Safetensors loading in "convert_diffusers_to_original_stable_diffusion"

[cafeai]

What does this PR do?

The convert_diffusers_to_original_stable_diffusion.py script currently only supports loading SD diffusers models in pytorch format. An update has been made to check to see if safetensors versions of the files exist before deserialization takes place. If safetensors versions exist, these are deserialized. If not, the pytorch version is attempted to be deserialized.

Usage Changes

No changes to the command line options are made. The detection happens behind the scenes.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[patrickvonplaten]

That is nice, it means the script automatically detects if the format is safetensors or not?

Note that we have an argument --from_safetensors should we maybe deprecate this one then?

[cafeai]

Yes, the script automatically detects if the diffusers folders contain safetensors or not, and will load accordingly.

I suppose from a security point of view, you could make the argument that it may be useful to have a command line option that will force the script to fail if the version is not safetensors, instead of loading it automatically.

I can refactor so that it falls in line with --from_safetensors, as there is an added security benefit.

Another option is to add an argument like --only-allow-safe-loading where the script fails if and only if a safetensors format isn't detected to be loaded (or a potential future "safe" format). That way people who don't care can just have the script load any file type, without having to know what the original format is, or enforce safetensors if they so desire.

Thanks for your time!

[patrickvonplaten]

Actually think this is very nice as is - let's merge it :-)