TA的每日心情 | 开心 2017-2-4 00:11 |
---|
签到天数: 4 天 [LV.2]偶尔看看I
至尊会员
- 积分
- 164586
|
本帖最后由 dtlnor 于 2022-4-15 15:12 编辑
bilinear毫无疑问是最不用考虑的算法,FSRCNNX应该<=NGU(没记错这个是ai滤镜来着。效果挺好的)。不过oversample似乎与我描述integer scaling有点不一样。我看了看mpv的描述
oversample
A version of nearest neighbour that (naively) oversamples pixels, so that pixels overlapping edges get linearly interpolated instead of rounded. This essentially removes the small imperfections and judder artifacts caused by nearest-neighbour interpolation, in exchange for adding some blur. This filter is good at temporal interpolation, and also known as "smoothmotion" (see --tscale).
实际上我需要的应该就是这个artifacts caused by nearest-neighbour interpolation 来达到integer scaling(四个像素为一组,同一个像素的内容)的目的。mpv这个oversample看起来更像是bilinear的亲戚( |
|