Streaming video generation, where an AI model produces a long clip block by block and can take new instructions as it goes, is expensive because every new block needs the model to run again. A new arXiv preprint describes APEX, a way to train such models to work in very few steps while using less memory during training. It is aimed at researchers and labs rather than end users, but it is the kind of efficiency work that eventually decides how cheap AI video tools become.
The technical problem is specific. A common method for making video models fast, called Distribution Matching Distillation, needs two helper networks during training: a frozen teacher that represents good output, and a separate fake-score network that keeps track of what the generator currently produces. For large video models, that second network costs memory and compute, and it has to be kept in step with the generator as it changes. APEX removes it. Using a trick called condition shifting, the same model plays both roles: the normal text prompt makes it generate, and a shifted version of the prompt makes it estimate its own output distribution. The frozen teacher from the earlier S2-Avatar system is kept, along with its Self-Replay Forcing training method.
The evaluation uses 180-second videos. Against S2-Avatar, APEX scores slightly higher on instruction fulfilment on the benchmark's Progressive track, 3.856 against 3.831, and on visual aesthetics, 0.641898 against 0.632606. In a 70-case human study, raters preferred APEX most clearly for following instructions, with a net preference of +30.0%, and by +4.3% each for consistency and video quality. Under matched training settings, it used 4.4% less peak GPU memory per GPU and ran 9.2% faster per training iteration than the standard distillation method.
These are modest savings, which is the honest headline. The gains in automatic scores are small, and the clearest win is that the video follows changing instructions better, which matters for interactive uses such as avatars and live storytelling.
The limits are visible in the abstract itself. Human raters preferred S2-Avatar on motion quality, audiovisual synchronisation was mostly a tie, and usability did not change. The human study is small, the comparison is against one earlier system, and this is a preprint that has not been peer reviewed. No code or product release is described.