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

Wave Forcing co-designs causal wavefront generation, three-stage training, and Wave Runtime to stream video at one chunk per concurrent forward round after pipeline fill while retaining all denoising updates.

Existing chunk-wise autoregressive methods serialize denoising and clean-history cache refresh, so even one-step denoising needs two forward passes and sequential execution gives NF1C. Longer rollouts can degrade quality and cause camera drift. The paper targets 1F1C: one completed chunk per concurrent forward round after pipeline fill while retaining every denoising update.

Wave Forcing uses a bounded-width causal wavefront: chunk-wise causal attention lets newer, noisier chunks read earlier, cleaner chunks before those chunks finish denoising, but not vice versa. Denoising stages and the store stage are assigned to full-model replicas, forming a pipeline that emits one chunk per round after fill. Training follows three stages: mixed-noise denoising, teacher-guided motion supervision with off-policy DMD, and causal adaptation using self-generated history. WaveRT implements Wave Parallelism by publishing layer K/V early, overlapping communication with computation, writing K/V into paged attention slots, and pipelining VAE decoding across load-balanced GPUs.

• The 5-step model leads VBench Total 81.64 and Quality 84.40, with VBench-Long Quality 81.26. §5.1
• Wave Forcing raises long-video Dynamic Degree to 91.45 and Subject Preservation to 97.10 while Camera Drift is 0.37; Rolling Forcing has Dynamic Degree 36.63, Subject Preservation 92.75, and Drift 0.17. §5.1
• 4-step WF with BF16/FA3 reaches 147.6 DiT FPS and 134.8 end-to-end FPS on eight H200 GPUs, a 7.07× end-to-end speedup over the 19.06-FPS single-GPU baseline. §1
• Across six model scales, WP with Sage+W8A8 reaches 1.025–2.530× the throughput of the faster CP baseline. §5.5
• At 14B, E2E rates are 25.47, 28.64, 28.62, and 33.50 FPS versus a 5.51-FPS single-GPU reference; best relative gains are 7.07× at 1.3B and 6.08× at 14B. §5.3
• Sync-to-Paged lowers BF16 tick p50 from 91.96 to 78.88 ms, 14.2%, at 1.3B and from 464.27 to 413.80 ms, 10.9%, at 14B. §5.4

• The paper states that evaluation establishes streaming generation under fixed conditioning and the tested model, sampling, and hardware configurations (§7).
• New 1.3B and 14B system tests use dummy DiT weights and a real-weight VAE to evaluate execution cost, not generation quality (§C.3).
• Explicitly retained video context is limited to the anchor and recent history (§B.8).
• The resident launcher handles one queued request at a time, and interaction and world-model extensions have not been evaluated with WF (§B.3 and §E).

• WF targets 1F1C: one completed chunk per concurrent forward round after pipeline fill while retaining every denoising update. §1
• Chunk-wise causal attention keeps full attention within each chunk and lets query slot s see K/V slot j iff j >= s, a lower-triangular mask over self and cleaner predecessors. §3.1
• Per-layer context order is anchor, bounded clean history, live store, then denoising K/V in increasing j with self first. §3.1
• Training uses mixed-noise initialization, teacher-guided motion supervision via off-policy DMD, and causal adaptation with self-generated history. §3.2
• WaveRT assigns a full-model replica to each denoising stage and store; GPU allocations are 4+1+3 for 4-step and 5+1+2 for 5-step WaveRT, both totaling eight GPUs. §B.1
• WaveRT publishes K/V after QKV projection and RoPE and checks readiness immediately before attention; delivery is fully hidden only when every edge satisfies tau_e <= Delta_e. §4.2
• Paged delivery writes K/V directly into ordered attention slots and signals readiness before use, avoiding remote concatenation under fixed local conditions. §4.2
• VAE decoding overlaps DiT and is balanced across contiguous decoder partitions using profiled streaming times; the first output still incurs pipeline fill latency. §4.3