Wave Forcing: Towards Speed-of-Light Streaming Video Generation

Yijie Jin, Yuxuan Liu, Jiajun Xu, Ye Huang, Chuzhi Wang, Jiayang Li, Ruiyang Chen, Chenling Hong, Xu Wang, Zhijie DengarXiv 2026 · arXiv

TL;DR

Wave Forcing (WF) is a co-designed framework that achieves "speed-of-light" streaming video generation by pipelining multi-step denoising across multiple GPUs. By allowing newer video chunks to "look ahead" at cleaner, partially denoised versions of preceding chunks, it achieves a throughput of one completed chunk per forward pass (1F1C), delivering up to 134.8 end-to-end FPS and superior long-video stability compared to existing autoregressive methods.

What Problem Does This Paper Address?

Generating long, high-quality videos in real-time is a massive computational challenge. Most current state-of-the-art models use chunk-wise autoregressive generation, where the video is generated in small segments (chunks).

However, these methods face two major hurdles:

  1. Efficiency Bottlenecks: Standard methods require at least two forward passes per chunk—one for denoising and one to "store" the clean result into a Key/Value (K/V) cache for future chunks. This results in an "NF1C" (N Forward rounds for 1 Chunk) schedule, which is inherently slow.
  2. Quality Degradation: As the video gets longer, errors accumulate. This often leads to "camera drift" (the background moves unnaturally) or "subject loss" (the main character disappears or transforms into something else).

Methods

The authors propose a holistic co-design of the generation schedule, the training process, and the hardware execution.

1. Causal Wavefront Scheduling

Instead of waiting for chunk $A$ to be fully denoised before starting chunk $B$, Wave Forcing uses a staggered pipeline. In any given "round," different GPUs work on different stages of denoising for different chunks.

  • GPU 1 might be doing the first denoising step for Chunk 5.
  • GPU 2 is doing the second step for Chunk 4.
  • GPU 3 is doing the final "store" pass for Chunk 1.

This creates a "wavefront" that moves through the video. Crucially, they use chunk-wise causal attention, allowing a noisy chunk to attend to earlier chunks that are "cleaner" (further along in their denoising process), even if they aren't finished yet.

Figure 1: Causal wavefront and continuous output (S = 4). (a) Denoising and store advance on anti- diagonals, with dashed cells marking the next round. (b) One later-layer K/V edge arrives before use. Other edges and initial waits are omitted. (c) Both schedules retain four denoising forwards plus store per chunk. Over five rounds, one SF replica emits one chunk, whereas five WF replicas emit five after fill. Wall-clock gains depend on stage costs and overlap.

2. Three-Stage Training

To make the model robust to this "mixed-noise" context, the authors developed a specialized training recipe:

  • Stage 1: Mixed-noise initialization. Teaches the model to handle bidirectional attention across chunks at different noise levels.
  • Stage 2: Off-policy Distillation (DMD). Uses teacher videos to provide strong motion and subject guidance.
  • Stage 3: Causal Adaptation. Fine-tunes the model on its own self-generated history to prevent error accumulation during long rollouts.
Figure 2: Training flow. Stage 1 applies DMD to one sampled RF or SF rollout. Stage 2 re-noises teacher chunks and applies DMD and LPIPS to the same prediction bYθ. Stage 3 samples one self-generated RF, WF, or SF rollout for on-policy DMD.

3. Wave Runtime (WaveRT)

To translate this theoretical scheduling into actual speed, the authors built WaveRT. It uses Wave Parallelism (WP) to overlap communication and computation. As soon as a "producer" GPU finishes calculating K/V values for a layer, it immediately sends them to the "consumer" GPUs in the pipeline, rather than waiting for the entire forward pass to finish.

Datasets

The researchers evaluated their methods primarily using:

  • Wan2.1-T2V-1.3B: The base Diffusion Transformer (DiT) model used for implementation.
  • VBench & VBench-Long: Standard benchmarks for video quality, semantic consistency, and motion.
  • MovieGen Prompts: A set of 200 prompts used to test long-video (100-second) generation.
  • Subject-focused Subset: A newly curated 100-prompt subset specifically designed to test "Subject Preservation" and "Camera Drift" over long durations.

