[Feature Request] Standalone version

[ruSauron]

Is it possible to make a separate launcher that doesn't require automatic1111 webui to run? So that even when webui updates and breaks everything, you can still use your script. May be without interrogators if its hard to implement them. So you could run a separate .py script without making a big load of video memory from the whole webui. Or better yet, make possible to use this extension as a utility for bmaltais/kohya_ss training script

[toshiaki1729]

Since this extension is mostly independent from webUI scripts, it is not too difficult to make it standalone as long as using gradio.

The dependent features are as follows;

• independent from webUI main feature (these can be implemented)
  • BLIP, DeepDanbooru
  • CLIP (raw text tokenizer)
• dependent on future webUI update (impossible to support permanently)
  • hijacked CLIP tokenizer (with embedding, emphasis, …)

However, actually I want to use something other than gradio (like streamlit) if I make the standalone version. This is because gradio cannot accept custom component (component is something like button or textbox) without changing gradio library folder itself (this makes distribution of custom component difficult). Custom component is an easy way to add more user-friendly UIs. I need this because I don't have any skill to make UI working on web browser (and it seems really tough even if I had the skill).

[toshiaki1729]

As you say, there are many mysterious bugs (like #66 #67), so I would like to create a stand-alone version as soon as possible. (issue 66 turned out to be caused by some browser extension)

I'm going to make gradio version first (with gradio v3.22.1, based on tag-editor v0.1.2), and then move on to streamlit version.

[ShaguarWKL]

with a better UI than 1111 UI. This extension is so great for pre-training prep but 1111 UI drives me nuts. Hope this happens in the future.

[Kadah]

Standalone version would make sense. Not seeing what benefit, other than maybe ease of install, being an extension of A1111 has. Would be nice to have the full window for the UI.

[ruSauron]

fake "standalone version" (actually it's fresh automatic1111 installation), you need 10-11 gb on your hdd/ssd. Copy the text below into notepad and save as install.bat in the folder where you want to have a separate version of dataset-tag-editor. The installation will take about 5-30 minutes (depends on your download speed, CPU and SSD) Next time use the file dataset-tag-editor\dataset.bat to start it (startup time 15-120s) I chose port 7862 because 7860 is the standard webui port and I usually run bmaltais/kohya_ss on 7681

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git dataset-tag-editor
cd "dataset-tag-editor"
git reset --hard 2f01814
cd extensions
git clone -b v0.1.2 https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editor.git
cd ..
echo title dataset-tag-editor > dataset.bat
echo set COMMANDLINE_ARGS= --lowvram --skip-version-check --autolaunch --theme dark --port 7862 >> dataset.bat
echo call webui.bat >> dataset.bat
echo httpx==0.24.1 >> requirements_versions.txt
echo torchmetrics==0.11.4 >> requirements_versions.txt
call dataset.bat

And copy any SD 1.4-1.5 ckpt or safetensors checkpoint to dataset-tag-editor\models\Stable-diffusion

[IkariDevGIT]

please a standalone version! @toshiaki1729

[toshiaki1729]

Thank you for your patience! Still WIP and not tested well but released the Standalone Version https://github.com/toshiaki1729/dataset-tag-editor-standalone

Here is difference from this Extension Version (Pros)

• Avoid bugs caused by specific version of gradio
• Much faster startup and running
• Added some captioning interrogators (GIT (paper (arXiv)), BLIP-2 (paper (arXiv)): BLIP-2 may extremely heavy), which is mentioned in #33

(Cons)

• Hijacked CLIP tokenizer is not available (which has SDwebUI specific features like embedding, emphasis, …)
• No more support for old Python < 3.9