Auto 1111 SDK vs Huggingface Diffusers
The only other type of SDK that exists to run diffusion models is Huggingface Diffusers(https://huggingface.co/docs/diffusers/index). However, there are significant limitations it has that Automatic 1111 solves:
Diffusers takes up considerably more RAM than Auto 1111 SDK.
Diffusers doesn't have all the samplers in Automatic 1111.
Diffusers has small limitations like a 77 prompt length token limit.
Diffusers will Cuda out of memory/perform very slowly for huge generations, like 2048x2048 images, while Auto 1111 SDK won't.
Running Text-to-Image, Image-to-Image, Inpainting, Outpainting, and Stable Diffusion upscale can all be performed with the same pipeline object in Auto 1111 SDK, whereas with Diffusers, you must create a pipeline object instance for each action, severely increasing the memory/RAM used.
Diffusers doesn't have upscaler pipelines to load any Esrgan upscaler weights, nor does it have a dedicated RealEsrgan pipeline
Diffusers doesn't have support for Outpainting, directly downloading models from Civit AI, Prompt Attention, Composable Diffusion.
Diffusers accepts different parameters than the original Automatic 1111 Web UI, rendering it impossible to replicate images generated on Civit AI/Automatic 1111 Web UI on Diffusers.
Extension: Many of Automatic 1111's most powerful features exist as extensions, and Diffusers doesn't support any of these. With Auto 1111 SDK, we will be adding support for these soon, and our goal is to eventually have an 'extensions' hub where developers can publicize any extension they develop. Some examples of extensions are Controlnet Tile Diffusion, Clip Interrogator, and Composable Lora's. You can view a comprehensive list of Automatic 1111 extensions here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Extensions
Last updated