add module : scene-referred blurs #9521
Conversation
Add support for motion, lens and gaussian blurs in linear RGB.
|
Any chance of running this in reverse to undo motion blur? You have what looks like a nice interface to set up the point spread function used by most (all?) deconvolution algorithms. |
Wow, definitely not. Deconvolution is a whole different game than this, and the methods that actually work ok are super slow and a lot more work. |
|
Howeverrrr if one were to optimize it (like diffuse got optimized ;) ) this are the words of the Aurthor:
Just a note for optimizers ;) |
There was a problem hiding this comment.
Some minor comments.
You also need to update po/POTFILE.in
And as for diffuse you need to add this new module into some of the module group presets.
See src/libs/modulegroups.c where presets are defined.
You need to at least add it in "module: all" preset and maybe some others.
Can you also do that for diffuse ?
TIA.
|
@TurboGit I will deal with the module presets in a separate PR |
| @@ -186,6 +186,7 @@ src/iop/basicadj.c | |||
| src/iop/bilat.c | |||
| src/iop/bilateral.cc | |||
| src/iop/bloom.c | |||
| src/iop/blurs.c | |||
There was a problem hiding this comment.
I'll do it just after merging, but for reference you always have two entries to add, one for the module itself and one for the generated introspection code.
|
@aurelienpierre I just tested it and it works great, but I wonder why the radius is limited to 128 px? For such a scene.... original:
blur:
...I had to use three instances. But in itself, with new blurs module you can do cool things :) |
|
A radius of 128 px means a full blur of 256 px, which means at each pixel, you need to compute 256×256 multiplications and additions. Given that the module is not fully optimized right now, that's already insanely slow. |
elstoc
added
documentation-complete
Add support for physically-accurate motion, lens and gaussian blurs in linear RGB.
This is slow code for now, doing convolution products in spatial space, CPU and OpenCL variants provided. A faster convolution approach has been tested (commented out in the code), using FFT, but crashes with no explanation, and I don't have time to investigate that now. That shall be investigated later, the current code is simple and functional.
Warning : to test, in lens blur mode, to get proper stars, you need to ensure
concavity < blades - 2. If that condition is not met, you get degenerated shape (which can have merits too, but is not what you expect).Example pics by @johnny-bit :
Original:
