Anaconda Environment Setup

Create and activate an Anaconda environment using

conda create -n SDXL python=3.11
conda activate SDXL

Install PyTorch using:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126

Newer CUDA versions are available on the PyTorch website. Note that a GPU is a must for reasonable inference time. Inference using SDXL requires around 15 GB of VRAM. You can also use Google Colab.

Install other packages:

pip install -U accelerate diffusers transformers
pip install pandas

Inference

To generate 300 images using 300 prompts, run:

bash run.sh

The images will be saved in ./inf/. The prompts are written in metadata.csv. The fine-tuned UNet for SDXL kumo24/sdxl_nuclear is available on HuggingFace.

Note on Reproducibility

To get a fully reproducible pipeline, you should use deterministic algorithms as explained here. However, deterministic algorithms may be slower than nondeterministic ones, and you may observe a decrease in performance. Therefore, you should expect different images on different GPU hardware or different pipeline initializations.