PaperlayerView exact revision on arXiv

Source-linked AI summary

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 Deng

arXiv:2611.00859v1cs.CVcs.DC

TL;DR

Chunk-wise autoregressive video generation serializes denoising and clean-history refresh, requiring two forward passes even with one-step denoising. Wave Forcing combines causal wavefront generation, staged training, and overlapped execution; it improves motion while maintaining overall quality and reports high-throughput streaming.

  • Problem

    Chunk-wise autoregressive methods must sustain playable output with consistent appearance and natural motion, yet serialized denoising and clean-history refresh require two forward passes even for one-step denoising.

    Source: §1 · p. 1

  • Method

    Wave Forcing uses chunk-wise causal attention to let newer chunks read cleaner predecessors, trains for mixed-noise contexts and self-generated history, and pipelines denoising and cache refresh across full-model replicas.

    Source: §1 · pp. 1–2

  • Results

    WF improves motion while maintaining overall generation quality; 5-step WF raises Dynamic Degree from 36.63 to 91.45 on 100-second videos versus RF.

    Source: §1 · p. 2, §9 · p. 11

  • Takeaways & Limitations

    The reported results demonstrate that jointly designing generation context, training, and execution can improve motion while maintaining quality and achieving high-throughput streaming video generation.

    Source: §9 · p. 11

    The fixed-conditioning experiments do not establish low input-to-visible-output latency or evaluate prompt switching, camera/action control, closed-loop tasks, or multi-stream deadline guarantees.

    Source: §7 · p. 10

Abstract

from arXiv · show

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).

1 Introduction

Chunk-wise autoregressive generation serializes denoising and clean-history refresh, requiring two forward passes even with one-step denoising. Wave Forcing combines causal wavefront generation, staged training, and overlapped execution to improve motion while maintaining quality and sustaining streaming throughput.

1 Introduction

  • 5-step WF matches RF overall quality while raising Dynamic Degree from 36.63 to 91.45 on 100-second videos.VBench Total is 81.64 versus 81.15, and VBench-Long Quality is 81.26 versus 81.18; Subject Preservation rises to 97.10 versus 92.75 while Camera Drift remains 0.37 versus 0.17.Source: §1 · p. 2
  • 4-step WF reaches 134.8 end-to-end FPS on eight H200 GPUs, a 7.07× speedup over the 19.06-FPS single-GPU baseline.It also achieves 147.6 DiT FPS with the BF16/FA3 backend.Source: §1 · p. 2
  • WF pipelines staggered denoising and clean-history refresh across full-model replicas to produce one chunk per concurrent forward round after fill while retaining every denoising update.Newer chunks condition on earlier, cleaner chunks before their denoising is complete.Source: §1 · pp. 1–2
  • A three-stage recipe combines mixed-noise initialization, teacher-guided motion supervision, and causal adaptation with self-generated history for long-rollout quality comparable to RF.The reported outcome includes strong motion and subject preservation.Source: §1 · p. 2
  • WaveRT uses early layer-wise K/V delivery and communication–computation overlap, alongside optimized local computation and load-balanced VAE decoding, to sustain end-to-end throughput.WP efficiency is validated across models from 1.3B to 14B.Source: §1 · p. 2

2 Preliminaries

Each chunk undergoes history-conditioned denoising updates, and the clean-history convention adds a store pass to refresh K/V. Sequential execution therefore uses one more forward than the number of denoising steps.

2 Preliminaries

  • A chunk begins from noise and undergoes S history-conditioned denoising updates at progressively lower noise timesteps.The VAE decodes latent-frame chunks into video.Source: §2 · p. 3
  • Each denoising step predicts clean latents that are re-noised for the next timestep; a separate full-model store forward refreshes clean-history K/V.The store retains K/V and discards its prediction, making sequential generation N = S + 1 forwards per chunk.Source: §2 · p. 3
  • At each filled runtime tick, active denoising stages advance and fresh noise is admitted, allowing 1F1C after pipeline fill without changing per-chunk work.The schedule also includes a store stage.Source: §2 · p. 3

3 Wave Forcing

Wave Forcing advances successive chunks before their predecessors finish denoising by combining causal wavefront attention with training for mixed-noise contexts and self-generated history.

