acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-time long-video generation requires fast, continuous output with consistent appearance and natural motion. Chunk-wise autoregressive methods serialize denoising and history-cache refresh, requiring two forwards even with one-step denoising. Their long rollouts also often suffer from quality degradation and camera drift. We propose **Wave Forcing (WF)**, a co-design of generation, training, and execution for continuous output with full multi-step denoising. WF pipelines chunks along a causal wavefront at staggered denoising stages, using chunk-wise causal attention to access earlier, cleaner chunks before they finish. Distributing denoising and cache-refresh stages across full-model replicas yields **1F1C**: one chunk per concurrent forward round after pipeline fill. Three-stage training develops joint mixed-noise denoising, strengthens motion and framing, and adapts the model to causal inference with self-generated history. **Wave Runtime (WaveRT)** uses **Wave Parallelism (WP)** for early K/V delivery and communication–computation overlap, complemented by kernel optimization and load-balanced VAE decoding. On Wan2.1-T2V-1.3B, 5-step WF maintains overall quality comparable to Rolling Forcing with higher Dynamic Degree scores on short and long videos. With SageAttention, our 4-step model reaches 117.7 E2E FPS on eight H200 GPUs, delivering a speedup over the 4-step single-GPU baseline.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Streaming Video Generation, Efficient Video Generation, Machine Learning Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3jXGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Wave Forcing pipelines video chunks across denoising stages using causal attention and early K/V delivery. WaveRT implements this schedule with communication overlap and pipelined VAE decoding. The paper reports up to 117.7 FPS on eight H200 GPUs and evaluates generation quality on roughly 100-second videos.

Critical strengths

The model and runtime are designed together, and the system gains are backed by useful ablations. There are also matched-GPU DiT throughput comparisons with Ulysses and Ring, beyond the headline comparison against a single GPU.

Critical weaknesses
  1. I could not find playable video results in the submission. The paper makes claims about natural motion and long-term stability, but sparse frames cannot show flicker, repeated motion, or discontinuities between chunks. The quantitative results are useful, but without full videos I cannot properly assess the quality of the generated sequences.

  2. The scope of the streaming speedup could be clearer. Appendix B.4 already notes that pipelining does not itself reduce first-output latency. A related limitation matters for interactive use: several chunks are in flight, so faster output does not necessarily mean a faster response to new input. A brief discussion would help clarify which applications benefit from the higher throughput and where response latency may still be a constraint. This is a request for clarification, not a reason for rejection, and I do not expect additional interactive experiments.

Questions for authors
  1. Appendix D explains how live K/V exchange differs from the stage-local caches in HiAR and LiveAvatar. What is the main benefit of this choice, and which existing results best support it?

  2. For Subject Preservation and Camera Drift, how many videos and how much video duration actually contribute to each method’s score? Are there substantial differences in coverage across methods?

Reviewer_RWciOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Wave Forcing (WF) for real-time streaming video generation. Starting from Rolling Forcing (RF), WF replaces RF's bidirectional attention among chunks in the rolling window with chunk-wise causal attention, allowing each denoising stage to run on its own full-model GPU in a pipeline. After the pipeline fills, the system produces one chunk per forward round (1F1C).

The proposed training contains 3 stages: RF-style mixed-noise DMD, off-policy DMD with an LPIPS loss on staggered re-noised teacher videos, and on-policy DMD. The runtime optimization further improves performance.

