FLUX LoRA training
A few days ago, I started experimenting with local AI image generation using ComfyUI – building workflows in a web interface and running current open-weight models such as FLUX.1-dev entirely on my own hardware.
The results were good, but not adapted to any specific subject. LoRA training addresses precisely this.
LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method, originally developed for large language models, that keeps the base weights frozen and learns low-rank updates to selected weight matrices. Rather than retraining the full model, a small adapter is trained that encodes a particular style, subject, or concept, and is then applied on top of a base model such as FLUX within ComfyUI.
For training, I used FluxGym:
- A simple web UI for training FLUX LoRAs
- Designed for low-VRAM setups (12–20 GB)
The workflow itself consists of five steps:
- naming the LoRA;
- defining trigger words;
- uploading the training images (automatic captioning is supported);
- training;
- loading the resulting LoRA into ComfyUI for inference.
Results
As a personal project, I trained a LoRA on roughly 150 images of my cat Carli:
- Sleeping, playing, sitting, running
- Different lighting and environments
The resulting outputs are shown in the video below.
A side note on hardware
- In FluxGym's low-VRAM mode, training on 20–30 images runs fine on my local RTX 3080 (12 GB)
- For the full 150-image dataset, local training would have taken far too long
- Training on the full dataset was therefore carried out on an NVIDIA A100 (40 GB) on an HPC cluster
- Training time: approximately five hours
The experiment indicates that locally operated generative models are practicable on consumer hardware for inference and small-scale adaptation, although the full 150-image run still required an HPC GPU.