How to prevent Color Banding with GIMP?

Just doing a simple bump_map operation in python GIMP. But the result has color banding!

import os, glob, sys, time
from gimpfu import *

image = pdb.gimp_file_load(img_path, img_path, run_mode=RUN_NONINTERACTIVE)
gray = pdb.gimp_file_load(gray_path, gray_path, run_mode=RUN_NONINTERACTIVE)

pdb.plug_in_bump_map(image, image.active_layer, gray.active_layer, 
    135.0, 20.00, 30, 0, 0, 0.0, 0.0, 1, 0, 0)
pdb.file_png_save_defaults(image, image.active_layer, out_path, out_path)

If I remove the "plug_in_bump_map" operation, there is no color banding in my final result saved. If I apply a gaussian blur to the grayscale image, the color banding disappears in the final result, but leaves an undesirable blurriness, what should I do? All images are PNG.

OK, so it's banding :)

But if you look closely, the banding is already in your bump map, and is sort of unavoidable since you only have 256 levels in your PNG.

Banding occurs when there is a significant change in value between two uniform areas with a somewhat linear border between them. So the two mitigations are

• spreading the change around the border:
  • higher precision image, so the changes can be smaller than 1/256
  • blurring the image
• making the border non linear to the eye (spread noise)
• dithering, which does a bit of both, since it introduces intermediate values, but places them at random.

The best way to fix the problem is at the source, so if openCV can compute on more that 256 levels, have it do so and export the bump map in some high-bit depth format (16-bit/channel PNG could be enough, otherwise TIFF).

Otherwise if you have to start with an 8 bit bump map:

1. Load in Gimp
2. Convert to high-precision (32bit FP linear)
3. If you look at the histogram at that point you still have only 256 values:

4. You then add "dithering" with the RGB noise filter to spread the values to fill the gaps. This is best done with the Histogram dialog up, and cranking up the values until the "hair comb" structure nearly disappears:

5. You can also use Spread noise to "shake the pixels" a bit (2-3px), so that transitions are less linear and so less visible:

6. You can then apply the bump mapping. You trade banding for a bit of graininess:

7. You can reduce the graininess using Median blur without having to use values that make your image lose too much sharpness.

It is also possible to replace steps 4 & 5 by a selective Gaussian blur.

Edit: tried your 16bpc:

1. Promote to 32bit FP linear
2. Apply bump-map without pre-processing

"native" result is already much better:

It can be improved with:

1. color selection of the "flats" color with a threshold of 0
2. Selective Gaussian blur