Discord Bot - Stable Diffusion

A discord bot that runs the Stable Diffusion text-to-image model on a local GPU (the computer hosting the discord bot). This uses a fork of stable diffusion repository to do the text to image generation.

---

Requirements

• git
• Anaconda or miniconda python distribution (link)
• An Nvidia GPU with CUDA support (ideally with 8GB+ VRAM)
  • Nvidia drivers

Installation instructions

1. Clone this repo (and its submodules) using

     git clone --recurse-submodules [email protected]:Haydeni0/discord-stable-diffusion
   

2. Navigate to the repository in terminal

3. Follow the instructions here to set it up

   1. Or if using linux/WSL2, run the command

        . install.sh
      

      to automatically set things up (requires an installation of conda).

   2. Download the weights for the model (check the latest here)

      • At the time of writing, the latest one is checkpoint version 1.4 and can be downloaded under "Files and versions" from the Hugging Face page for stable-diffusion-v-1-4-original (note that this requires a signup to the website).

   3. Move and rename the model checkpoint file to

      ./model.ckpt

4. Create a discord application and bot using the first two chapters of this guide (Creating an Application, Creating a Bot)

   Click to view discord bot configuration

   1. Make sure to enable all the "Privileged Gateway Intents"
   2. Add the bot to the discord server with these OAuth2 permissions by opening the URL generated by the OAuth2 URL generator
   3. Generate a discord bot token

5. Add the discord bot token to a file named config.ini in the root directory of the repository (you may have to create this yourself). The contents of the file should be:

   # .ini
   [auth]
   DISCORD_TOKEN=PASTE_YOUR_TOKEN_HERE
   

6. Run the python script bot.py using the command

     python bot.py
   

   or in Windows by using the full path to python, for example:

     C:\ProgramData\Anaconda3\envs\ldm\python.exe bot.py
   

7. Run the bot on the discord server using "/" commands in a discord channel, e.g:

   • /help
   • /txt2img <your prompt here>

---

To-do

Features

• img2img
• Add capability to change txt2img options from discord
• Add option that automatically prefixes prompts with stuff like "4K, 8K, high resolution, award winning, ..."
  • To automatically improve quality of image outputs without typing
• Upscaling
• Make a command to cancel the bot generation if it's taking too long

QOL

• Capture the console output so that the loading bar can be shown in discord, the lstein fork does this somehow
• Unload the model after some time

Development

• Add error handling to the commands

• Don't let txt2img automatically save the images, put that functionality in a separate routine

• Make sure two parallel txt2img things dont occur, give a warning if a txt2img is called again without finishing the previous one

  • Or maybe allow it by putting them in a queue (with a max length), and run them serially so the gpu isn't working on 2 at once

Fixes

• Fix img2img mode producing garbage result because of low resolution
  • Fix the rescaling to use the volume rather than max of the width and height

Problems