Please add support to upload textual inversion embedding

[abdbarho]

Please add support to upload textual inversion embedding

[AbdBarho]

Thanks @fpgaminer I managed to get it up and running with the correct cli args, I also updated the Readme.

@sausax Can you try from again with version 0.1.0?

[AbdBarho]

I will give it a look today after work!

[AbdBarho]

@sausax Unfortunately, the UI in its latest version does not seems to provide a file upload for the pre-made embedding, unlike what is mentioned here, or at least I could not find it, therefore I cannot implement / test this feature.

if you know where I can find it that would be great! otherwise I would suggest you get in contact with author of the library, either here https://github.com/hlky/sd-enable-textual-inversion or here https://github.com/hlky/stable-diffusion

[sausax]

This is weird. The code is present here in the original repo: https://github.com/hlky/stable-diffusion-webui/blob/master/frontend/frontend.py#L99

But it is not coming up in the ui by the docker container.

[AbdBarho]

The model needs to have the an embedding_manager attribute, which seems to be done here, However, the LatentDiffusion class is NOT instantiated in the WebUI.

[fpgaminer]

https://github.com/hlky/sd-enable-textual-inversion/blob/08f9b5046552d17cf7327b30a98410222741b070/configs/stable-diffusion/v1-inference.yaml#L20

The sd-enable repo also has a modified config, which includes a personalization_config key. Does that perhaps get the WebUI to change its behavior?

[AbdBarho]

@fpgaminer it might be that cp -rf that I used is not enough, I will dig deeper!

[AbdBarho]

I have made sure that the the new configuration is included. In the attached log you can see that the personalization_config is part of main repo.

log.txt

Yet still, I can't see anything new in the UI

This is the code I used in the dockerfile to copy the stuff over, feel fee to experiment with it

RUN apt-get update && apt install rsync -y && apt-get clean
RUN git clone https://github.com/hlky/sd-enable-textual-inversion.git && cd sd-enable-textual-inversion && \
git reset --hard 08f9b5046552d17cf7327b30a98410222741b070
RUN rsync -a /sd-enable-textual-inversion/ /stable-diffusion/
RUN cat /stable-diffusion/configs/stable-diffusion/v1-inference.yaml

[fpgaminer]

Remove the --optimized and --optimized-turbo flags and the UI will show:

EDIT: The reason for this is because webui.py loads the optimized config and optimized models with those flags. I guess the textual tweaks were only implemented in the stock models for now. They'd need to be ported over. But yeah, for now, copy everything over like the repo says, drop the optimized flags, and it looks like it'll work (I'm training an embedding right now so can't directly test).

[fpgaminer]

Okay, got it fully working. Also need to add the --no-half flag, otherwise it errors out during generation. Tested using my own embedding. FYI, textual inversion doesn't work great for SD right now anyway. I think they have it tweaked for plain Latent Diffusion, not the specific Stable Diffusion model. So we'll need to wait for another update there to get better results.

[sausax]

@fpgaminer can you create a PR with the fix ?

[fpgaminer]

I'll be a bit busy so probably not. Sounds like @AbdBarho already has a working Dockerfile update; just needs the README updated to mention removing those flags to get embeddings working.

[sausax]

@AbdBarho Getting a new error when trying to build the latest master

$ docker-compose build
Building model
Sending build context to Docker daemon 9.216kB
Error response from daemon: dockerfile parse error line 33: unknown instruction: GIT
ERROR: Service 'model' failed to build : Build failed

[AbdBarho]

@sausax can you try again with docker compose instead of docker-compose? one of them is v1 the other one is v2.

I am using Heredoc syntax which is modern feature, you need a relatively modern docker version (from the last half year or so)

[sausax]

That worked :) Thanks @AbdBarho