Compositing tricks to improve image quality, your tricks

(Ace Dragon) #1

And I’m not talking about the render at twice the size and scale it down thing.

Here’s one.

Say you rendered a .jpg, uh oh, a bunch of tiny little artifacts and some banding, what do you do?

In my newest clouds image I was faced with the problem of small .jpg artifacts when looking close and some banding, so this is what I did in the compositor

1). Bring in the image
2). Scale it up to 4500x4500 pixels
3). Gaussian blurred it with the values at 12
4). Scaled it back down
5). Took the scaled up image, drew it through another compositor branch an edge filter and a colorband and joined that with the former by plugging it into the size input for the blur.

I found from some testing that it made the banding look a little better and wiped out the tiny artifacts while giving very little loss of sharpness. The effect was most noticable under high contrasts and after I tested edge enhancing in a paint program, in the initial image, the edges shown were blocks, after the compositing was done, it was a fine grain.

From 100% scale it looks almost identical, but zoom in closely and you’ll need a sort of very fine grain where little artifacts used to be in areas with little contrast. The advantage of this is that rendering it at twice the size is not needed, you can render it at the origional size then use the trick to do what I described above.

EDIT: Found my trick also reduced sharpness on most images, no worries, I simply added another step to keep as much sharpness as possible while getting rid of artifacts.

So you found any tricks to improve quality without rendering at twice the size?


(PapaSmurf) #2

Using RGB curve to adjust the contrast and range of the colors in an image is a biggie, simulating the use of different colored light in the original. Making the sky bluer for example, if the shot was a hazy day.


(harkyman) #3

Here’s something I do (in Photoshop) as a pretty standard image process to give a more “realistic” effect. And by “realistic” I mean “what we’re used to seeing from current digital imaging technologies.”

  1. Apply 1% noise to entire image (Gaussian, color).
  2. Apply Gaussian blur at 0.5%
  3. Standard Sharpen filter. Keep doing it until you see it’s gone one too far, then undo that.
  4. Apply 0.5% noise (Gaussian, color).

What it does is try to emulate the CCD capture and processing regimen. It’s not called for with every image, but something that you want to look like it was recorded, not seen with the eyes, can benefit from the subtle shifts this will give it.


(ChojinDSL) #4

It would be cool if some of you could post some “before” and “after” images which highlight the effect your describing.


(buddamint) #5


upscaling and downscaling the image through fiters = antialiasing. if you have the blocky edges and not alot of fine detail you want to keep, it should be enough to simple double the image size and then resize it to its original size again. the best filters are; mitchell for upscaling (making larger) and lanczos/sinc for downscaling.

to deband an image;
-apply a bit of blur (for PAL/NTSC I recommend 8 pixels)
-apply a mix operator to mix between the blurred and original, around 30-70% should be best
-apply noise (as posted above) to the image (again with a mix operator).

what this does; the blur smoothes the gradients (the ‘plateaus’ of banding in the image), and the noise operator varies the pixel values beyond the limited bit depth of the original image-- which effectively raises the bit depth of the image.