22/01/25 04:59:18.65 QDq8f09f.net
スナイアーズ博士による-E optionの解説あった
URLﾘﾝｸ(encode.su)
>Each channel separately does mean it's missing opportunities to use previous channels as context for the current one, e.g. for a 10-channel image you could give it as a 10-channel image to cjxl -q 100 -e 9 -E 9 and the -E option means it will take up to 9 previous channels as potential Extra properties when learning the MA tree context model.
Also color decorrelation could be useful - RCTs can be applied to arbitrary channels, though the current encoder will only try that on the first 3 channels (since it assumes those are RGB or CMY and thus quite correlated).