tokenspace / README.sd-embeddings.md
ppbrown's picture
Update README.sd-embeddings.md
5b4c792 verified

Summary of Stable Diffusion embedding format

This file is to be a quick reference for SD embedding file formats.

Note: there are a bunch of files here that have "embedding" in their names. However, they cannot be used as Stable Diffusion Embeddings.

I do include some tools, such as generate-embedding.py and generate-embeddingXL.py, that are intended to explore the actual inference tool formatted embedding file types. Therefore, I'm taking some time to document the little I know about the format of those files

Stable Diffusion v1.5

Note that SD 1.5 has a different format for embeddings than SDXL. And within SD 1.5, there are two different formats

SD 1.5 pickletensor embed format

I have observed that .pt embeddings have a dict-of-dicts type format. It looks something like this:

[
"string_to_token": {'doesntmatter': 265}, # I dont know why 265, but it usually is
"string_to_param": {'doesntmatter': tensor([][768])},
"name": *string*,
"step": *string*,
"sd_checkpoint": *string*,
"sd_checkpoint_name": *string*
]

(Note that string can be None)

SD 1.5 safetensor embed format

The ones I have seen, have a much simpler format. It is a trivial format compared to SD 1.5:

{ "emb_params": Tensor([][768])}

According to https://github.com/Stability-AI/ModelSpec?tab=readme-ov-file there is supposed to be metadata embedding in the safetensor format, but I havent found a clean way to read it yet. Expected standard slots for metadata info are:

"modelspec.title": "(name for this embedding)",
"modelspec.architecture": "stable-diffusion-v1/textual-inversion", 
"modelspec.thumbnail": "(data:image/jpeg;base64,/9jxxxxxxxxx)"

SDXL embed format (safetensor)

This has an actual spec at: https://huggingface.co/docs/diffusers/using-diffusers/textual_inversion_inference But it's pretty simple.

summary:

{
"clip_l": Tensor([][768]),
"clip_g": Tensor([][1280])
}