This is a Plain English Papers summary of a research paper called APEX Uses Condition Shifting to Simplify Few-Step Streaming Video Generation. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
APEX removes DMD’s extra fake network and improves instruction following
APEX adapts few-step streaming video generation to a shared-network design: it estimates the generated video distribution inside the generator instead of maintaining DMD’s separate fake-score network. Against S2-Avatar at three backbone evaluations per block, it scores higher on Progressive-track video instruction fulfillment (3.856 vs. 3.831) and visual aesthetics on both benchmark tracks (0.641898 vs. 0.632606). In a 70-pair human study, its clearest advantage is instruction following, with +30.0% net preference; matched training measurements also show 4.4% lower peak GPU memory and 9.2% shorter iteration time than DMD.
The method targets long-form streaming, where every new video block incurs repeated model evaluations. APEX retains S2-Avatar’s frozen reference teacher, Self-Replay Forcing (training on segments replayed from the model’s own causal rollouts), and perceptual regularization, while replacing only the independent fake estimator.

Long-form samples pair Progressive and Interactive benchmark prompts with varied actions and storytelling. They give concrete context for the streaming setting and the instruction-following claims evaluated quantitatively....
Top comments (0)