[grishka]

He encodes bits as signs of DCT coefficients. I do feel like this is not as optimal as it could be. A better approach IMO would be to just ignore the AC coefficients altogether and instead encode several bits per block into the DC. Not using the chrominance also feels like a waste.

[brandonli28]

This actually won't work against YouTube's compression. The DC coefficient is always quantized, rounded, scale, and any other things. That means that these bits are pretty much guaranteed to be destroyed immediately. If this is the case for every single block, then data is unrecoverable. Also, chrominance is not used on purpose, because chrominance is compressed much more aggressively compared to luminance.

[grishka]

I meant choosing multiple values, e.g. 4 to represent 2 bits. Say, 0.25, 0.5, 0.75, and 1. Then when decoding you would pick the closest valid value, so for example for 0.20 it would be 0.25. Not using AC coefficients would mean that theoretically you would get more bitrate for the DC ones.

[brandonli28]

I’ve been told this many times in the comments, but this again is not reliable. Simply put, compression doesn’t necessarily follow a pattern, so specifying “ranges” or rounding to a specific place will not work. Compression optimizes for the eye, and doesn’t do the same thing for every value. It will round some down, some other mores, others less. Giving a range is simply not enough.