Oriel — AI & technology worth noticing, every day.

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

Paper · https://arxiv.org/abs/2611.00859

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).

First seen by Oriel on .

AI summary

Wave Forcing co-designs training and distributed execution for continuous multi-step video generation, reporting 1F1C chunk output and higher throughput while maintaining quality comparable to Rolling Forcing.

  • Chunk-wise causal attention lets the model access earlier, cleaner chunks before their denoising is complete.
  • Wave Runtime combines Wave Parallelism for early K/V delivery and communication-computation overlap with optimized kernels and pipelined VAE decoding.
  • A three-stage training procedure adds mixed-noise denoising, motion supervision, and causal adaptation using self-generated history.
  • On Wan2.1-T2V-1.3B, 5-step WF achieves comparable overall quality to Rolling Forcing and higher Dynamic Degree on short and long videos.
  • On eight H200 GPUs, 4-step WF reaches 147.6 DiT FPS and 134.8 end-to-end FPS; Wave Parallelism delivers up to 2.53x baseline throughput across model scales from 1.3B to 14B.

#video generation #diffusion models #distributed inference #wave parallelism

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hugging Face Papers: 126, latest observation ; +30 over observed period since

Comments

Comment count reported by the source.

  • Hugging Face Papers: 6, latest observation ; +3 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Papers2611.00859https://arxiv.org/abs/2611.00859 Read discussion