[Don't merge] PR to show that diffusers and SGM gives 1-to-1 the same result

[patrickvonplaten]

This PR was opened only to show that the default settings of SGM gives 1-to-1 the same results as the branch sd_xl (soon in "main"): huggingface/diffusers#3859

You can verify this as follows:

git clone https://github.com/patrickvonplaten/generative-models.git
git checkout same_result_diff

Then make sure to add the SD-XL checkpoint to the checkpoints folder under "./generative-models/checkpoints/sd_xl_base_0.9.safetensors". Then:

cd generative-models && python scripts/demo/sampling.py

This should generate an image that looks more or less like the following (results might vary depending on the GPU). The following image was created on a RTX4090:

Now you should get 1-to-1 the same result with diffusers. Simply do:

pip install git+https://github.com/huggingface/diffusers@sd_xl

and then run:

#!/usr/bin/env python3
from diffusers import DiffusionPipeline
from pytorch_lightning import seed_everything
import torch

pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
pipe.to("cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
seed_everything(0)
image = pipe(prompt=prompt).images[0]
image.save("/path/to/save")

which should give an image of the following:

Note that when downloading the images you can see that they are not 100% identical but for the human eye one cannot make a difference.

[benjaminaubin]

Thank you for the great comparison. I am closing this PR since it is not intended to be merged.