Real-time long-video generation is bottlenecked by two things: chunk-wise autoregressive diffusion must serialize denoising and refresh a history cache — two forward passes even with one-step denoising — and longer rollouts degrade quality while the camera drifts. On October 31, DENG Lab at Shanghai Jiao Tong University posted Wave Forcing to arXiv; three days later it topped Hugging Face Daily Papers as paper of the day (86 upvotes), with a thorough redesign of this pipeline.

Later chunks start before earlier ones finish

The core idea is wavefront scheduling: a later video chunk does not wait for earlier chunks to finish denoising. Through chunk-wise causal attention, it accesses the cleaner intermediate states of earlier chunks and feeds that cleaner-to-noisier context into the current chunk. Once the reverse dependencies that block pipelining are removed, the system reaches 1F1C — after pipeline fill, every concurrent forward round steadily produces one output chunk. Output stops arriving in batches and becomes an even stream.

Training and runtime are co-designed

  • Three-stage training: mixed-noise denoising, then motion supervision, then causal adaptation, all using self-generated history so the model adapts to generating from half-finished context.
  • WaveRT runtime: Wave Parallelism assigns different chunk-stage tasks to full-model replicas, publishes layer-wise K/V states in one direction, and prefetches communication to hide it under the longer attention computation on the critical rank — plus optimized kernels and pipelined VAE decoding.

The numbers (self-reported by the paper and repo)

The team distilled WaveForcing-1.3B from the Wan2.1-T2V-1.3B base:

  • Throughput: on 8×H200 at 832×480, the 4-step checkpoint reaches 134.8 end-to-end FPS — 1,593 frames decoded in about 11.8 seconds, a 7.07× speedup over the matched 19.06 FPS single-GPU baseline; peak steady-state throughput is 157.8 FPS.
  • Quality: 5-step Wave Forcing stays comparable to Rolling Forcing overall, while on 100-second videos Dynamic Degree improves from 36.63 to 91.45 and Subject Preservation from 92.75 to 97.10.
  • Scalability: in DiT-only tests with matched SageAttention+W8A8 backends, step counts and GPU counts, Wave Parallelism delivers up to 2.53× the throughput of the faster of the Ulysses/Ring baselines across six model scales (1.3B-14B); a 14B system prototype reaches 33.5 end-to-end FPS (note: random-initialized weights — this measures system scaling only, not generation quality).

Open-source substance

WaveRT inference code is Apache 2.0; both the 4-step and 5-step checkpoints are on Hugging Face with a ModelScope mirror; the S1/S2/S3 training recipes are released too (under a separate academic-use license). The GitHub repo currently sits at 91 stars.

My take

This is not "yet another video model" — it is a systems redesign that fuses scheduling, training objective, and runtime into one loop. Speed-of-Light in the title points at latency, not quality SOTA. The property that really matters is 1F1C: once generation becomes a steady stream, live-style content, interruptible interactive video, and even real-time rendering for world models finally have a foundation. Stay calm, though: every number above is self-reported, eight H200s is a real barrier to entry, and until third-party reproductions land, "regime change" conclusions should wait.

When generated video starts being scored in FPS rather than "does it look right", has the second half of this race already begun?