logo
pub

Using FluxGym for Low VRAM Flux LoRA Training

Introduction

FluxGym is a web UI for training Flux LoRAs with low VRAM (12G, 16G, etc). Developed by Black Forest Labs using Kohya-ss/sd-scripts, it aims to simplify the training process for users with less powerful GPUs. This note summarizes common issues users face and offers solutions.

Problem Description

Users experience various issues while trying to use FluxGym, including high VRAM usage, implementation problems, and discrepancies in training results.

Common issues include:

  • High VRAM requirements even after optimizations
  • Training discrepancies
  • Errors caused by environment configurations
  • Slow training times on low VRAM setups

Issue 1: High VRAM Usage

Even with the optimizations, some users find VRAM usage still too high. For instance:

UPDATE: Just learned that the Florence-2 Auto-caption was not clearing the cache... this alone seems to shave off 4GB VRAM! Now, the 20G option runs with just 16G.

Link: GitHub Change

Solution:

  • Ensure to pull the latest version from GitHub.
  • Use the torch.cuda.empty_cache() function to manually clear cache where applicable.

Issue 2: Training Configurations

Some users are unclear on how to adjust training parameters for better results. Example:

so this base config is optimal for a 4090? or is there more speed to be squeezed out of it with a diff config?

Comparison:

  • A4500: 58 minutes (1300 steps)
  • 4090: 20 minutes (1200 steps)

Solution:

  • Use the advanced tab to customize settings such as epochs, learning rates, and resolutions.
  • Example of settings adjustment:
    https://x.com/cocktailpeanut/status/1832113636367876446

Issue 3: Environment Setup Problems

Errors often stem from environment and dependency issues. Example:

return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass" for over an hour and half.

Solution:

  • Ensure Python compatibility and verify that all dependencies are installed correctly.
  • Review this discussion for help:
    https://github.com/pinokiofactory/factory/discussions/6

Issue 4: Discrepancies in Training Outcomes

Users report inconsistent results when training. Example:

does anyone know how to adjust settings for better results when original training does not match expectations?

Solution:

  • Review and adjust parameters like steps and epochs based on dataset quality and size.
  • Use higher resolutions for detailed work:
    https://x.com/cocktailpeanut/status/1832098084794356081

Additional Tips

Tip 1: Keep Up with Updates

Check for updates regularly to benefit from new optimizations.

Tip 2: Use Proper Datasets

High-quality and well-tagged datasets result in better training outcomes.

Tip 3: Engage with the Community

Participate in relevant forums and discussions for support and advice.

By following these strategies and tips, you can make the most out of FluxGym for training Flux LoRAs with low VRAM. Happy training!

SHARE