How do you handle edges on textures that join adjacent textures?

Question

This question explains how to use borders to avoid texture bleeding. The following image is taken from the excellent answer to that question.

But the problem I don't understand is, imagine you have an 800x800 texture, and to alleviate the colour bleeding you use a one pixel border to the right. Now when the sampler samples from the last pixel it can blend between the second last pixel and the last pixel (the padding border you put).

However now imagine at a lower mip level. Now picture the image (or atlas, or whatever you want to imagine), instead of being 800 pixels across, is only 4 pixels across. Now the last pixel is located at U value is located at the centre of the fourth pixel, 0.5 + 0.125 + (0.125 /2). Anyway, U value after that (the centre of the fourth pixel) is going to want to be blended with an equally large pixel to the right. That one or two pixel border you put or however big really to the right isn't going to do anything because it was only 2 pixels in an 800 pixel image, and meanwhile in the 4x4 mipmap each pixel covers the equivalent of 200 pixels in the original image.

Please explain.

Answer 1

You've accurately identified the reason why it's better to store textures like this in a texture array / array texture, where supported, rather than a texture atlas.

In a texture array, each image wraps and mips independently, so samples near the border of one never end up pulling in data from a different image, and you don't need any borders or padding at all.

When you don't have texture arrays available, or you need to pack images of different dimensions, then about the best you can do is add more padding. Take the deepest mip level you want to be able to address, subtract one, raise two to the resulting power, and that's how much padding you need to ensure that even at the deepest mip, your samples never blend between two distinct images.

That can get excessive, so you can also accept some bleeding at deep mips, on the assumption that these objects must be very far or viewed very obliquely for that mip to even be sampled, and so the bleeding will likely be less noticeable.