/discord-diffusion

A self-hosted discord bot running Stable Diffusion 1.4

Primary LanguagePython

Discord Diffusion

A self-hosted discord bot running Stable Diffusion 1.4.

Setup

  1. Get a discord bot token, connect it to your server. TODO: document
  2. Get a huggingface token. TODO: document
  3. Run the docker image jamesrom/discord-diffusion:latest

Example:

docker run --rm -it --gpus=all \
    --volume $PWD\huggingface:/home/huggingface/.cache/huggingface \
    --volume $PWD\output:/home/huggingface/output \
    --env DISCORD_TOKEN=$BOT_TOKEN_HERE> \
    --env HUGGINGFACE_TOKEN=$HF_TOKEN_HERE \
    jamesrom/discord-diffusion:latest

Configuration

Configure by environment variables:

Variable Default Description
DISCORD_TOKEN required — Discord token.
HUGGINGFACE_TOKEN required — Huggingface token.
DEFAULT_HEIGHT 512 Must be divisible by 8.
DEFAULT_SEED Random Set the default seed for reproducibility. Must be an integer.
DEFAULT_WIDTH 512 Must be divisible by 8.
FILE_FORMAT jpeg Must be jpeg or png.
MAX_COUNT 5 Set the maximum number of generated images per command. If set to 1, will disable the command option altogether.
MAX_HEIGHT 2048 Maximum height of generated images.
MAX_WIDTH 2048 Maximum width of generated images.
NSFW_GLOBAL_ENABLE false Bypass the built-in safety check globally.
NSFW_IN_AGE_RESTRICTED_CHANNELS true Disable the safety check only in Age-Restricted Channels.
NSFW_USER_OVERRIDE false Allow the user to disable the safety check with the nsfw command option.
SAVE_IMAGE_TO_DISK false Whether to save a copy of the generated image to disk.
SD_ATTENTION_SLICING false Use less memory at the cost of speed.
SD_MODEL CompVis/stable-diffusion-v1-4 The huggingface model to download and use.
SD_REVISION main The model revision to use.

Volume mounts

TODO

Usage

Type /text2img and enter your prompt. Optional parameters can be given:

  • width - The desired width of the generated image. Must be divisible by 8 and smaller than the configured MAX_WIDTH.
  • height - The desired width of the generated image. Must be divisible by 8 and smaller than the configured MAX_HEIGHT.
  • count - The number of images to generate. If MAX_IMAGE_COUNT is 1, this option will be unavailable.
  • nsfw - Disable the built-in safety check. This option will be unavailable if GLOBAL_NSFW_ENABLE is true, or ALLOW_NSFW_USER_OVERRIDE is false.

nolifer

Memory requirements

At least 8GB recommended. Here's some things you can do if you see memory errors:

  1. Set REVISION=fp16, this uses half-sized tensors (16 bit instead of 32 bit) and should reduce memory footprint by about half.
  2. Enable attention slicing with ATTENTION_SLICING=true. This lowers memory requirements at the cost of speed.
  3. Lower DEFAULT_WIDTH and DEFAULT_HEIGHT, similarly MAX_HEIGHT and MAX_WIDTH

Building from source

Easy:

docker buildx build . --tag my-cool-bot

Thanks

Thanks to the following prior work for inspiration: