search

nicolai256 / Stable-textual-inversion_win

MIT License
241 stars 43 forks source link

Broken? embeddings in other forks #29

Open specblades opened 1 year ago

specblades commented 1 year ago

Hi! I trained embeddings with single placeholder and single init word, 11 pictures, 12k iterations and no --init_word in cell. Its going nice. Pictures in train folder are good. txt2img demo in your colab works well, but. When i try to use embeddings in another forks, it wont produce me a similar result. I read a bunch of threads but cant find answer. It rly doing nothing and i dont know why. Can you advise something?

For example in this repo - https://github.com/AUTOMATIC1111/stable-diffusion-webui

UPD. Apparently, if you do not use the line --init_word "XX" in the cell, then embedding will not work in other forks.

GucciFlipFlops1917 commented 1 year ago

I think the lstein repo has a more readily interconvertible output set for textual inversion, not requiring that line you cited.

CodeExplode commented 1 year ago

Both repos appear to save embeddings in the exact same format, with only minor differences like using quotation marks vs apostrophes for string declarations

https://github.com/nicolai256/Stable-textual-inversion_win/blob/main/ldm/modules/embedding_manager.py#L131

https://github.com/invoke-ai/InvokeAI/blob/main/ldm/modules/embedding_manager.py#L210

They both also only override the init_word from the yaml if it's in the command line the same way:

https://github.com/nicolai256/Stable-textual-inversion_win/blob/main/main.py

https://github.com/invoke-ai/InvokeAI/blob/main/main.py#L721

Is it possible that you deleted the initializer words setting in the settings yaml specblades?

https://github.com/nicolai256/Stable-textual-inversion_win/blob/main/configs/stable-diffusion/v1-finetune.yaml#L27