mspronesti / stable-diffusion

Stable Diffusion in PyTorch without any huggingface token!

Stable Diffusion in PyTorch

No Hugginface Token Required

Implementation of Stable Diffusion in PyTorch, for personal interest and learning purpose.

Weights are stored on a huggingface hub repository and automatically downloaded and cached at runtime.

Usage

The easiest way to give it a shot is using one of the following Colab Notebooks:

• GPU Text2Image
• GPU Img2Img

Example output

textual prompt	output
"Italy with cyberpunk style"
"Galileo in Disney style"
"DSLR photograph of an astronut riding a horse"
"Realistic Gandalf playing the piano"
"Eren from Attack on Titan drinking wine"

---

Acknowledgements

The implementation is based on this repository, which is based on these repositories:

• divamgupta/stable-diffusion-tensorflow
• CompVis/stable-diffusion
• huggingface/transformers
• huggingface/diffusers
• AUTOMATIC1111/stable-diffusion-webui

Difference with the original implementation

• the ClipTokenizer has been rewritten to exploit LRU Caching
• the weights and vocabulary download process has entirely automated so that no manual action is required from the user
• add a progress bar to track the generation process and provided a generation script
• the code has been