Stable Diffusion XL Image Batch Generator

A Python script that uses Stable Diffusion XL (SDXL) to generate batches of images based on your prompt. I made this script around the time SDXL first came out and decided to upload it now.

• Supports SDXL base 1.0 and SDXL refiner 1.0 models
• Stable Diffusion Video (SVD) and SVD XT are implemented but not currently working
• Prompt weighting is supported with the syntax: "(token_string)weight_float"
• You can optionatly load LoRAs for use with the base model
• The diffusers library is set to offline mode in the script, so you have to manually download models
• Tested and working on Linux Mint 21.3 with a NVIDIA GTX 1070 (8GB VRAM)

Copyright 2024, Benjamin Steenkamer - see LICENSE for details

Types of Batch Runs

Run different types of image generation batches by commenting/uncommenting calls in the __main__ function. All the main configuration variables (prompts, number of inference steps, number of images, etc.) are set in __main__ and you can change them as needed. This script supports:

• SDXL base model image generation, with or without LoRAs
• SDXL image in-painting: load an input and image mask, white areas of the mask will be transformed in the input image, black areas will be kept the same
• Image to Image conversion: load all the images in the output folder and transform into new images
• SDXL refiner model image generation to make refined images that are closer to the original prompt than with just the base model alone
• SDXL refiner model image generation to make images with more fine details

VRAM Limitations

• This script is optimized to run on "low" VRAM GPUs like the NVIDIA GTX 1070 with only 8 GB of VRAM
• If you have more VRAM than this, you can enable an extra feature to improve processing speed: comment out the line enable_model_cpu_offload() and uncomment to("cuda")
• I do not know what will happen if you try running this script with less than 8 GB of VRAM

File Locations

• models/: Place SDXL models here, each model must be in its own sub-folder
• models/lora/: Place LoRA safe tensor files here
• outputs/: Finished and latent (intermediate step) images will be saved here; output name is seed-random_uuid.png

Model Links

• SDXL Base 1.0: Download all folders and model_index.json; you don't need the 3 big safetensors
• SDXL Refiner 1.0: Download all folders and model_index.json; you don't need the 2 big safetensors

Requirements

python==3.11
pillow
tqdm
torch
diffusers
compel
accelerate
omegaconf
xformers
peft