Lossless Float Image Compression

Summary

Back in 2021 I looked at OpenEXR lossless compression options (and I think my findings led a change of the default zip compression level, as well as change of the compression library from zlib to libdeflate. Yay blogging about things!). Then in 2023 I looked at losslessly compressing a bunch of floating point data, some of which might be image-shaped. Well, now a discussion somewhere else has nerd-sniped me to look into lossless compression of floating point images, and especially the ones that might have more than just RGB(A) color channels. Read on! Four bullet point summary, if you’re in a hurry: Keep on using OpenEXR with ZIP compression. Soon OpenEXR might add HTJ2K compression; that compresses slightly better but is worse compression and decompression performance, so YMMV. JPEG-XL is not competitive with OpenEXR in this area today. You can cook up a “custom image compression” that seems to be better than all of EXR, EXR HTJ2K and JPEG-XL, while also being way faster. My use case and the data set What I wanted to primarily look at, are “multi-layer” images that would be used for film composition workflows. In such an image, a single pixel does not have just the typical RGB (and possibly alpha) channels, but might have more. Ambient occlusion, direct lighting, indirect lighting, depth, normal, velocity, object ID, material ID, and so on. And the data itself is almost always floating point values (either FP16 or FP32); sometimes with different precision for different channels within the same image. There does not seem to be a readily available “standard image set” like that to test things on, so I grabbed some that I could find, and some I have rendered myself out of various Blender splash screen files. Here’s the 10 data files I’m testing on (total uncompressed pixel size: 3122MB): OpenEXR OpenEXR is an image file format that has existed since 1999, and is primarily used within film, vfx and game industries. It has several lossless compression modes (see my prev...