Intra-prediction is how a codec compresses a single frame without looking at any other frame. Instead of storing every pixel directly, the encoder predicts each small block from the pixels just above and to the left (which the decoder has already reconstructed), and stores only the small correction between the prediction and reality. It's the workhorse of i-frame compression, and the reason a JPEG-like still image inside a video is smaller than an actual JPEG of the same picture.
The intuition: most natural images have lots of regions where neighbouring pixels are similar — a clear sky, a brick wall, smooth skin. The encoder runs a handful of standard prediction modes (extend the pixels from above downward, extend the pixels from the left rightward, blend both, use a diagonal pattern, etc.), picks whichever one matches the actual block best, and stores only the small leftover difference. For a block of sky, the prediction is almost perfect; for a block of detailed grass, less so — but in both cases the result compresses better than coding the raw pixels.
Modern codecs have made intra-prediction dramatically more sophisticated. H.264 had 9 directional modes for 4×4 blocks; HEVC has 35; AV1 has 56 plus newer non-directional modes and tools like "intra block copy" that lets a block in a screenshot reference another part of the same frame. The aggregate effect is that an HEVC I-frame is around 25 % smaller than an H.264 I-frame, and an AV1 I-frame is smaller still. For product teams, this is invisible technology that quietly powers thumbnail generation, video previews and the first frame of every stream.
