Only key frames and no frames that are in any way predicted by what came before it. That's actually the major point that differentiates formats intended for con

Only key frames and no frames that are in any way predicted by what came before it. That's actually the major point that differentiates formats intended for consumption and those intended for editing. If you usually only watch something from start to end, you can vastly increase compression ratio by re-using parts of earlier frames that changed very little. However, this means that to get to a particular frame you have to find the last keyframe before that and then decode from there.