This repository is for running MiniGPT-4 on Google Colab as a way to label your images for fine tuning Stable Diffusion models

The system is split into two parts: a server and a client. The server will be serving MiniGPT-4 using oobabooga's text generation ui through its API functionality. Since it'll be hosted on Google Colab, it'll use Cloudflare to open a tunnel that'll allow other computers to access it.

The client will take a folder of image files and send unlabeled ones to the server to be labeled. Although you don't need this client to use MiniGPT-4 per se, without it, you'll be manually querying every image individually and copying and pasting the resulting labels into text files. The client provides a way of automatically labeling your images without taking up much of your time and attention.

Open either the server_latest.ipynb or server_failsafe.ipynb notebook on Google Colab (there should be a link when you open either file.) The latest one might not work if future updates to its dependencies break it but the failsafe version should, albeit an older version especially when you're reading this. If you're not willing to troubleshoot and debug, I'd stick with the failsafe version.

The follow along with the instructions, making sure you copy the URL to the non-streaming (not streaming) server which you'll need to input later into the client notebook. Make sure the links to the Gradio interfaces have also appeared before you start running the client (you don't need to use the Gradio interface but they won't appear until the server has completely started up.

Occasionally, errors might prevent either the streaming or non-streaming server from starting but as long as the non-streaming server has started, you should be able to copy the URL and paste it in the respective box in the client notebook. If the non-streaming server has failed to start, you may need to restart the last cell (there's no need to restart the entire notebook.)

Open the client notebook and click the appropriate button to launch it on Google Colab. Make sure you fill in the fields as instructed. You'll also need to upload the images you want to label to the Google Account that you're running the client from (it can be a different Google Account than the one the server is running on.) Make sure you've copied the non-streaming server's URL from the server notebook and pasted it into the correct field. Make sure you've also entered the path where you've uploaded your images in your Google Drive. If you check the "resume" box, the script will skip the files that already have labels, saving you time though you may want to uncheck this if you've decided to change the prompt.

Once you're done labeling the images, you'll be able to fine tune a Stable Diffusion model (or LoRA) using, say, Kohya-ss or Everydream.