detect Textual Inversion type or get it from the prompt token

[ssube]

The code for Textual Inversions cannot currently detect whether it's an sd-concept from the HF hub or a tensor file with embeddings. The blend code is not fully integrated with the model fetch code, and does its own fetch from HF hub.

Inversions used with the <inversion:name:1.0> token will always be loaded as embeddings, with no way to use a concept-style inversion by token.

[ssube]

The HF concepts do not really need a second file and contain their own token as the only key in the tensor, making it pretty easy to detect them once the inversion code was set up to use the regular model fetch code.

Using an HTTPS or local path source for an inversion will behave normally, but huggingface:// paths will only download the learned_embeds.bin file, like https://huggingface.co/sd-concepts-library/concept-art/blob/main/learned_embeds.bin

[2023-03-20 01:35:23,724] DEBUG: 349112 140247140089856 onnx_web.convert.diffusion.textual_inversion: detected Textual Inversion concept: ['<concept-art>']                               
[2023-03-20 01:35:23,724] DEBUG: 349112 140247140089856 onnx_web.convert.diffusion.textual_inversion: found embeddings for 2 tokens: ['concept-art', '<concept-art>']                     
[2023-03-20 01:35:23,918] INFO: 349112 140247140089856 onnx_web.convert.diffusion.textual_inversion: blending Textual Inversion ../models/inversion/cloud with weight of 0.25 for token cloud