A sequence consists of all the pictures that follow a sequence header till a sequence_end_code.
Encoding and displaying parameters are transmitted with the sequence header.
The sequence header can be repeated in order to allow random access, but all the data elements of the repeated sequence header, except those concerning quantisation matrices, must have the same values as in the first sequence header.
The repeated sequence header must precede in the bitstream either an I-picture or a P-picture. In the case that random access is performed to a P-picture, it's possible that the decoded pictures may not be correct.

As the GOP layer is optional the management of the pictures is performed at the sequence layer.
An important item is the fact that the frames are not coded in the order in which they are displayed.
In particular the B_frames, that use references "from the future", are always coded after the P_frame (or the I_frame) used for backward predictions.



The frame reordering causes a delay in the coding and decoding processes. On the coder side the delay is given by the number of B_frames, that must wait for the following P_frame (or I_frame), while the decoder should wait for having full the two frame-memories before starting the display.
For special application is possible to code the sequence without any B_frame with the mode low_delay, set at the sequence layer. In this case the decoder needs only one frame-memory.