[Yoshi]

Although the following quotes are from years ago, since they are still up do date and worth to be discussed, I want to bring up the following again, especially based on what Madshi posted at that time.

Originally Posted by madshi

@Mark, all the resampling filters are a compromise between aliasing, sharpness and ringing. There is no filter which is best in every category.

Considering that video signals are essentially the same as audio signals, I wonder why in practice it seems to be such a challenge to get it right in the video domain. I might be corrected but I think when it comes to audio, taking any voodoo/audiophile arguments aside, filters which leave the usable bandwidth quite intact while suppressing any other range which could cause aliasing are rather easily achievable using cheap and decent technics nowadays.

Originally Posted by madshi

Do you want higher frequency response and live with ugly jaggies? Or do you want smooth jaggies and can live with a lower frequency response?

So what is the reason that already when it comes to low-pass-filtering of images it seems to be such a trade-off between sharpness and aliasing to be made still in 2017?

For anyone considering my comparison between video and audio far-fetched and "unorthodox": both are the same phenomenon at least in their waveform appearance - amplitude change over time. So any video signal, be it 4K in BT2020 or whatever nowadays fancy color space could be represented by a pure analog waveform to be digitized. Hence I wonder why a downscale is still such a challenge. Couldn't one apply the same already proven tools theoretical in the analog domain of the signal to archieve better results then?

Originally Posted by madshi

However, dithering down to 8bit should produce very good results. I don't think anyone could see a difference between dithered 8bit and 10bit.

I'd be interesting in your opinion on the following:

Wouldn't it be possible to reproduce any color space and any gamut and any brightness and gradient with any bitdepth, be it one bit only given ideal dither implementation? The only drawback of lower bitdepth should be a higher noise floor, effectively preventing darker tones to be resolved as they would be masked by the dither and quantization noise of course. Hence I'm not saying that the bitdepth is irrelevant but I'd claim that any color or tonal variation should be able to effectively displayed with whatever bitdepth down to the noise floor at least.

Audio domain teaches that when applying either, there are no stair steps or discrete loudness levels depending on the bitdepth but only the noise floor which is getting higher of about 6dB per bit. That the smoothness of an audio signal depends on the sampling rate or bitdepth is a very common misconception so I wonder why it is still so common with pictures and video. Everywhere, one can read "the higher the bitdepth, the higher the available steps", etc.

Originally Posted by madshi

Yes, please. madVR will most probably never do any kind of intermediate frame interpolation. So this discussion doesn't really belong here.

Well, at least from an admittedly quite theoretical point of view, the signal theory would dictate you to do so because the individual frames of a movie are snapshots or samples according to Nyquist/Shannon which - just as pixels - were never supposed to be shown separately but reconstructed as a continuous movement.

Most people would say that the higher a frame rate, the smoother the movie and movement of a movie will become. Theoretically, that's not true: the only thing which changes is the nyquist frequency. In that case: what speed of movement can be reconstructed without aliasing. Anything else, including the frame-by-frame display of a movie - strictly speaking - is wrong to begin with.

Originally Posted by madshi

Basically it's the same as with dithering vs debanding: Dithering tries to make sure that madVR processing does not add new banding artifacts.

I recognized that when making a screenshot of a movie when using madVR and dithering to 1-bit only, while error diffusion dithering results in images containing only up to 8 colors just as to be expected, with random dithering, several thousand colors are counted and I wonder why. Apparently, some fusion between neighbour pixels takes place, raising the amount of colors, but why?