3 Wave Forcing

  • WF lets newer, noisier chunks read earlier, cleaner states while preventing later chunks from conditioning earlier ones.It couples this causal wavefront with training for mixed-noise context and self-generated history.Source: §3 · p. 3

3.1 A bounded-width causal wavefront

  • The chunk-wise causal mask exposes each query to itself and cleaner predecessors, while store refreshes a chunk’s K/V for subsequent history.Store occurs at the final stage of the chunk’s schedule.Source: §3.1 · p. 3
  • Without post-training, short-video Quality and Dynamic Degree fall in the reported evaluation, motivating adaptation to the causal conditioning change.The causal mask removes access to later, noisier active chunks, changing RF’s learned conditioning.Source: §3.1 · p. 3

3.2 Training for the deployment graph

  • The 4-step student starts from SF’s ODE initialization, whereas the 5-step student reuses released RF weights and skips Stage 1.The causal mask changes RF’s learned conditioning, motivating the training adaptation.Source: §3.2 · p. 3
  • Training progresses from mixed-noise RF/SF DMD, through teacher-chunk DMD and LPIPS, to on-policy DMD over self-generated RF, WF, or SF rollouts.In the final stage, WF uses store-refreshed history from its own outputs; only WF is deployed.Source: §3.2 · pp. 3–4

4 Wave Runtime

Wave Runtime implements Wave Parallelism by coordinating full-model replicas, early K/V delivery, and overlapping computation. Local operator optimization and pipelined VAE decoding complete the path to sustained output.

4 Wave Runtime

  • Wave Runtime uses stage-owned full-model replicas to publish K/V early and overlap delivery with computation when cleaner producers run ahead.Wave Parallelism is the runtime mechanism for realizing this concurrency.Source: §4 · p. 4, §4.1 · p. 4

4.1 Stage placement and dependencies

  • Attention waits for the store and all legal cleaner ranks, while WF’s causal mask removes the reverse dependencies on noisier ranks present in RF.The synchronized reference instead uses a per-layer all-rank K/V gather.Source: §4.1 · p. 4

4.2 Producer lead and asynchronous delivery

  • Cleaner ranks read fewer keys, permitting later-layer producer lead; delivery is fully hidden only when every edge’s cost fits the available lead.WaveRT publishes K/V after QKV projection and RoPE, then checks readiness immediately before attention.Source: §4.2 · p. 5
  • Paged delivery writes K/V directly into ordered attention slots, while K/V quantization reduces payloads and faster local operators alter producer–consumer progress.With weights, noise draws, positions, and local operators fixed, attention consumes the same K/V values as in synchronized execution.Source: §4.2 · p. 5, §4.3 · p. 5

4.3 Local computation and streaming decode

  • WaveRT balances contiguous VAE partitions across dedicated GPUs using profiled streaming times, while preserving request-local causal caches.The first decoded output still incurs pipeline-fill latency.Source: §4.3 · p. 5
  • Execution width depends on denoising-step count; logical K/V fan-out remains quadratic in stage count, and attention workloads vary across stages.Source: §4.3 · p. 5

5 Experiments

Experiments assess generation quality and motion, training ablations, end-to-end runtime, system components, and context-parallel comparisons. Wave Forcing balances quality, motion, and stability, while Wave Runtime and Wave Parallelism improve streaming efficiency.

5.1 Generation quality and motion

  • 5-step WF leads VBench Total (81.64) and Quality (84.40); on 100-second videos it reaches Dynamic Degree 91.45, Subject Preservation 97.10, and Camera Drift 0.37.Rolling Forcing has the lowest Camera Drift (0.17) but a long-video Dynamic Degree of 36.63; Causal Forcing’s high motion accompanies subject loss and camera drift.Source: §5.1 · pp. 6–7, Table 1 · p. 7

5.2 Training ablations

  • The full 4-step model balances Quality (80.89), Dynamic Degree (74.97), Subject Preservation (91.88), and Camera Drift (0.36).Skipping Stage 2 drops Quality from 80.89 to 76.68 and Dynamic Degree from 74.97 to 15.73; removing training branches also trades motion against subject stability.Source: Table 2 · p. 7, §5.2 · pp. 7–8

5.4 System ablations

  • Paged has the lowest tick p50 in all six groups; Sync-to-Paged lowers BF16 tick p50 by 14.2% at 1.3B and 10.9% at 14B.The largest adjacent reduction varies by group; at 14B BF16 it is Overlap-to-One-sided.Source: §5.4 · p. 8, Figure 7 · p. 9

5.3 End-to-end inference efficiency

  • At 1.3B, WaveRT + Sage reaches 134.79 E2E FPS versus 147.78 DiT FPS in the same runs; the best single-GPU-relative E2E gains are 7.07× at 1.3B and 6.08× at 14B.The tests use dummy DiT weights; the eight-GPU setup allocates four GPUs to denoising, one to store, and three to VAE decoding.Source: §5.3 · p. 8, Figure 7 · p. 9

