A web ui and deployable API for 🤗 diffusers.
< under development, request features using issues, prs not accepted atm >
If something doesnt work as expected, or if you need some features which are not available, then create request using github issues
- text to image
- image to image
- instruct pix2pix
- textual inversion
- inpainting
- outpainting (coming soon)
- image info
- stable diffusion upscaler
- gfpgan
- clip interrogator
- more coming soon!
- text to image
- image to image
- instruct pix2pix
- textual inversion
- inpainting
- outpainting (via inpainting)
- more coming soon!
To install bleeding edge version of diffuzers, clone the repo and install it using pip.
git clone https://github.com/abhishekkrthakur/diffuzers
cd diffuzers
pip install -e .
Installation using pip:
pip install diffuzers
To run the web app, run the following command:
diffuzers app
To run the api, run the following command:
diffuzers api
Starting the API requires the following environment variables:
export X2IMG_MODEL=stabilityai/stable-diffusion-2-1
export DEVICE=cuda
If you want to use inpainting:
export INPAINTING_MODEL=stabilityai/stable-diffusion-2-inpainting
To use long prompt weighting, use:
export PIPELINE=lpw_stable_diffusion
If you have OUTPUT_PATH
in environment variables, all generations will be saved in OUTPUT_PATH
. You can also use other (or private) huggingface models. To use private models, you must login using huggingface-cli login
.
API docs are available at host:port/docs
. For example, with default settings, you can access docs at: 127.0.0.1:10000/docs
.
❯ diffuzers app --help
usage: diffuzers <command> [<args>] app [-h] [--output OUTPUT] [--share] [--port PORT] [--host HOST]
[--device DEVICE] [--ngrok_key NGROK_KEY]
✨ Run diffuzers app
optional arguments:
-h, --help show this help message and exit
--output OUTPUT Output path is optional, but if provided, all generations will automatically be saved to this
path.
--share Share the app
--port PORT Port to run the app on
--host HOST Host to run the app on
--device DEVICE Device to use, e.g. cpu, cuda, cuda:0, mps (for m1 mac) etc.
--ngrok_key NGROK_KEY
Ngrok key to use for sharing the app. Only required if you want to share the app
❯ diffuzers api --help
usage: diffuzers <command> [<args>] api [-h] [--output OUTPUT] [--port PORT] [--host HOST] [--device DEVICE]
[--workers WORKERS]
✨ Run diffuzers api
optional arguments:
-h, --help show this help message and exit
--output OUTPUT Output path is optional, but if provided, all generations will automatically be saved to this
path.
--port PORT Port to run the app on
--host HOST Host to run the app on
--device DEVICE Device to use, e.g. cpu, cuda, cuda:0, mps (for m1 mac) etc.
--workers WORKERS Number of workers to use
If you want to use private models from huggingface hub, then you need to login using huggingface-cli login
command.
Note: You can also save your generations directly to huggingface hub if your output path points to a huggingface hub dataset repo and you have access to push to that repository. Thus, you will end up saving a lot of disk space.