Sign in
Pending benchmarksTracking since Nov 3, 2026

Is Wave Forcing 1.3B state of the art?Pending benchmarks.Announced Oct 31, 2026; no independent results yet.

SJTU DENG Lab's streaming text-to-video diffusion model on the Wan2.1-T2V-1.3B architecture: later video chunks start denoising before earlier ones finish, so a multi-GPU pipeline emits one chunk per round. Open 4-step and 5-step checkpoints (Apache 2.0) with the WaveRT runtime code.

Announced facts

The details so far

Lab
SJTU DENG Lab
Announced
Oct 31, 2026
API
Not yet · None; self-hosted
Open weights
Published · Hugging Face SJTU-DENG-Lab/WaveForcing-1.3B, not gated; code at github.com/SJTU-DENG-Lab/WaveForcing
Parameters
1.3B class (1,418,996,800 BF16 parameters in each checkpoint; text encoder and VAE not included)
Type
Streaming, chunk-by-chunk text-to-video diffusion; overlapping chunks are denoised together under a causal attention mask
Base model
Wan2.1-T2V-1.3B; the 4-step model starts from Self Forcing's ODE initialization and the 5-step model from released Rolling Forcing weights
Checkpoints
4-step and 5-step EMA generators in native Wan format; text encoder, tokenizer and VAE load from Wan-AI/Wan2.1-T2V-1.3B-Diffusers
Training
Three stages: mixed-noise denoising, off-policy DMD for motion, then causal adaptation on self-generated history (the 5-step model skips stage 1)
Output
832x480 at 16 FPS in the paper's evaluation, including 100-second videos
Runtime
WaveRT (built on SGLang Diffusion's model layer) spreads denoising stages across full-model GPU replicas with one-way K/V exchange (Wave Parallelism), plus separate GPUs for VAE decoding
Earlier version
SJTU-DENG-Lab/WaveForcing-T2V-1.3B-5step-Preview: a different 5-step export, linked from the public project page by Aug 21, 2026, which the lab keeps for reference only and says not to use for evaluation
License
Apache 2.0 (weights and code)

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: the 4-step model reaches 147.6 DiT FPS and 134.8 end-to-end FPS on eight H200 GPUs, 7.07x the end-to-end throughput of single-GPU rolling inference.

Lab claim: the 5-step model keeps overall VBench quality comparable to Rolling Forcing (Total 81.64 vs 81.15) while raising Dynamic Degree from 51.39 to 94.44 on short videos and from 36.63 to 91.45 on 100-second videos.

Lab claim: in DiT-only tests with dummy weights, Wave Parallelism delivers up to 2.53x the throughput of the faster of Ulysses and Ring context parallelism across six model scales (1.3B to 14B).

Receipts

Sources

PAPERWave Forcing: Towards Speed-of-Light Streaming Video Generation (arXiv 2611.00859)arxiv.org/abs/2611.00859 · Oct 31, 2026LABDrew Jin (DENG Lab) announcement thread on Xx.com/drew_jin_0827/status/2117475319150256518 · Nov 3, 2026LABWave Forcing project pagesjtu-deng-lab.github.io/WaveForcing/LABProject page snapshot of Aug 21, 2026 linking the preview checkpoint (Wayback Machine)web.archive.org/web/20260821172804/https://sjtu-deng-lab.github.io/WaveForcing/REPOSJTU-DENG-Lab/WaveForcing-1.3B on Hugging Facehuggingface.co/SJTU-DENG-Lab/WaveForcing-1.3BREPOSJTU-DENG-Lab/WaveForcing on GitHubgithub.com/SJTU-DENG-Lab/WaveForcing

Tracking continues until an independent board lists it.