JPEG-XL allows the use of client-side synthesized grain. A method pioneered by Netflix/AV1 I believe. Compression algorithms struggle with high frequency detail, which often introduce visible artifacts. JPEG-XL allows to decouple the grain component from the actual image data. This allows for significantly more efficient compression of images that inherently require noise, such as those in gnome-backgrounds — smooth gradients that would otherwise be susceptible to color banding.
So basically dithering is on its own separate channel?