APEX: Few-Step Streaming Video Generation via Condition Shifting

Streaming generation produces video incrementally, often in causal temporal blocks. Each block must remain consistent with prior frames, respond to evolving instructions and preserve motion and audiovisual coherence. Reducing denoising steps can lower inference cost, but distillation must preserve quality while training the few-step generator. DMD improves few-step sampling but its separate fake-score network introduces another large model component and an additional moving optimization target.

The method keeps three important pieces from the S2-Avatar training setup: a frozen pretrained reference teacher, Self-Replay Forcing for causal replay blocks, and perceptual regularization. The paper therefore should not be described as eliminating all auxiliary computation or all teacher-model cost. Its narrower claim is that it removes the independent fake-score network by reusing the generator parameters for fake estimation.

A practical reproduction needs careful separation of the generation and estimator conditioning paths, stable handling of detached replay samples, consistent noise/time sampling for teacher and fake predictions, and explicit tracking of memory and iteration time. Validate that the shifted-condition branch learns the intended estimator behavior rather than leaking into ordinary generation. Monitor training stability, replay-distribution drift, gradient norms, teacher-estimator agreement and quality metrics by dimension.

The reported human study contains 70 cases, so fine-grained subgroup conclusions may be uncertain. The automated comparison uses a specified StreamAV-Bench setup and does not establish performance on every video domain. The resource reductions are modest and measured under matched training settings; they do not quantify all end-to-end deployment costs. The results also show trade-offs in motion quality and synchronization. Independent reproduction, broader backbones, diverse prompts and larger human studies would strengthen confidence in generality.

The method is relevant to long-form, instruction-driven video generation where training cost and streaming consistency both matter: narrative clips, product demonstrations, multi-step object interactions and videos that must respond to evolving prompts. The value proposition is strongest when the extra fake-score network is a material training bottleneck. Teams should verify that their own workload values instruction adherence and efficiency enough to justify any quality trade-offs.