This paper proposes a new adaptive prediction algorithm which exclusively selects, on a pel-by-pel basis, either intraframe or motion-compensated interframe prediction. Prediction selection is made referencing previous samples, thus transmission of the selection information is not needed. Compared with motion-compensated interframe coding, it has been shown that coding performance is improved to a great extent, by about 40% for pictures with very fast motion, which is beyond motion compensation capability, and by 50% for scene change where motion compensation is, in principle, of no use.