jxl support

[yggdrasil75]

jxl are faster decompression than png, and much smaller file sizes while being lossless, using that would benefit quality over jpg and speed over png. converting would be useful. alternatively, for similar sizes as the png while still lossless, qoi is significantly faster. this could allow reading from disk instead of keeping the images in memory with less speed impact.

at the bare minimum, it would speed up the initial startup time and caching process while saving space on the drive.

[without-ordinary]

OpenCV should support jxl if its a "recent" version. And PIL can support jxl with pillow-jpegxl-plugin and import pillow_jxl.

Another trainer I've been using does not explicitly have jxl support, but between not have a very old version of opencv and importing the pil jxl plugin, and adding ".jxl" to the list of image files for it to look for, its been working. I imagine hacking it support for it in kohya_ss would be similar.

Edit: Looks like upstream added/fixed jpeg-xl support last year. May already work. 🤔