Fluxgym setup: train Flux.1 Dev LoRAs on Windows

Uploaded on September 11, 2024, this tutorial covers Fluxgym for Flux.1 Dev LoRA training on Windows. No Fluxgym release number is supplied; the installation steps and interface reflect the setup shown in that upload.

The speaker introduces a local AI training interface with 12, 16 and 20 GB VRAM settings. They describe its frontend as based on AI Toolkit's Gradio UI and its training backend as Kohya's SD scripts. The manual installation walkthrough covers cloning both repositories, creating a Python virtual environment and installing dependencies, including Torch packages. It also shows how to reuse downloaded model files from ComfyUI, with fp16 text encoders, the Flux model and VAE placed in the expected folders.

The training example sets a LoRA name and trigger word, uploads character images and uses Florence 2 to generate captions. The speaker notes that the first captioning run downloads Florence 2. They keep 10 repeats per image and 16 epochs, then follow progress in the browser interface.

Using the 16 GB setting on a card with 24 GB available, the speaker reports completion in 3 hours and 5 minutes. That is a result from this demonstration, not a general timing guarantee. The final section transfers the saved LoRA to ComfyUI and tests character likeness with a pose-guided ControlNet workflow and without ControlNet.