Wave Forcing: Turning Long-Video Generation into Continuous Streaming

Real-time long-video generation is not simply a matter of producing individual frames faster. A practical system must keep output continuous while preserving a subject’s identity, the scene’s appearance, and natural motion over a long horizon. Existing chunk-wise autoregressive pipelines often process video chunks in strict sequence: a chunk is denoised, its history cache is refreshed, and only then can the next chunk proceed. This creates extra forward passes and leaves long rollouts vulnerable to quality degradation and camera drift.

Wave Forcing (WF) tackles the problem by redesigning generation, training, and execution together. Its central idea is wavefront scheduling. Instead of waiting for an earlier chunk to finish every denoising stage, a later chunk can start its own denoising earlier and use cleaner intermediate states from the preceding computation. Chunk-wise causal attention provides the temporal constraint needed for streaming while allowing the model to access useful context before the entire preceding pipeline has completed.

The execution schedule distributes staggered denoising and cache-refresh stages across multiple full-model replicas. After the pipeline is filled, each concurrent forward round produces one output chunk, a pattern described as 1F1C: one forward pass, one chunk. This changes the bottleneck from fully serialized chunk processing to a continuously operating pipeline.

Wave Runtime, or WaveRT, is designed to make that schedule practical. Its Wave Parallelism mechanism sends key/value states ahead of when they are consumed and overlaps communication with model computation. Optimized kernels reduce operator overhead, while pipelined VAE decoding helps prevent video reconstruction from becoming the next bottleneck. The result is a system-level design in which model execution, context movement, and decoding are treated as parts of the same streaming process.

The model also needs training that matches this unusual inference pattern. WF uses three successive stages. Mixed-noise denoising exposes the model to history chunks with different levels of cleanliness. Motion supervision focuses on temporal dynamics rather than appearance alone. Causal adaptation then uses self-generated history, reducing the mismatch between clean training inputs and the imperfect, continuously produced context encountered during rollout.

Experiments with Wan2.1-T2V-1.3B show that five-step WF achieves overall quality comparable to Rolling Forcing while producing higher Dynamic Degree on both short and long videos. In the 100-second evaluation, Dynamic Degree rises from 36.63 to 91.45, while Subject Preservation increases from 92.75 to 97.10. In system tests, four-step WF reaches 147.6 DiT FPS and 134.8 end-to-end FPS on eight H200 GPUs, equivalent to 7.07 times the single-GPU end-to-end throughput. Under matched attention backends, denoising steps, and GPU counts, Wave Parallelism reaches up to 2.53 times the throughput of the faster Ulysses or Ring baseline.

The broader contribution is a shift in how streaming video generation is optimized. Rather than relying only on fewer sampling steps or a smaller model, Wave Forcing coordinates when chunks are generated, how intermediate context is exposed, and how computation is distributed. The reported results are tied to a specific model and hardware setup, so gains may vary with scale, resolution, and decoding costs. Even so, the work presents a clear systems path toward continuous, long-form video generation.

Source: Hugging Face Daily Papers