How to make more aggressive profile? #181
-
|
I want to delete all text, even sounds and onomatopoeia. But in the default profile it only cleans text in bubbles. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 3 replies
-
|
Try enabling the inpainter, it's off by default. Also, the text detector has frequent false positives in the raw mask where it thinks random patterns here and there are text, so it does a verification step with the raw box data it generates (the output right after this). Parts of the raw mask that don't fall within boxes where text was actually detected, are deemed to be false positives and thus discarded. So check what actually gets marked by the original boxes output and you'll see what's possible. Disabling this verification does a lot more harm than good (and thus not changeable without modifying the source code), so hopefully with inpainting you'll get the results you want, otherwise that's simply as good as this AI model can get, it isn't perfect. |
Beta Was this translation helpful? Give feedback.
-
|
This helped me. By the way, it would be more convenient if you could choose which parts of the text to keep when confirming. Is there any feature I could turn on for that? |
Beta Was this translation helpful? Give feedback.
-
|
You mean being able to manually disable boxes in the review screen? Or also painting in regions manually? Those aren't available. I think you're better off doing the real review in a real image editor that has proper support for painting over stuff, but if it's only the former, I may consider it. |
Beta Was this translation helpful? Give feedback.
-
@VoxelCubes Adding a way to manually toggle individual detection boxes would be a huge plus. I sometimes run into false positives or stylized 'art text' that I want to keep; being able to enable or disable specific boxes before processing would solve that perfectly! 
|
Beta Was this translation helpful? Give feedback.
Try enabling the inpainter, it's off by default.
Also, the text detector has frequent false positives in the raw mask where it thinks random patterns here and there are text, so it does a verification step with the raw box data it generates (the output right after this). Parts of the raw mask that don't fall within boxes where text was actually detected, are deemed to be false positives and thus discarded. So check what actually gets marked by the original boxes output and you'll see what's possible. Disabling this verification does a lot more harm than good (and thus not changeable without modifying the source code), so hopefully with inpainting you'll get the results you want, otherwise…