Gimp: How to optimize the resize of images

Hello folks,

I want to show you how to resize images for web presentation without losing to much sharpness. The original image is the tower bridge in London. What i want to demonstrate is, that images loose a great deal of detail and sharpness when carelessly resized in GIMP. I don’t know if this applies to other image manipulation programs to. The results are even more drastic if you take an even bigger base image.

The image has a resolution of 1280×1024 and shall be resized down to 450×360. If you open the image in gimp and scale it directly down to the target size you get the following output.

The resize algorithm used in this example is Sinc (Lanczos3). The result looks very pixeled and distorted.

If you use the cubic algorithm the result is kind of blurred and the details washed out a bit.

Whenever I want to resize my 10-12MP images I make with my 40D for web publishing I do the following.

1. I always resize using the Sinc (Lanczos3) algorithm but do it in several steps.
2. I resize the image by 50%
3. Repeat this step (usually 1-3 times) until the resolution of the next step would be below target image size. In this example 1280×1024 resized by 50% will be 640×512. If i would resize this image again by 50% the result would be 320×256 which is already smaller than 450×360 so I don’t repeat the step.
4. This time the image will be resized (from 640×512) to the target size 450×360 pixels.

As you can see, the result is something in between the cubic and the Sinc results. It doesn’t show the distorted dots like in the first image and has more detail than the second image. The result will be even better if you apply a very slight USM (unsharpen mask) to the original image before you start resizing. A sharpening after the resizing depends on your image.

I did the same steps to the following image of model Wendla:

Sinc (Lanczos3) resize: Pay close attention to the eye lashes, the lipps and the hair around her right ear.

Cubic resize

Procedural resize

The differences might be minimal and if you can’t see them you probably shouldn’t care. Click the next image to enlarge it and see a 200% view of it.

Left: Sinc(Lanczos3)
Middle: Cubic
Right: described method