I suspect that better sounds could be made from images if the data was treated as if it were in frequency domain instead of time domain, because then some filtering could make more sense. This could open up some more possibilities, such as even transforming color distribution data into musical notes (I wonder if nice-looking color chords can be transformed into nice-sounding musical chords).
This is very interesting, and I love both coding and music, so I guess I'll get on making something like that :)
I'm not sure I understand your comment, but I think that's what is already happening. Consider this: WAV is a sequence of samples - BMP is sequence of RGB values. So you can more or less treat it the same.
If you would try to read JPEG (i.e. frequency domain data) as WAV it would be a mess. The "appropriate" format would then be something like MP3 but you would run into issues - MP3 contains more metadata in the streams, JPEG encodes 8x8px 2D images while MP3 encodes ~576 1D samples...
This is very interesting, and I love both coding and music, so I guess I'll get on making something like that :)