Key Insights and Ideas

  • Intermediate States are Useful Context: The core "aha" moment is that a partially denoised chunk is a "good enough" conditioning signal for the next chunk's earlier denoising stages. You don't need to wait for perfection to maintain temporal flow.
  • 1F1C is the Efficiency Limit: By assigning a full-model replica to each denoising stage, the system reaches the theoretical limit of one chunk output per concurrent forward round.
  • Co-design is Mandatory: You cannot simply take a standard model and run it in a wavefront pipeline; it will produce garbage. The training must be aware of the specific causal mask and noise distribution used during inference.

Key Results

  • Throughput: On eight NVIDIA H200 GPUs, 4-step Wave Forcing achieved 134.8 end-to-end FPS, a 7.07x speedup over a single-GPU baseline.
  • System Efficiency: Wave Parallelism outperformed traditional context-parallel baselines (like Ulysses or Ring Attention) by up to 2.53x across various model scales (1.3B to 14B parameters).
  • Long-Video Stability: On 100-second videos, WF maintained a Subject Preservation score of 97.10 (compared to 92.75 for Rolling Forcing) and significantly higher "Dynamic Degree," meaning the videos had more natural, active motion without losing the subject.

Limitations and Future Work

  • Training Complexity: The current multi-stage training process is complex and computationally expensive. Simplifying this is a priority.
  • Fixed Sampling Budget: The degree of parallelism is currently tied to the number of denoising steps. If you want a 50-step high-quality render, you'd need 50+ GPUs to maintain the 1F1C throughput.
  • Interactive Latency: While throughput (FPS) is high, the initial latency (time to first frame) is still limited by the pipeline fill time.
  • Future Directions: The authors aim to extend this to interactive world models, where user inputs (like camera controls) can affect the "in-flight" chunks of the wavefront in real-time.

Summary updated

Abstract

Real-time long-video generation requires continuous output with consistent appearance and natural motion. Existing chunk-wise autoregressive methods face challenges in both generation efficiency and long-video quality: serialized denoising and history-cache refresh require two forward passes even with one-step denoising, and longer rollouts can degrade quality and cause camera drift. We propose Wave Forcing (WF), which co-designs generation, training, and execution for continuous video output with multi-step denoising. WF uses chunk-wise causal attention to access earlier, cleaner chunks before their denoising is complete, and distributes staggered denoising and cache-refresh stages across full-model replicas to achieve 1F1C: one output chunk per concurrent forward round after pipeline fill. To adapt the model to this continuous generation process, we develop a three-stage training procedure that successively introduces mixed-noise denoising, motion supervision, and causal adaptation using self-generated history. To execute this process efficiently, Wave Runtime (WaveRT) uses Wave Parallelism (WP) for early K/V delivery and communication-computation overlap, together with optimized kernels and pipelined VAE decoding. Experiments on Wan2.1-T2V-1.3B show that 5-step WF maintains overall quality comparable to Rolling Forcing while achieving higher Dynamic Degree on both short and long videos. Our system efficiency experiments show that 4-step WF achieves 147.6 DiT FPS and 134.8 end-to-end FPS on eight H200 GPUs, yielding 7.07x the single-GPU end-to-end throughput. In DiT-only tests with matched SageAttention+W8A8 backends, denoising steps, and GPU counts, WP delivers up to 2.53x the throughput of the faster of the Ulysses and Ring context-parallel baselines across six model scales (1.3B-14B).
BibTeX
@misc{jin2026wave,
  author = {Yijie Jin and Yuxuan Liu and Jiajun Xu and Ye Huang and Chuzhi Wang and Jiayang Li and Ruiyang Chen and Chenling Hong and Xu Wang and Zhijie Deng},
  title = {{Wave Forcing: Towards Speed-of-Light Streaming Video Generation}},
  year = {2026},
  eprint = {2611.00859},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00859},
}
Figure