lora-vast-ai

This repo is for documentation purposes.

Getting Started

Prerequisistes:

• vast.ai account
• 10-15 images
• labels with descriptions as txt files (in txt labels I always used "[trigger]")

Training

Start a vast.ai container with enough GPU ram. At least for vast.ai it was helpful to use directly a pytorch template with cuda installed. For the installation follow the instructions for FLUX.1 Training.

I needed to install another package:

sudo apt-get update
sudo apt-get install libgl1-mesa-glx -y

My repo uses the flux schnell weights.

data structure

└── ai-toolkit/
    └── dataset/
        ├── image-1.jpg
        ├── image-1.txt
        ├── image-2.jpg
        ├── image-2.txt
        ├── ...
      └── .../

At least for me with around 10-15 images it took me around 3 hours to train the lora models (with around 4 epochs).

Inference

For the inference I used another ComfyUI FLUX.1 on Vast.ai.

It takes a while until you can access the workflow UI (~ around 5-10 minutes). After that, you can just upload the workflow.json file from this repo and start creating new images with your own prompts.

Acknowledgments

Big thanks to ostris for creating ai-toolkit.