Wave Forcing Triples Long-Video Motion Quality Over Rolling Forcing
Shanghai Jiao Tong University's DENG Lab introduces Wave Forcing, a pipelined denoising schedule that raises 100-second video dynamics from 36.63 to 91.45 on VBench, achieves 81.64 overall score, and delivers 7x throughput on 8 H200 GPUs while preserving subject consistency at 97.10.
Shanghai Jiao Tong University's DENG Lab presents Wave Forcing , a new inference-time scheduling strategy for long video diffusion models that dramatically improves motion dynamics while maintaining subject consistency.
Core Idea: Staggered Denoising Pipeline
Traditional chunk-based video diffusion (Rolling Forcing) processes each chunk sequentially: fully denoise a chunk, run a forward pass to write clean latents into the history cache, then start the next chunk. This creates a two-forward-pass gap between adjacent chunks, limiting motion coherence.
Wave Forcing overlaps these steps like a diagonal wave: older chunks continue later denoising stages while newer chunks begin earlier stages. A new chunk only reads history that is already cleaner than itself, never noisier. Once the pipeline is full, each concurrent forward pass produces one chunk, and every denoising stage plus the final cache refresh occupies a full model replica.
Training Procedure
Stage 1 – Mixed-noise distillation: Train the student on examples where chunks at different noise levels appear together, so the model learns to handle the staggered noise distribution.
Stage 2 – Teacher-guided distillation: Use a teacher video to supervise both distribution matching and perceptual quality; perceptual loss weight is 0.25 .
Stage 3 – Self-distillation: The model generates its own history and continues distillation. Rolling Forcing and Self Forcing appear only in this stage; deployment uses pure Wave Forcing. The 5-step student initializes from the already-released Rolling Forcing weights, skipping Stage 1.
Quantitative Results (Wan2.1-T2V-1.3B base, 832×480, 16 fps)
100-second video dynamics (VBench): Wave Forcing 91.45 vs Rolling Forcing 36.63 (2.5× improvement).
Short video dynamics: from 51.39 to 94.44 .
VBench total score: 81.64 vs Rolling Forcing 81.15 .
Subject consistency: 97.10 vs Rolling Forcing 92.75 .
Camera drift (lower better): 0.37 vs Rolling Forcing 0.17 .
Throughput Benchmark (4-step, 8× H200)
DiT forward pass: 147.58 FPS .
End-to-end (including VAE decode): 134.81 FPS .
Single-card baseline (19.06 FPS) → 7.07× speedup .
GPU allocation: 4 GPUs denoising, 1 GPU cache refresh, 3 GPUs VAE decoding.
Measured with randomly initialized DiT weights to isolate scheduling throughput.
Practical Takeaway
For teams already using chunked video diffusion, the authors recommend: (1) only attend to cleaner history across chunk boundaries, (2) parallelize all denoising stages, and (3) switch the final training phase to self-generated history.
References
Paper: https://arxiv.org/html/2611.00859
Project page: https://sjtu-deng-lab.github.io/WaveForcing/
Code example
👀点击关注
AI前沿快速FollowSigned-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
