Flux has changed the captioning - how does that affect taggui?

[StableLlama]

Taggui is perfect for captioning in the style of tags (as the name promises).

But (not only) with Flux the way to caption the images has changes and full sentences are required now. This creates two issues I see with taggui:

• commas inside a sentence doesn't separate tags anymore
• Captions can an will span multiple lines

As I want to continue to use taggui for captioning tasks (and especially to run auto captioners) I wonder how this can and will be reflected by taggui.

What is planned already? What are your thoughts?

[jhc13]

I think allowing \n (newline) as a tag separator will solve those problems.

[StableLlama]

Actually I'm not sure about that. I think it's getting more complicated here and we are on a way to the next level were we have to rethink the assumptions that have worked in the past - where everyone had used a list of tags as that was the way it worked. The new models are either pure sentences or hybrid with both.
And when I have training images that I'm tagging I have to think about many different target models now, i.e. I need all styles at the same time during image preparation - but not any more during image training. There I need the format the current model wants to have. E.g. tags for SD1.5 and SDXL and sentences for Flux.

This results in two issues: data storage and user interface.

Data storage

I guess it'll be required to switch to a database solution. Which would be quite bad as it would contradict the light weight approach that taggui has and with is a very strong point of it.

So in my reasoning I thought that it might be good to have a per image text file "database" that contains the tag and captions but only in a way that suits taggui. And on the press of a button taggui would output the usual text file in the style that was selected, i.e. export only tags, export only captions or export both in the order "tag, caption" or "caption, tag" just as selected.
This would also allow to store some meta information which could allow advances features like having timestamps for when a tag/caption was added to allow advanced filtering.

User interface

The current GUI is highly specialized for tags. So this would need an extension to show paragraphs of text.
Hopefully it'll be just some minor size adjustments.
What I have no clue is how filtering should work with captions.

[jhc13]

You can now use newline as a tag separator in v1.31.0.

[yggdrasil75]

generally I use | (pipe) to separate tags and have around 10 different prompts. however, this only affects loras trained on the base model or finetunes that also use nlp. sdxl similarly used nlp instead of tags when it was released, but most finetunes used tagging because it was easier. the only time I see pipe in any conversation is programming related, so I figure that is the best option.