The best Halloween.

The best Nanakusa-chan.

內封繁體官字。

感謝 Rakka-Aria 代爲發佈。

slow.pics compare: https://slow.pics/c/L2oVDl3c.

The source is bad, and we couldn't find a good way to save it, so we opt for a rather conservative filtering, minimising damage* to the image while doing as much as we can.
We first use a protective*, albeit strong vsdenoise.wnnm to denoise, followed by a decent vsTAAmbk.TAAmbk for aa, and a rather conservative* vsdehalo.dehalo_sigma for dehalo.
With masking and other methods, we try to avoid damaging the very stylish noise in OP as much as we can. The G and B shifting in this anime is very interesting, and we try to make it look nicer.

*: Not really.

For encoding, this anime is really demanding. We discard the normal method of metric-based boosting, because that way the ED alone will take up almost half of the filesize for the entire episode, while the ED itself doesn't have a good quality to begin with.
This anime also has very weak lineart in character's hair, which is always extremely challenging for SVT-AV1 derived encoders.
To deliver the best result we can, we use both --crf based and ROI map based character boosting. In episode 01, we have 113 scenes out of a total of 425 scenes with base Q index starting at Q 112, and we have 144 scenes with base Q index for characters starting at the equivalent of Q 26 in the keyframe.
We use 5fish's SVT-AV1-PSY v2.3.0-C --tune 3 to encode. To protect the weak character lineart as much as possible, we use --max-32-tx-size 1 --complex-hvs 1 --psy-rd [3.3 ~ 4.0] --spy-rd 0.
This anime features a lot of night scenes. To make sure the night scenes will look good as well, we use --frame-luma-bias 28.

This is the bitrate of the final encode as an excuse, and the metric result of the encode with compared to the filtered source:

Thanks to all the people creating the tools, plugins and vsfunc's that make this possible, and thanks to everyone in the community for teaching me, pointing out issues, and giving me advice.