Streaming video generation has a scheduling problem as much as a model-quality problem: generating a chunk and refreshing history sequentially can require separate forward passes, limiting output rate even when denoising takes only one step. The key insight here is to let denoising stages overlap across chunks and replicas, so steady-state throughput is no longer tied to finishing each chunk’s work in isolation.
Key Findings
- One output chunk per concurrent forward round after pipeline fill: Staggered denoising and cache refresh across full-model replicas target continuous output rather than serial chunk completion.
- Motion without the usual stability trade-off: On Wan2.1-T2V-1.3B, the five-step model scores 81.64 on VBench Total and 91.45 on long-video Dynamic Degree, while retaining 97.10 Subject Preservation and a 0.37 Camera Drift score. These results suggest that more motion need not mean losing the subject or letting the camera wander.
- System gains are substantial but configuration-specific: Four-step generation reaches 134.8 end-to-end FPS on eight H200 GPUs, or 7.07× single-GPU end-to-end throughput. In DiT-only tests, Wave Parallelism reaches up to 2.53× the faster Ulysses or Ring baseline under matched settings.
The mechanism depends on chunk-wise causal attention that can read earlier, cleaner chunks before their denoising is complete. Training progressively adapts the model to mixed noise, motion supervision, and self-generated causal history; Wave Runtime then overlaps early K/V delivery, computation, and VAE decoding. This co-design matters because a fast denoiser alone does not guarantee a fast streaming system.
Who Should Care?
A good fit for researchers and systems engineers working on continuous, long-form video generation, especially where motion and subject consistency both matter. Look elsewhere if you need evidence for interactive prompt changes, low input-to-visible-output latency, or performance across varied hardware: the reported evaluation uses fixed conditioning and specific model and GPU configurations, and the multi-stage training process is complex.
