About the checkpoint file.

[suzukimain]

Hello.
I have a rudimentary question,
Could you please tell me the exact names of the checkpoints that can be loaded by from_pretrained and from_single_file, respectively?
I also checked convert_diffusers_to_original_stable_diffusion.py and others, but I couldn't find much.
Thank you.

diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py

Line 1 in 8e7d6c0

# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.

[asomoza]

Hi, you don't say for which model type you're referring. from_pretrained it's for loading multifolder diffusers format checkpoint, so whenever you see a model that has multiple directories inside like vae or text_encoder you can use the from_pretrained. You can also search in the hub for a specific pipeline, for example: https://huggingface.co/models?other=diffusers%3AStableDiffusionPipeline

from_single_file is for loading, as the name says, a single file checkpoint which is the format that the webuis use, an example of this would be any model that you download from civitai. So whenever you encounter a model that only has one unique single file, you can use this method to load it.

[suzukimain]

Sorry. I did not explain well enough.
I would like to know how to describe each file type in the README.
For example, I could write Stable Diffusion checkpoint for the file to be loaded in from_single_file.
On the other hand, I could not figure out how to describe the model format to be read in from_pretrained.

[asomoza]

Both are the same, both are Stable Diffusion checkpoints, one has all the modules inside one file and the other keeps them separated in subdirectories. Depends on the model architecture but for example, for a SD 1.5 both have a text_encoder, a vae and a unet.

As I told you, there's a lot of model architectures so the from_pretrained varies from model to model, you can read what each pipeline requires in the corresponding pipeline class, for example for SDXL is here

[suzukimain]

I see. Thank you for your time.