Wave Forcing: Towards Speed-of-Light Streaming Video Generation
Real-time video generation has a scheduling problem as much as a modeling problem. A chunk-wise autoregressive system must repeatedly denoise each segment and refresh the clean key/value (K/V) history used by later segments. Even when denoising takes one step, that clean-history refresh can require another full model pass. As videos become longer, sequential execution also makes motion, subject identity, and camera stability harder to maintain.
In this paper, Yijie Jin, Yuxuan Liu, Jiajun Xu, Ye Huang, and colleagues propose Wave Forcing (WF), a joint design for video generation, training, and inference. The central goal is 1F1C: after pipeline fill, one completed video chunk per concurrent forward round. The authors do not claim that WF reduces the amount of work per chunk. Instead, it distributes that work across multiple full-model replicas so that denoising stages and cache refreshes proceed concurrently.
How the wavefront works
Traditional chunk-wise generation finishes one chunk before moving on, or uses a more broadly connected mixed-noise schedule. WF arranges chunks and denoising stages as a moving wavefront. An earlier chunk may be further along in denoising while a newer chunk is still noisy. Newer chunks can read earlier, cleaner intermediate states, but the reverse dependency is blocked. This creates a directed, causal context graph.
For a model using S denoising steps, WF assigns a full model replica to each denoising stage and adds another replica for the clean-history “store” operation. The store replica refreshes K/V history from a completed chunk while other replicas continue denoising. After the pipeline fills, the system can advance several chunks in every logical round rather than waiting for all operations on one chunk to finish.
The paper’s runtime, called Wave Runtime (WaveRT), addresses the communication costs created by this arrangement. Its Wave Parallelism approach publishes K/V data as soon as it is produced, overlaps transfers with local computation, and writes data directly into attention-buffer slots. The authors also describe optimized kernels and a pipelined variational autoencoder (VAE) decoder. These mechanisms are intended to reduce exposed waiting; they do not remove the underlying dependency or communication traffic.
Training is part of the method
Changing the inference graph changes what the model sees during generation, so the authors train WF rather than simply applying the schedule to an existing model. Their three-stage recipe begins with mixed-noise initialization, using distribution-matching distillation from RF or SF-style rollouts. A second stage adds teacher-guided motion and perceptual supervision on staggered, re-noised chunks. The final stage performs causal adaptation with self-generated histories, sampling Rolling Forcing, Wave Forcing, or Self Forcing rollouts.
The ablations support the authors’ argument that the recipe matters. In their 4-step long-video experiments, removing off-policy DMD reduced VBench-Long Quality from 80.89 to 76.68 and Dynamic Degree from 74.97 to 15.73. Removing the WF branch preserved subjects better but reduced Dynamic Degree, while removing the RF branch increased motion but produced substantially worse subject preservation and camera drift. These results suggest that the final balance comes from the combination of training branches, not from causal scheduling alone.
What the experiments show
The main quality evaluation uses Wan2.1-T2V-1.3B at 832×480 resolution and 16 frames per second. In the authors’ comparison, 5-step WF achieved a VBench Total score of 81.64, compared with 81.15 for 5-step Rolling Forcing. VBench-Long Quality was similarly close: 81.26 for WF versus 81.18 for Rolling Forcing.
The larger difference appeared in motion and stability. On 100-second videos, WF’s Dynamic Degree reached 91.45, compared with 36.63 for Rolling Forcing. Subject Preservation was 97.10 for WF and 92.75 for Rolling Forcing, while Camera Drift was 0.37 versus 0.17; lower is better for drift. The comparison also illustrates why motion scores need context. Causal Forcing achieved higher long-video quality and dynamic scores in some measurements, but the paper reports severe subject loss and camera drift for that configuration.
For system performance, the authors report 4-step WF running at 147.6 DiT frames per second and 134.8 end-to-end frames per second on eight H200 GPUs. This corresponds to 7.07 times the single-GPU end-to-end baseline reported in the paper. In DiT-only tests, Wave Parallelism reached up to 2.53 times the throughput of the faster Ulysses or Ring context-parallel baseline across the tested 1.3B to 14B scales and compute configurations.
Important limitations
These results should not be read as proof of universal real-time video generation. The quality checkpoint evaluation centers on Wan2.1-T2V-1.3B, while the runtime tests at several scales use dummy DiT weights and measure execution rather than generated-video quality. The authors did not train a 14B WF model because of compute constraints.
The reported end-to-end rates also depend on a particular eight-GPU H200 setup, denoising-step count, VAE allocation, backend, and resolution. At 1.3B, the gap between DiT and end-to-end throughput shows that decoding remains a meaningful bottleneck. Full-model replication consumes substantial GPU capacity, and the pipeline’s width is tied to the number of denoising steps. The authors note that the communication pattern grows quadratically with the number of stages.
Finally, the long-video stability diagnostics cover a fixed subject-focused subset and rely on tracking and geometric heuristics. The authors explicitly say that their sensitivity to thresholds and their agreement with human judgments remain to be established. High throughput also does not establish low first-output latency, prompt-switching responsiveness, or reliable interactive control; those scenarios were not evaluated.
Practical implications
For researchers building streaming video systems, WF’s main lesson is architectural: generation quality, training data flow, and serving infrastructure should be designed together. A causal wavefront can expose concurrency, but useful speed depends on stage balance, K/V delivery, decoder capacity, and enough workload to amortize pipeline fill and drain.
For practitioners, the reported gains are most relevant to fixed-conditioning workloads with access to multiple high-end GPUs. Applications requiring prompt changes, strict latency guarantees, or many concurrent streams would need additional scheduling and interaction experiments. The paper’s prospective roadmap—better VAE profiling, hybrid parallelism, deadline-aware serving, and richer human-calibrated evaluation—marks the work as an important systems direction, but not a complete solution to interactive long-video generation.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.