Critical strengths
  • RF's bidirectional in-window dependencies do prevent a stage-wise pipeline across GPUs. The motivation to enable the cross-GPU pipeline is well-motivated.
  • It takes a huge engineering effort.
  • The paper points out that a high Dynamic Degree can cause camera drift.
Critical weaknesses
  • The core modeling idea is not novel and the most closely related work is not cited. The core design, chunk-wise causal attention over chunks at different noise levels, is already used by MAGI-1 and SkyReels-V2. The paper does not cite either. The few-step DMD recipe is taken from Self Forcing, Rolling Forcing, and LPM 1.0, and distributing denoising stages across GPUs is standard pipeline parallelism, already used for rolling-window denoising by FIFO-Diffusion. The authors should compare WF with these works.
  • The systems side combines standard techniques. The pipeline parallelism over denoising steps, early K/V publication, CUDA IPC, fixed-slot buffers, SageAttention, and W8A8 are standard practice and not a contribution.
  • The NF1C bottleneck is not real. APT2 and Causal-Forcing have demonstrated the feasibility of frame-wise AR video generation.
  • The 7.9x speedup includes Sage-Attention, an orthogonal kernel that the single-GPU BF16 baseline does not use. The 5.7x speedup utilizes 8 GPUs vs 1, so per-GPU efficiency is not as expected.
  • Presentation
    • Abbreviations are dense and sometimes undefined. The paper relies on WF, WP, WaveRT, 1F1C, NF1C, E2E, and many labels. Several are introduced without definition.
    • Section 3.1 is hard to follow. It opens directly with index arithmetic without first stating the intuition. The core idea, that noisier chunks attend only to cleaner chunks, is never stated plainly.

Teng, H., Jia, H., Sun, L., Li, L., Li, M., Tang, M., ... & Li, Y. (2025). Magi-1: Autoregressive video generation at scale. arXiv preprint arXiv:2505.13211.

Chen, G., Lin, D., Yang, J., Lin, C., Zhu, J., Fan, M., ... & Zhou, Y. (2025). Skyreels-v2: Infinite-length film generative model. arXiv preprint arXiv:2504.13074.

Kim, Jihwan, et al. "Fifo-diffusion: Generating infinite videos from text without training." Advances in Neural Information Processing Systems 37 (2024): 89834-89868.

Lin, S., Yang, C., He, H., Jiang, J., Ren, Y., Xia, X., ... & Jiang, L. (2026). Autoregressive adversarial post-training for real-time interactive video generation. Advances in Neural Information Processing Systems, 38, 41061-41086.

Zhu, H., Zhao, M., He, G., Su, H., Li, C., & Zhu, J. (2026). Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation. arXiv preprint arXiv:2602.02214.

Questions for authors
  1. Can you report RF (bidirectional) trained with the same Stage 2 off-policy teacher supervision?

  2. How does WF differ from Diffusion Forcing, FIFO-Diffusion, and MAGI-1 in the training recipe? Can you compare it against them?

  3. What is the speedup against a single-GPU baseline that also uses SageAttention? What is the per-GPU efficiency? e.g., per-GPU throughput?

  4. Which values in Table 1 were “unconfirmed,” as the caption states?

Reviewer_D8TjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Wave Forcing, a framework for streaming video generation that combines a causal wavefront denoising schedule, staged training, and a multi-GPU runtime called WaveRT. Consecutive video chunks advance through staggered denoising stages, allowing newer chunks to access earlier, cleaner chunks before they finish. WaveRT maps these stages onto model replicas and overlaps layer-wise K/V communication with computation.

Critical strengths
  • improvements. WaveRT addresses practical bottlenecks in multi-GPU streaming video generation through early K/V publication, communication–computation overlap, and balanced VAE decoding. The system ablations support the usefulness of these optimizations.
  • Clear presentation. The paper is clearly written and well organized. The figures and explanations make the generation schedule, dependency structure, and runtime design easy to follow.
Critical weaknesses

My main concern is the insufficient discussion of Diffusion Forcing (Chen et al., NeurIPS 2024), which is not cited in the current manuscript. Causal Diffusion Forcing already explores denoising sequences with different noise levels across temporal positions, including staggered or pyramid schedules in which earlier tokens are cleaner than later ones. These ideas appear closely related to the central wavefront formulation presented here. This conceptual overlap weakens the novelty of the proposed generation schedule. The useful contributions of WaveRT should be acknowledged, but the manuscript needs to explain more precisely what the underlying generation formulation adds beyond Causal Diffusion Forcing.

Questions for authors

My main concern is the wavefront denoising schedule's improvement comparing to diffusion forcing. I would therefore like to see an explicit comparison of their noise schedules, causal conditioning, and inference procedures, together with a direct empirical comparison against Causal Diffusion Forcing under comparable model and compute settings. This would help establish the contribution beyond the earlier framework.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.