huggingface/diffusers vs AUTOMATIC1111/stable-diffusion-webui comparison

[camenduru]

https://www.youtube.com/watch?v=VfjITfEXRNY

probably huggingface/diffusers has some bug idk, do you have any idea?

comparison codes

!pip install -U -qq diffusers transformers
import torch
from diffusers import StableDiffusionPipeline

generator = torch.cuda.manual_seed("4609990744076363")
pipe = StableDiffusionPipeline.from_pretrained("nitrosocke/mo-di-diffusion", safety_checker=None).to("cuda")
image = pipe("donald duck", height=512, width=512, num_inference_steps=50, guidance_scale=7.5).images[0]
print(torch.cuda.initial_seed())
display(image)

!git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
%cd /kaggle/working/stable-diffusion-webui

!sed -i -e 's/numpy==1.23.3/numpy==1.21.6/g' requirements_versions.txt
!sed -i -e 's/fairscale==0.4.9/fairscale==0.4.6/g' requirements_versions.txt

!wget https://huggingface.co/nitrosocke/mo-di-diffusion/resolve/main/moDi-v1-pruned.ckpt -O /kaggle/working/stable-diffusion-webui/models/Stable-diffusion/moDi-v1-pruned.ckpt

!python launch.py --share

[patrickvonplaten]

You can speed up the Hugging Face implementation by adding torch_dtype=torch.float16 to from_pretrained(...) as is shown here: https://huggingface.co/nitrosocke/mo-di-diffusion#%F0%9F%A7%A8-diffusers -> this should speed up the inference quite significantly :-)