We use the 8-bit quantised version of T5-XXL encoder following the tutorial on https://huggingface.co/blog/sd3. This allows inference in less than 16GB of memory.
Before running the scripts, make sure to install the library's training dependencies:
conda env create -f env.yaml
conda activate sd3
Then run
python src/generate_sd3.py
You can run multiple prompts by specifying them on configs/prompts.txt
cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k
Shrek showing a sign that says 'Happy birthday John' to Donkey.
Other configs are specified in configs/sd3.yaml
You can find the generated images at output folder