Train a Flux LoRA on a rented GPU, step by step

A Flux LoRA teaches the Flux image model a new face, product or style from a few dozen pictures. The result is a small file you load next to the model in ComfyUI.

Ostris's AI Toolkit, a free trainer under the MIT licence, does this on one rented GPU. Its example settings for FLUX.1 [dev] are written for a card with 24 GB of memory.

The steps below follow AI Toolkit's README and example settings, read on 27 October 2026.

What you need

• Pictures in JPG or PNG, each with a caption in a text file of the same name. AI Toolkit's README says you do not need to crop or resize them.
• Only photos you own, or that the people in them have agreed to. Never train on someone's face without their consent.
• A Hugging Face account with a read token. FLUX.1 [dev] is gated: request access on its page first, as the README says.
• A rented NVIDIA GPU with 24 GB or more, running Linux.

Check the model's licence before you use the results commercially. FLUX.1 [dev] comes under Black Forest Labs' non-commercial licence; FLUX.1 [schnell] and FLUX.2 [klein] 4B base use the Apache 2.0 licence.

AI Toolkit trains all three, by its README. For FLUX.1 [schnell] it has its own example settings, which need a training adapter the file names.

Step 1: prepare the pictures

Put the pictures and their captions in one folder. Each caption describes the picture and contains a trigger word, a made-up word the LoRA will learn:

my_dataset/
  photo01.jpg
  photo01.txt     # a photo of [trigger] smiling on a beach
  photo02.jpg
  photo02.txt     # a photo of [trigger] in a grey suit, studio light

If you set a trigger word in the settings, AI Toolkit puts it wherever a caption says [trigger], by its README. WebP files currently have issues, it adds, so convert them first.

Step 2: rent the GPU and install AI Toolkit

Connect to the server over SSH and follow the README's manual install for Linux. It pins the PyTorch version it tests:

git clone https://github.com/ostris/ai-toolkit.git
cd ai-toolkit
python3 -m venv venv && source venv/bin/activate
pip3 install --no-cache-dir torch==2.13.0 torchvision==0.28.0 torchaudio==2.11.0 \
  --index-url https://download.pytorch.org/whl/cu130
pip3 install -r requirements.txt
huggingface-cli login      # paste your read token

Copy your dataset folder to the server, for example with scp, into the ai-toolkit folder.

Step 3: set up the training file and train

Copy AI Toolkit's example file for a 24 GB card and change the lines below; they sit in different parts of the file:

cp config/examples/train_lora_flux_24gb.yaml config/my_lora.yml
nano config/my_lora.yml

name: "my_first_flux_lora"        # the output folder and file name
trigger_word: "p3r5on"             # uncomment and set your made-up word
folder_path: "/path/to/ai-toolkit/my_dataset"
steps: 2000                        # 500 to 4,000 is a good range, says the file
name_or_path: "black-forest-labs/FLUX.1-dev"

The example also keeps rank 16, a learning rate of 0.0001 and 8-bit loading of the model, which fits it on 24 GB. Then start the run:

python run.py config/my_lora.yml

Every 250 steps it saves the LoRA and draws sample images, so you can watch it learn. You can stop with Ctrl+C and resume later; the README warns not to stop while it is saving.

Or use the web interface

AI Toolkit also has a web interface, which needs Node.js 20 or newer. The README advises a password on any server, set with AI_TOOLKIT_AUTH:

cd ui
AI_TOOLKIT_AUTH=choose-a-long-password npm run build_and_start
# on your own computer:
ssh -L 8675:127.0.0.1:8675 user@your-server    # then open http://localhost:8675

It serves on port 8675. The SSH tunnel keeps that port off the internet while you use it.

Step 4: test the LoRA in ComfyUI

The finished LoRA is a .safetensors file in output/my_first_flux_lora/. Copy it into ComfyUI's models/loras folder, then add a LoRA loader to a Flux workflow.

Write prompts that use your trigger word, and compare the saved versions from different steps. The example keeps the four most recent saves, so you can keep the one you like best.

Our guide to running ComfyUI in the cloud shows how to install it on the same rented server.

Which card to rent

• 24 GB, such as an RTX 4090 or an L4: what AI Toolkit's example Flux settings are written for.
• 32 GB, such as an RTX 5090: more room for higher resolutions.
• 48 GB, such as an L40S: room for larger batches, or to turn off 8-bit loading.

A run costs its GPU hours times the hourly price. Watch the time between the first two saves, then multiply for the full run before you leave it going.

Our GPU memory guide for image and video models shows what each Flux model needs to make images.

How we checked

We read AI Toolkit's README and example settings, and the Flux model pages on Hugging Face, on 27 October 2026. The settings quoted are AI Toolkit's own.

Our own measurements will be added here, with their dates, when our test bench runs them.

Back to ComfyUI in the cloud