[GH-ISSUE #159] @_@ #116

New issue

Closed

opened 2026-10-30 21:19:36 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-30 21:19:36 +00:00

Owner

Originally created by @bluelovers on GitHub (Apr 13, 2023).
Original GitHub issue: https://github.com/DominikDoom/a1111-sd-webui-tagcomplete/issues/159

should be @_@

Originally created by @bluelovers on GitHub (Apr 13, 2023). Original GitHub issue: https://github.com/DominikDoom/a1111-sd-webui-tagcomplete/issues/159 ![image](https://user-images.githubusercontent.com/167966/231689000-e46eb3f2-507b-46a3-822f-d7d6bebcb6f8.png) ![image](https://user-images.githubusercontent.com/167966/231689053-3571fab6-f732-439f-aa5f-bab62d528211.png) should be `@_@`

bsakaguchi closed this issue

2026-10-30 21:19:37 +00:00

bsakaguchi commented

2026-10-30 21:19:37 +00:00

Author

Owner

@DominikDoom commented on GitHub (Apr 13, 2023):

Most anime models (at least that I know of) are trained with the underscores replaced with spaces, since CLIP by default is better at encoding natural language rather than "special" tags. This is why the "replace underscores" option is on by default. The issue is that with the data I have from danbooru, there's really no way to separate a tag using underscores to separate words from a tag where underscores are part of the "final product". Booru sites themselves mostly don't replace the underscores with spaces, so they don't face this issue.

However, while it might look odd, these models most likely didn't handle those tags separately from the rest, so they actually never used the version containing underscores in training. So what gets inserted here is actually correct concerning the models. I haven't tried if there's a quality difference though, but for such short tags I imagine it will be very similar either way.
The only exception might be for special Loras that use the underscore version as a trigger word, but that is out of this extensions knowledge.

@DominikDoom commented on GitHub (Apr 13, 2023): Most anime models (at least that I know of) are trained with the underscores replaced with spaces, since CLIP by default is better at encoding natural language rather than "special" tags. This is why the "replace underscores" option is on by default. The issue is that with the data I have from danbooru, there's really no way to separate a tag using underscores to separate words from a tag where underscores are part of the "final product". Booru sites themselves mostly don't replace the underscores with spaces, so they don't face this issue. However, while it might look odd, these models most likely didn't handle those tags separately from the rest, so they actually never used the version containing underscores in training. So what gets inserted here is actually correct concerning the models. I haven't tried if there's a quality difference though, but for such short tags I imagine it will be very similar either way. The only exception might be for special Loras that use the underscore version as a trigger word, but that is out of this extensions knowledge.