Video to video ai - hbd (high bit depth)

[anton_foy]

Anyone in the know about how and if its possible to generate video to Video in high bit depth (10-bit at least) or some workaround (looking at you avisynth and VapourSynth users out there, remember stacked?)?

Why? Because 8-bit with often such hard contrast it outputs has clipped highlights and darker footage will be even harder to further color grade or filter.

The likes of "Luma ai" and "Runaway ai" are good examples but they only output 8-bit 420. I like to have control and not use online services like the mentioned and perhaps I could train the models myself if necessary. Any ideas are welcome.

[rgr]

You can convert your YUV420P8 video to YUV420/YUV422P10 in AviSynth using ConvertBits(10). Also, use ffmpeg.
But I don't think that will solve your problems.

[hello_hello]

I often (especially if the source is 8 bit) encode the video as 10 bit while applying 8 bit dithering. It's not always necessary, but numerous times I've found 10 bit is still prone to banding, even after debanding and dithering. It may be due to the fact that my monitor is 8 bit and banding is being created (or increased) when the video is converted to 8 bit RGB on playback, but dithering to 8 bit tends to prevent it. Especially for black and white video, or after applying denoising, as in both cases the likelihood of banding is increased. For absolute best results, you'd probably apply any filtering in 16 bit, but naturally it's slower, although generally faster than 16 bit stacked. Something like:

ConvertBits(16)
# Insert any filtering here
GradFun3Plus() # Dogway's function
ConvertBits(10, Dither=0, Dither_Bits=8)

Nothing is going to save a video with dark areas already crushed to black, which seems to be common these days, I assume in an attempt to give the appearance of more contrast. Often it's terribly over-done when old black and white movies are remastered.

[anton_foy]

Thank you both. I was not clear in my first post, I have the input video in 12-bit prores (slog3) but I am trying to find a way to generate video2video in 10-bit or more since it it usually only outputs 8-bit and introduces banding, clipping of shadows and highlights etc. The dithering is absolutely useful.

[anton_foy]

Thank you both. I was not clear in my first post, I have the input video in 12-bit prores (slog3) but I am trying to find a way to generate video2video in 10-bit or more since it it usually only outputs 8-bit and introduces banding, clipping of shadows and highlights etc. The dithering is absolutely useful.

Edit: @hello_hello
I agree and regarding the BW remastered versions they could easily choose the distribution of the luminance curve to avoid clipping/crushing.