High speed, low complexity, and interoperability are just three of the main advantages turning the MPEG stream watermarking into a hot research topic. Unfortunately, viable solutions (in terms of robustness, data payload and transparency) are yet to be found. In their previous work, the authors computed general models for the watermarking attack effects (StirMark, linear & nonlinear filtering, rotations) in the MPEG 4 AVC stream. These models (expressed as noise matrices) are now the starting point for evaluating three classes of watermarking insertion techniques (substitutive, additive, and multiplicative). For each class, a specific set of noise matrices is first computed by particularising the general model. Secondly, the corresponding capacity values (i.e. the largest data payload which can be inserted for prescribed transparency and robustness) are computed. The paper is concluded with a comparison among these method performances. The experiments are run on a video corpus of 10 video sequences of about 25 minutes each