quickdif

Quick and easy CLI inference that just works™ for a variety of Diffusers models

Including

• Verified working across many models
  • Stable Diffusion (1.5/2.1)
  • Stable Diffusion XL Base
    • No refiner
  • Stable Diffusion 3
  • Stable Cascade
  • Wuerstchen
  • Kandinsky22
  • Pixart (Alpha/Sigma)
  • Hunyuan
  • Kolors
  • AuraFlow
  • Other Diffusers format models will likely work to varying degrees
• Multi-lora fusion using peft for minimal performance loss
• Extremely small 1-shot script using accelerate for hot loading models
• Features all of the most common generation parameters
• Includes many advanced generation parameters such as latent augmentation
• Iterate over most parameters and create grids
• Expand prompts with photo of {all|of|these} or photo of [one|of|these]
• Load settings from JSON, TOML, PNG
• Targeted AMD/ROCm optimizations
  • NVIDIA/CUDA does not need additional optimizations

Not Including

• ControlNet + other Stable Diffusion extensions
• Multi-stage models: DFIF Stage 2, SDXL Refiner
• Server/API for a perpetual instance
• 100% maximum throughput
• 1-click installer
• MacOS/MPS and Intel OneAPI may not function properly

Installation

This project only supports pip based installs on Python 3.11+

Basic setup

> git clone https://github.com/Beinsezii/quickdif.git
> cd ./quickdif/

Create a venv in the quickdif folder

> python3 -m venv ./venv
# alternatively, it's recommended to use the full virtualenv if you have it
> virtualenv ./venv

Install dependencies

# replace with appropriate activation script for other shells
> source ./venv/bin/activate
# It's recommended to first install torch using the recommended commands from https://pytorch.org/get-started/locally/
> pip install torch --index-url https://download.pytorch.org/whl/rocm6.0 # AMD example
# finally
> pip install -e .
> deactivate
> ./quickdif.sh "kitten"
# Windows users will instead have to invoke Python directly
> python quickdif.py "kitten"

As a module

Additionally, you may use the project as a python module

pip install git+https://github.com/Beinsezii/quickdif.git
python -m quickdif "kitten"

Usage

# See all options. Always refer to the script help over the other examples in this README
> ./quickdif.sh --help

# Run with defaults
> ./quickdif.sh "high resolution dslr photograph of pink roses in the misty rain"
# Custom model
> ./quickdif.sh -m "ptx0/terminus-xl-gamma-v1" "analogue photograph of a kitteon on the beach in golden hour sun rays"
# Single files work for Stable Diffusion
> ./quickdif.sh -m ./checkpoints/sd15/dreamshaper-v6.safetensors "colorful fantasy artwork side profile of a feminine robot in a dark cyberpunk city"
# Four dog and four cat images at twenty steps
> ./quickdif.sh "photo of a dog" "painting of a cat" -B 4 -s 20
# Colored latent for an offset noise effect
> ./quickdif.sh "underwater photograph of a dark cave full of bioluminescent glowing mushrooms" -g 9.0 -s 30 -C black -c 0.8
# Compile for a long job
> ./quickdif.sh $(cat prompts.txt) --compile
# Export favorite settings to the defaults JSON
> ./quickdif.sh -m "stabilityai/stable-cascade" -s 20 -n "blurry, noisy, cropped" --json ./quickdif.json
# Save a style to a custom JSON
> ./quickdif.sh "fantasy artwork of a kitten wearing gothic plate armor" -g 10 -G 0.5 --json ./epic_kitten.json
# Merge multiple configs
> ./quickdif.sh -I underwater_cave.png epic_kitten.json

F.A.Q.

Question	Answer
Why not X popular UI?	SD.Next's diffusers backend is extremely buggy/broken in multiple areas and InvokeAI (+non-diffusers UIs) only really supports Stable Diffusion.
Windows?	The python script should work just fine but you'll need to set up the powershell/CMD stuff on your own.
Gradio?	No. If a UI ever gets made for this it'll be its own separate entity that interfaces via API. Cramming a bad gradio interface into the main script wont do anyone any favors
Feature XYZ?	Maybe. Things in the Not Including list may come eventually™ if this script winds up being useful enough