FLUX LoRA Training
A few days ago, I started experimenting with local AI image generation using ComfyUI - building workflows via a web interface and running cutting-edge models like Flux-1-dev entirely on my own hardware.
That worked surprisingly well.
But I wanted more personalized results.
👉 That’s where LoRA training comes in.
LoRA (Low-Rank Adaptation) is a lightweight approach for fine-tuning large image generation models. Instead of retraining the full model, you train a small add-on that teaches the base model a specific style, subject, or concept - which can then be stacked on top of models like Flux inside ComfyUI.
For training, I used FluxGym:
- A simple web UI for training FLUX LoRAs
- Designed for low VRAM setups (12-20 GB)
The workflow is refreshingly straightforward:
- Name your LoRA
- Define trigger words
- Upload images (automatic captioning supported)
- Train
- Drop the resulting LoRA into ComfyUI and start generating images in your own style.
Watch the result!
Personal project: I trained a LoRA using ~150 images of my cat Carli:
- Sleeping, playing, sitting, running
- Different lighting and environments
The results are honestly impressive - have a look on the attached video!
Side note on hardware
- With FluxGym’s low-VRAM mode, training with 20–30 images works on my local RTX 3080 (12 GB)
- For the full 150-image dataset, training would have taken far too long locally
➡️ I used an NVIDIA A100 (40 GB) on an HPC cluster
➡️ Training time: ~5 hours
This experiment really highlights how far local, customizable generative AI has come - from inference to personalization.