Encoder configuration is affected by the choice of lossless or lossy coding.  In the case of lossless coding, the decoder can be assumed to have an exact local copy of the anchor picture (frame or field) used for prediction (ignoring transmission errors, of course). In this case, the encoder can simply encode the difference between a new input frame and another input frame used as an anchor.  In lossy coding,the decoder does not have an exact copy of any video frame, and instead, the encoder must have a copy of  the actual decoded anchor frame(s), from which to calculate the error prediction signal.

If the encoder would not use the decoded anchor frame, coding noise would propagate and increase with each successive predicted frame, until the next I picture was transmitted.  This prediction "drift" would cause a considerable degradation of image quality.  Because the decoder has some characteristics of a temporal integrator, the result would be an amplification of low temporal frequencies in the coding errors.  

The encoder must therefore include a standard decoder to generate its anchor frames for prediction.  The overall process is then one of "noise feedback", which is feasible because the "noise" is known (unlike thermal noise in analog systems, which cannot be predicted).

Notice that one effect of this arrangement is that still scenes will not be coded as inactive (that is, with prediction error = 0 on every frame), since each successive prediction will attempt to cancel out the quantizing noise in the previous anchor frame.  At low bit rates, this may be noticeable as a random "crawling" or "boiling" motion in textured areas or a time varying, blocky mottling of relatively smooth areas.  The exact nature of the effect will depend on the interaction of motion compensated prediction (if used) and the quantizer.

Other aspects of the coded signal that are transmitted without prediction may be generated in a variety of ways, as long as the result is compatible with the capabilities of decoder.  For example, motion values may be estimated in any one of several ways.  While the accuracy of the estimation affects ultimate picture quality at a given bit rate, any motion value within the decoder's range of motion compensation is acceptable.  It is areas like this that leave room for improvement of coding within the MPEG 2 standard, for example by trading off the quality of motion estimation vs. the number of bits required to send the error prediction signal.

Example - one-dimensional lossy system