5.5 Comparison with traditional context parallelism

  • WP leads BF16 at five of six scales (WP/CP ratios 0.985–1.862×) and Sage+W8A8 at all six (1.025–2.530× the faster CP baseline).BF16 throughput tests use 5 steps/6 GPUs for 1.3B–9B and 4 steps/5 GPUs for 14B; the Sage+W8A8 measurements use the same matched-backend comparisons.Source: §5.5 · pp. 8–9, Figure 7 · p. 9

6 Related Work

Related work spans streaming-video training and inference, blockwise generation in other domains, parallel execution, and video evaluation. WF is positioned around training video chunks to use cleaner intermediate context and assessing long-stream stability beyond benchmark scores.

Efficient streaming video generation

  • Prior streaming-video work addresses exposure bias, historical K/V supervision, distillation, mixed-noise denoising, and long-horizon context, motivating joint study of training and inference context.Examples include Self Forcing [Huang et al., 2025a], SGF [Zhuang et al., 2026], and Helios [Yuan et al., 2026].Source: §6 · p. 9

Multi-chunk generation

  • Blockwise diffusion spans language, action, and multimodal generation; WF specifically trains video chunks to consume cleaner intermediate context through a directed causal wavefront.Examples include D2F [Wang et al., 2025], BlockVLA [Wang et al., 2026c], and Multimodal Diffusion Forcing [Huang et al., 2025c].Source: §6 · pp. 9–10

Parallel execution

  • Ulysses and Ring partition individual forward passes, whereas WP schedules chunk–stage tasks on full-model replicas.The related work also situates WP alongside pipelined execution systems and optimized attention runtimes.Source: §6 · p. 10

Video evaluation

  • VBench scores can coexist with subject disappearance and camera drift, so the paper supplements benchmark scores with targeted stability diagnostics.The discussion connects remaining methodological and system limits to future evaluation and interactive-generation directions.Source: §6 · p. 10

7 Limitations

The evaluation is limited to fixed conditioning and tested configurations, while quality diagnostics, output-path attribution, training complexity, and pipeline scaling leave important boundaries.

Evaluation scope

  • Streaming generation is evaluated only under fixed conditioning and the tested model, sampling, and hardware configurations.Online interaction and broader serving conditions are outside this evaluation scope.Source: §7 · p. 10

Quality evaluation and diagnostic coverage

  • VBench-based scores do not fully characterize long-stream quality, and the targeted diagnostics rely on a 100-prompt subject-focused subset and geometric heuristics.Their sensitivity to thresholds and missing observations, coverage of intentional scene changes, and agreement with human judgments remain unestablished.Source: §7 · p. 10

Output-path capacity and attribution

  • Aggregate timings do not isolate VAE compute, transfers, queueing, or backpressure, and system DiT timing may reflect decoder interference.The observed DiT/E2E gaps are consistent with output-path limits, but matched VAE-on/off runs and stage-level measurements are needed to identify bottlenecks.Source: §7 · p. 10

Training complexity

  • WF currently uses a complex multi-stage training procedure that the authors aim to simplify while preserving generation quality and long-video stability.Source: §7 · p. 10

Scalability at a fixed sampling budget

  • At a fixed denoising-step count, WF's pipeline has S denoising stages plus one cache-refresh stage, bounding stage parallelism and limiting single-pipeline scale.The experiments also show that WP's throughput advantage over context parallelism narrows for larger models.Source: §7 · p. 10

Interaction and serving scope

  • High steady-state throughput does not establish low input-to-visible-output latency, and fixed-context experiments do not evaluate online condition changes.Prompt switching, camera/action control, closed-loop tasks, and multi-stream deadline guarantees require dedicated evaluation.Source: §7 · p. 10

8 Future Work

The proposed agenda addresses execution and scaling, concurrent-stream serving, interactive conditions, world-model extensions, and broader quality evaluation.

