Textual inversion embeddings tokens are not reliabaly detected

[stduhpf]

The embeddings seem to only be detected when the trigger words are at either the end of the prompt, or just before a comma. I don't think this is consistent with the way it is supposed to work for prompt compatibility with stable-diffusion-webui style.

For example, either -p "image of [trigger_word]" or -p "image of [trigger_word], doing something" will load the embedding, but -p "image of [trigger_word] doing something" won't.