Flux LoRA Training Tools Compared

Four tools cover the Flux LoRA training landscape: the Civit website trainer, Kohya_ss, ai-toolkit, and fluxgym. Each targets a different skill level and hardware profile.

TL;DR

• Civit Trainer: Browser-based Flux LoRA training with zero local setup.
• Kohya_ss: GUI wrapper around Kohya's terminal training scripts; full parameter control.
• ai-toolkit: Open-source YAML-driven suite recommended by its README for consumer GPUs.
• fluxgym: Low-VRAM web UI targeting Flux LoRA training on constrained hardware.

Why Flux LoRA Training Has a Special Challenge

FLUX.1-dev was distilled using a reduced guidance approach. Trainers report this makes direct LoRA training on FLUX.1-dev diverge from the original training process (Flux-Dev2Pro model card). The common workaround is a two-stage approach: recover a more capable base by training on FLUX.1-dev or Flux-Dev2Pro, then train the LoRA on that recovered model.

Civit Website Trainer: The Browser-First Option

The Civit platform hosts a built-in LoRA trainer for Flux. One user compared three runs using the same dataset of AssassinKahb-style skeleton images (article 6748).

V3 used Kappa Neuro's Prodigy configuration, which sets network dimension to 2, learning rate to 0.0005, cosine scheduler with 3 cycles, noise offset of 0.1, and min-SNR gamma of 5 (article 6748). Each run produces a .safetensors LoRA file. Civit caps training at 10,000 steps (article 6748).

The Civit Trainer requires no local installation, making it the lowest barrier to entry.

Kohya_ss: GUI Wrapper for Kohya's Scripts

Kohya_ss is a GUI wrapper around Kohya's training scripts. The Civit setup guide lists these Windows steps (article 8073):

1. Clone: git clone https://github.com/bmaltais/kohya_ss
2. Enter folder, checkout master, pip install rich, setup.bat

The setup menu offers Accelerate configuration, GPU selection, and precision mode. On RTX 30/40/50 cards, select bf16; older hardware uses fp16.

Caution: The current Kohya_ss README does not mention setup.bat or pip install rich; its installer centers on uv or pip install -r requirements.txt, and lists python >= 3.10 with 3.12 recommended rather than Python 3.10.11 as a fixed requirement. Treat article 8073's steps as one reporter's procedure, not the project's current standard.

ai-toolkit: The Open-Source All-in-One Suite

ai-toolkit is a training suite for diffusion models. It supports FLUX.1-dev natively and provides both a CLI and a web UI (ai-toolkit README).

Installation

On Linux or Windows with Python >= 3.10, create and activate a venv, then install dependencies with pip install -r requirements.txt. The README shows Python 3.12 as the recommended version (ai-toolkit README).

Training a Flux LoRA

1. Copy the example config: config/examples/train_lora_flux_24gb.yaml to the config folder (ai-toolkit README).
2. Edit the config file for your model path, dataset directory, and output name.
3. Run: python run.py config/your_config.yml

Using Flux-Dev2Pro

Article 7809 describes swapping the transformer to Flux-Dev2Pro: replace the FLUX.1-dev/transformer directory with Flux1-dev2Pro/transformer inside your training tool's model folder (article 7809). This swap is not documented in ai-toolkit's README.

fluxgym: Low-VRAM Web UI

fluxgym is an alternative interface for low-VRAM Flux LoRA training. The repository describes it as a "dead simple FLUX LoRA training UI with LOW VRAM support" (GitHub). It targets users who want a GUI rather than Kohya's terminal workflow. Exact VRAM figures from community spreadsheets were not independently verified from opened documentation.

Comparison Summary

For related reading on Flux model architecture, see How Flux Diffusion Models Work. For GPU selection guidance, see Choosing a GPU for Diffusion Training. For a practical training workflow walkthrough, see Flux LoRA Training Step-by-Step.

VRAM Requirements: What's Documented

Article 7619 records that one early low-VRAM Flux trainer produced a LoRA in about 12 hours, and that the author tested resolutions from 256 to 1024 before settling on practical settings (article 7619). The article itself was later updated by a newer VRAM-management guide. Exact time-reduction percentages from the community spreadsheet are omitted here because they were not independently verified.

License and Model Considerations

The Flux1-dev2pro model is distributed under the FLUX.1 [dev] Non-Commercial License (Civitai model 654919, Flux-Dev2Pro model card). LoRA models trained with it inherit this restriction for commercial deployment.

The base FLUX.1-dev model is similarly licensed by Black Forest Labs under the FLUX.1 dev Non-Commercial License (Hugging Face repo). Production or commercial use of trained models requires a separate license from Black Forest Labs.

Common Errors and Fixes

FAQ

How do I train Flux LoRA on consumer GPUs? Use ai-toolkit with its example Flux config, or the Civit website trainer for a no-install browser workflow. On low VRAM, start at 512×512 with a small batch size.

What is the best VRAM setup for Flux training? Lower resolution and smaller batch sizes reduce VRAM linearly. Gradient accumulation lets you simulate larger batches without VRAM spikes.

How do I swap the transformer for Flux1-dev2pro? Article 7809 instructs replacing the FLUX.1-dev/transformer directory with Flux1-dev2Pro/transformer inside your training tool's model folder. This exact path is not documented in ai-toolkit's README.

What are recommended settings for Kohya_ss on Windows? Clone the repo, run the setup menu, select bf16 for RTX 30/40/50 series cards, disable DeepSpeed and distributed training, and enable NUMA for NVIDIA hardware. The current Kohya_ss README uses uv or pip install -r requirements.txt rather than setup.bat and pip install rich.

Is Flux1-dev2pro free to use? The model is publicly available under the FLUX.1 [dev] Non-Commercial License. You may use it for research and personal projects but must obtain a commercial license from Black Forest Labs for production deployment.

What is the difference between Kohya_ss and fluxgym? Kohya_ss is a GUI wrapper around Kohya's training scripts. fluxgym is a separate low-VRAM web UI that targets Flux LoRA training on constrained hardware.

Sources