A simple implementation of the UNet of StableDiffusion

This is a simpler implementation of the UNet of SD than diffusers. Researchers can test their new ideas on this implementation more quickly and conveniently.

It can directly load the SD model config and weights in the diffusers:

from diffusers import StableDiffusionPipeline
from simple_sd_unet.unet import Cond_UNet2DModel

pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
unet = Cond_UNet2DModel(pipe.unet.config)
unet.load_state_dict(pipe.unet.state_dict())

Reference

• https://arxiv.org/abs/2112.10752
• https://github.com/huggingface/diffusers