Suppose you want constant frame rate video over a bandwidth-limited channel with very simple encoding and decoding. Paeth’s predictor predicts that each new pixel will be equal to one of the three pixels N, NW, and W of it, based on an estimate of the gradient from those three pixels, and then you encode, normally, a precise residual from that prediction. This approach can be straightforwardly extended to use a three-dimensional prediction incorporating the previous frame as well.
However, instead of transmitting a precise residual, you could transmit a quantized approximate residual. For example, you could transmit a choice of 0, ±1, ±4, ±16, or ±64, which is 9 choices, 3.17 bits per pixel per color channel. If an image holds still for a little while, it’ll settle down to the correct image fairly quickly.
However, this still sucks, because the compression ratio is less than a factor of 3. You can do a little better by transforming to YUV (16-bit maybe) and subsampling the chroma, but to do a lot better by some such approach, you need to not transmit any bits for most pixels.