Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
GIMP-ML
#1
Check out these awesome machine learning plugins for GIMP 2.10
https://github.com/kritiksoman/GIMP-ML
Reply
#2
Awesome, as an adjective extremely impressive or daunting; inspiring awe maybe not.

I installed this about 10 days ago, so these comments are as a very new user. Anyone intending to use this software should check out the license terms on the github page. There are useage restrictions on some of the filters.

Using kubuntu 20.04 / Gimp 2.10.22 (appimage) and my old desktop (not very, but quick enough) computer.
Installation. This is a lot of plugin. 1.5 GB as installed. The good thing is it is self-contained and seems portable. I moved the installation folder out of my home partition onto a storage partition.

Try-outs on some of the filters:

Superresolution to upscale x4 : This a comparison with Gimp scale (nohalo) and superresolution. It is a bit sharper, I did see some strange edge artifacts on a different image. Not that much different from a regular upscale + some sharpening.  

   

Deepmatting: Not too much difference between Gimp FG extract and ML. ML has the edge, it uses a mask layer, easy to adjust and re-apply the filter.

   

Deepcolor: This very like the gimp_gmic_qt interactive colorize. Again uses a mask layer so easy to adjust and re-apply. Quick example.

   

Some of the others: Deblur, tried it on this one https://www.gimp-forum.net/Thread-Proces...xt-Legible and truthfully no better than the other suggestions  ...and others, might not be their intention but make quite good layer masks.
Reply
#3
Also: https://www.gimp-forum.net/Thread-Some-p...g-for-Gimp
Reply
#4
@rich2005 Thanks for the review. Few points :
-> GIMP FG select tool can be used to create trimap mask for deepmatting.
-> Deepcolor plugin can be used without layer mask as well. Original image and color mask should be the same greyscale image layer for this.
-> Super-resolution plugin results might not appear significantly different than NoHalo + wavelet-decompose on the particular example you showed. But the network that it uses is in-fact one of the state of the art (https://arxiv.org/abs/1609.04802) methods.
Please share any suggestions if you have any.
Thanks

@Ofnuts
-> Similarly, deblur plugin (https://arxiv.org/abs/1908.03826) is also one of the state of the art methods although it didn't work well on the particular image.
Reply


Forum Jump: