I'm not sure how far you'd get; since PNG's predictor selection is IIRC only done on a scanline basis you don't have as many opportunities to really leverage prediction. Lossy image compression typically operates using a combination of planar modifications (like subsampling) and block-level modifications.
On the flipside, part of why PNG does so well on lossless compression is due to the scanline-oriented predictor selection - that allows it to beat regular GZIP by identifying ways to losslessly compress 2D image patterns that aren't necessarily obvious in the 1D stream of bytes being sent to the compressor.
On the flipside, part of why PNG does so well on lossless compression is due to the scanline-oriented predictor selection - that allows it to beat regular GZIP by identifying ways to losslessly compress 2D image patterns that aren't necessarily obvious in the 1D stream of bytes being sent to the compressor.