trained LoRAs directly from the UI.
  • Docker & Pinokio one‑click installers – ready‑made containers and a 1‑click launcher for hassle‑free setup.
  • Caption‑file upload – supports paired .txt caption files matching image filenames.

Quick‑start overview (from the README)

  1. Install – either use the Pinokio 1‑click launcher, run the provided Docker compose, or clone the repo and set up a Python virtual environment.
  2. Run – python app.py (with the venv active) starts a Gradio server at http://localhost:7860.
  3. Train – in the UI:
    • Fill LoRA metadata.
    • Upload images (and optional .txt captions).
    • Optionally enable sample‑image generation and set prompts.
    • Click Start.
  4. Publish – after training, log in with a Hugging Face token and push the LoRA.

Installation details (as documented)

  • Manual: clone the repo, clone the sd‑scripts submodule (git clone -b sd3 https://github.com/kohya-ss/sd-scripts), create a venv, install sd‑scripts requirements, install the app’s requirements.txt, then install a PyTorch nightly build (cu121 or cu128 for RTX 50‑series).
  • Docker: same folder layout, set PUID/PGID if needed, then docker compose up -d --build.
  • Pinokio: click the provided Pinokio link for an automated installer.

Supported base models (auto‑downloaded)

  1. Flux‑1‑dev
  2. Flux‑1‑dev2pro (see linked Medium article for background)
  3. Flux‑1‑schnell (supported but not recommended for high quality)
  4. Custom models – add entries to models.yaml.

Extensibility

  • The UI builds the Advanced tab by parsing the latest kohya‑ss/sd‑scripts launch flags, so any new flag added upstream appears automatically.
  • Adding new base models only requires editing models.yaml and committing a PR if you want to share them.

Who’s using it

The README points to a community showcase page on Pinokio where users share their locally‑trained LoRAs.

Bottom line

Flux Gym is a genuine, open‑source tool that bridges a user‑friendly web UI with the full power of Kohya‑ss training scripts, making FLUX LoRA fine‑tuning accessible on consumer‑grade GPUs.

Related