End-to-end execution and scalability

  • The authors plan to profile decoder stages, transfers, queues, and backpressure, and study finer-grained execution or combinations with context parallelism.They also aim to simplify WF training while preserving motion and long-horizon stability.Source: §8 · p. 11

Serving concurrent streams

  • Future concurrent-stream work will study admission control, batching, and resource allocation while respecting context dependencies.Evaluation should include startup delay, output jitter, playback stalls, fairness, and quality alongside throughput.Source: §8 · p. 11

Interactive streaming generation

  • Interactive generation requires defining which active chunks receive prompt updates, refreshing affected state, and adapting training to condition changes.The authors propose evaluating prompt-to-visible-output latency, prompt adherence, and visual continuity.Source: §8 · p. 11

Interactive world models

  • Prospective interactive world-model directions include learned world prediction and systems where an engine maintains world state while a video model renders observations.Both routes require training for the control interface and evaluation of control alignment, state consistency, and closed-loop behavior.Source: §8 · p. 11

Richer quality evaluation

  • The authors aim to combine automatic diagnostics, understanding-based judgments, and human assessment for richer quality evaluation.Any automatic judge would require timestamped evidence, calibration against human annotations, and checks for missed failures.Source: §8 · p. 11

9 Conclusion

Wave Forcing jointly designs causal context, training, and multi-GPU execution for concurrent video-chunk denoising. The paper reports improved motion while maintaining overall generation quality, alongside practical streaming-efficiency gains.

Conclusion

  • Wave Forcing jointly designs causal context, training strategies, and multi-GPU pipelining to denoise multiple video chunks concurrently.Cleaner intermediate states of preceding chunks condition subsequent chunks, while full-model replicas produce one chunk per concurrent forward round after pipeline fill.Source: §9 · p. 11

AI Use Statement

AI assistants supported conceptual framing, manuscript organization, methodology discussion, interpretation of available evidence, text editing, and figure development.

AI Use Statement

  • AI assistance supported framing, manuscript organization, methodology discussion, evidence interpretation, translation and editing, and figure development.The authors state that this assistance was not independent verification and that they remain responsible for the paper's claims, citations, and disclosures.Source: AI Use Statement · p. 12

A Training Implementation Details

This appendix specifies WaveRT’s training implementation, execution semantics, and evaluation protocols. It details how its wavefront is scheduled and measured, including limits from pipeline fill, resource allocation, and output-path work.

A Training Implementation Details

  • Training preserves each model’s denoising schedule; the 4-step model uses three stages, while the 5-step model starts from released RF weights and proceeds directly to Stage 2.Stage 2 re-noises consecutive teacher chunks at staggered timesteps and applies DMD and LPIPS to the same clean prediction; Stage 3 samples RF, WF, or SF rollouts.Source: A Training Implementation Details · p. 17

B.1 Process ownership and reference execution

  • WaveRT assigns each denoising stage and the clean-state store to a full DiT replica, while dedicated GPUs own contiguous VAE decoder partitions.The 4-step allocation is 4+1+3 GPUs and the 5-step allocation is 5+1+2; store refreshes clean-state K/V rather than performing another denoising update.Source: §B.1 · pp. 17–18

B.2 K/V publication and buffer lifetime

  • WaveRT publishes layer K/V to legal cleaner-stage consumers and waits for readiness immediately before attention, placing remote data directly into fixed attention slots.Generation flags distinguish successive buffer uses, and outgoing copies must finish before a round boundary permits slot reuse.Source: §B.2 · p. 18, §B.3 · p. 18

B.3 Round schedule and cache semantics

  • The context order is anchor, bounded clean history, live store, self, then cleaner denoising slots; the live store page becomes retained history after each round.The first clean chunk remains a persistent anchor, while other pages reuse their stored RoPE-applied keys.Source: §B.3 · pp. 18–19

B.4 Output cadence, work, and fill/drain

  • The filled pipeline can emit faster than serial one-step inference, but it uses more replicas, does not reduce work or first-output latency, and may not fill for short requests.For 399 latent frames (133 chunks), S=4 gives 97.1% slot occupancy; requests with Nchunk ≤ S never fill the pipeline.Source: §B.4 · p. 19

B.5 Streaming VAE partitioning

  • VAE partitions are chosen to balance estimated local service, and the output interval includes decoder service and residual delivery not already overlapped.Slower decoding can grow queues or trigger bounded-buffer backpressure; the model excludes fill/drain and file encoding.Source: §B.5 · p. 19

B.6 Same-graph semantic preservation

  • With fixed model, schedule, inputs, and history rules, dependency-respecting asynchronous and synchronized execution produce the same latent sequence when attention receives the ordered legal K/V context.The timing contract includes producer readiness and delivery waits; concurrent durations should not be summed because that double-counts overlap.Source: §B.6 · pp. 19–20, §B.7 · p. 20

B.8 Bounded attention storage and communication

  • Attention-page storage is independent of generated video length, while steady-state K/V publication requires multiple logical payload destinations across the wavefront.The bound covers attention pages, not model weights or other request state; physical link traffic depends on transport and topology.Source: §B.8 · p. 20, Table 3 · p. 20

C.3 Runtime implementation and measurement protocol

The runtime measurements separate DiT execution from end-to-end video output and compare communication strategies under specified configurations. They report output-path constraints, same-graph correctness checks, and throughput comparisons whose scope depends on backend, steps, and GPU allocation.

