I'm not well versed at all in using things like Photoshop or GIMP, so just putting that out there. I figured this would be a good place to ask. I apologize in advance if I say anything dumb.

Anyway, I've tried searching which interpolation may be the best for this situation but I'm just not finding anything that leads me to believe a certain interpolation would be best. I'm downscaling images in GIMP to really small sizes (102x96 to be exact), changing them to 256 colors, and I'm having trouble deciding which interpolation would be best for this size. I'm downscaling images for a retro game that I'm modifying. As you would guess, at 102x96 the images get quite pixelated. I'm not really trying to retain the most pixelated look possible, because I think most images in the game have a slight blurred look to them. I guess I want to know what the best interpolation for the image size that I'm downscaling to is? It's just that I'm driving myself crazy comparing the 4 interpolation options GIMP has (Cubic, Sinc, Linear, and Nearest Neighbor). I keep finding that there's different things about the images that I don't like, and I can't seem to find a consistent interpolation that I like without second guessing myself.

Also, let me note that I have looked at Difference between None, Linear, Cubic and Sinc(Lanczos3) interpolation in image scaling?, but in my situation I still can't seem to come to a decision. Is it really a case-by-case basis and I can't settle on using any singular interpolation? If I'm downscaling to this size and changing it to 256 colors, what interpolation do you guys think is best at retaining details in the image?

I'm terribly sorry if this post is a mess. Again, I'm not knowledgeable at all in this arena, despite researching a lot about this stuff.

  • What size are you downscaling from?
    – Confused
    Commented Jan 22, 2017 at 13:40
  • @Confused Most images I'm downscaling from are 544x544. Some are between that and around 1000x1000, but most are 544x544.
    – Sutiko
    Commented Jan 22, 2017 at 23:37

1 Answer 1


After some test I can give the following advice:

Downscaling erases details and spoils smooth color gradients. How much of details and gradients - bicubic saves gradients better but hides more details, nearest neighbour saves sharp edges better, but destroys gradients. Which to use - that depends on what you consider important.

My opinion: Sharp edges are wanted in small game images. Use nearest neighbour algorithm or a sharpened version of bicubic.

I have a furthergoing suggestion. Use some cartoonizer or other filter that simplifies the image by replacing smooth areas by larger uniform colored ones. The simplifying can be done before downscaling, after smoothness savvy downscaling or both, but before the amount of colors is reduced to 256. The simplifying radically enhances the recognizability of the downscaled and reduced color count image. Top of this filter leaque is Topaz Simplify for Photoshop. See a comparison. The first is not simplified

enter image description here

Now a couple of different simplified versions:

enter image description here enter image description here

Note: Using "Posterize" is not a replacement for a proper simplifying filter, it makes the result even worse. But some into your program already built artistic filter probably do the job, if proper parameter values are set.

Here's also the original as about 1 megapixels

enter image description here

  • Thanks for the answer! It's funny that you say either Cubic or Nearest Neighbor because I was probably split between these two the most. I may just stick with Cubic though, since sometimes when I use Nearest Neighbor I feel like the images are too rough looking. (For lack of a better term)
    – Sutiko
    Commented Jan 22, 2017 at 23:31
  • I'll definitely try the cartoon-esque filters and see what results I get! That's an interesting suggestion but perhaps that'll work for me.
    – Sutiko
    Commented Jan 22, 2017 at 23:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.