C.3 Runtime implementation and measurement protocol

  • WaveRT measurements use one eight-H200 node, three fixed repetitions, and 399 latent frames after warmup; DiT timing includes fill/drain and may include concurrent VAE effects.The DiT-only parallel-strategy sweep disables the VAE, and the execution tests use dummy DiT weights rather than measuring generation quality.Source: §C.3 · pp. 22–23, Table 6 · p. 23

C.4 Throughput accounting and repeatability

  • In the 5-step, 8-GPU setup, BF16 and SageAttention reach 104.82 and 104.80 DiT FPS, while end-to-end rates are 96.46 and 96.48 FPS, respectively.The setup allocates five GPUs to denoising, one to Store, and two to a two-stage VAE; aggregate measurements do not isolate backpressure or establish a large hidden SageAttention speedup.Source: §C.4 · p. 23

C.5 Mechanism diagnostics

  • Synchronized, one-sided, and paged execution of the same 1.3B, 4-step WF graph produced bit-identical latent outputs in earlier checks.The fixed-compute 4-step, 14B tick p50 falls from 464.270 ms for Sync to 413.795 ms for Paged, with intermediate strategies reported separately.Source: §C.5 · p. 24, §C.6 · p. 24

C.6 Exposed communication, synchronization, and payload size

  • WaveRT targets exposed consumer-side waits through early K/V delivery, point-of-use readiness, and paged placement; reduced payload or assembly costs do not directly determine end-to-end impact.The impact depends on whether transfer time is exposed, and quantization can change execution timing.Source: §C.6 · pp. 24–25

D.1 Generation schedules and conditioning

  • WF preserves current-layer live K/V from legal cleaner chunks and store, distinguishing its conditioning from methods that expose noisier chunks or use different cache rules.Self Forcing completes one chunk before starting the next, while Rolling Forcing jointly advances mixed-noise chunks and retains active-window K/V dependencies in both temporal directions.Source: §D.1 · p. 25

D.2 Partitioning a forward versus distributing stages

  • The matched DiT-only comparison controls parameter scale, backend tier, GPU count, and denoising steps within each scale, but compares WP’s WF mask with CP’s RF active-window attention.WP relies on the WF graph and sufficient producer lead, at the cost of full-model replication; the comparison therefore combines execution strategy and attention visibility.Source: §D.2 · pp. 25–26, Table 10 · p. 26

E Extended Future Work

The extended-work discussion identifies interaction, serving, world-model integration, and quality evaluation as directions not yet evaluated with WF. Each would require additional specifications or validation beyond the fixed-context generation setting.

E Extended Future Work

  • WF has not been evaluated for the proposed interaction, serving, world-model, or understanding-assisted extensions.The paper identifies these extensions as future work following its stated training and execution limits.Source: E Extended Future Work · p. 27
  • For changing prompts during generation, an extension must specify prompt assignment for active chunks, cached-state refresh, and whether outstanding work completes or is discarded.The fixed-context equivalence argument does not establish correctness across condition changes; evaluation should measure time to visible response as well as adherence and continuity.Source: E Extended Future Work · p. 27
  • Deadline-aware serving remains to be studied, including admission control, stream priorities, resource allocation, and placement under live-context dependencies.Evaluation should add startup delay, output jitter, stall rate, and fairness alongside throughput and quality.Source: E Extended Future Work · p. 27
  • Integrating WF with controlled world evolution requires specifying when camera or action inputs reach in-flight chunks and validating dependencies and closed-loop control fidelity.The proposed assessment includes prediction consistency and policy performance in closed-loop tasks.Source: E Extended Future Work · p. 27
  • A renderer integration would need aligned structure/video training and temporal-conditioning validation; neither cited renderer establishes compatibility with WF’s wavefront.Evaluation should test whether rendered observations preserve underlying state and expose action effects quickly enough for feedback.Source: E Extended Future Work · p. 27
  • Broader quality evaluation could combine local defect checks with long-horizon, instruction-aware judgments, but requires human-calibrated validation and auditable temporal evidence.The proposed checks include agreement with human annotations and whether compressed memory loses brief defects or earlier facts.Source: E Extended Future Work · p. 27
Loading 2611.00859v1…