EditorNodesPricingBlog

WaveForcing & MovieRoll: Open Source Long Video AI

November 18, 2026
WaveForcing & MovieRoll: Open Source Long Video AI

Peking University & Tencent Hunyuan MovieRoll (arXiv:2611.21339) and SJTU DENG Lab WaveForcing (arXiv:2611.00859, Apache 2.0)

Share this post:

WaveForcing & MovieRoll: Open Source Long Video AI

WaveForcing (SJTU-DENG-Lab/WaveForcing, arXiv:2611.00859) and MovieRoll (Tencent-Hunyuan/MovieRoll, arXiv:2611.21339) are two November 2026 open source frameworks engineered to generate continuous, minute scale long takes without the color burn, contrast drift, or motion freezing that plague autoregressive video rollouts. Released under the Apache 2.0 license with 4-step and 5-step WaveForcing-1.3B weights on Hugging Face alongside MovieRoll's 5-step distilled SGLang Diffusion and vLLM-Omni pipelines, the two releases bring low latency streaming long takes to modern GPU workflows.

Studio availability. WaveForcing and MovieRoll are not available in AI FILMS Studio. This article covers public research releases from November 2026. To generate or extend long video sequences in the Studio, use WAN 3.0, Seedance 2.5 Video Extend, MiniMax H3 Video Extend, or FLUX 3 Video Extend on the Text to Video page or in the video workspace.

WaveForcing overview demonstrating wavefront pipelined streaming generation (unmute for audio)

Continuous 60 second single take rollout maintaining stable exposure and subject geometry

Why Autoregressive Video Models Drift After 10 Seconds

When an autoregressive video generator predicts chunk t+1 conditioned on its own previously decoded frames from chunk t, tiny quantization errors and high frequency noise accumulate exponentially. By second 15 or 20, colors oversaturate, textures turn plastic, and characters freeze in place.

Wait times compound the quality problem. Standard Diffusion Forcing schedules denoise chunks sequentially, leaving GPUs underutilized and forcing directors to wait minutes to inspect a 60 second blocking test.

WaveForcing (arXiv:2611.00859) and MovieRoll (arXiv:2611.21339, released November 12, 2026) solve streaming throughput and long horizon stability through complementary mechanisms.

  1. Wavefront pipelined causal scheduling (WaveForcing). SJTU DENG Lab formulates causal video diffusion as a diagonal wavefront across frame chunks and noise levels. In its 4-step and 5-step WaveForcing-1.3B checkpoints (built on a native Wan 1.3B bfloat16 EMA backbone), adjacent temporal chunks denoise concurrently at staggered noise levels (5+2 pipeline schedule), reaching real time streaming throughput without inter chunk seam flicker.
  2. Joint window denoising and hierarchical KV caching (MovieRoll). Developed by Peking University and Tencent Hunyuan, MovieRoll introduces a systematic pipeline from a 50 step supervised MovieRoll-Base model (using AnnealCFG across arbitrary 480p aspect ratios) to a 5-step CFG free MovieRoll-Distilled generator (480x864 and 864x480). Its hierarchical key value (KV) cache retains anchor scene tokens while compressing mid range history, stopping error accumulation across extended rollouts.
  3. Day-0 SGLang Diffusion and vLLM-Omni serving. MovieRoll ships native inference backends for both SGLang Diffusion (sglang_movieroll) and vLLM-Omni (vllm_omni_movieroll), allowing production studios to serve long take video generation behind standard high throughput inference engines.
MovieRoll systematic long form video rollout architecture from supervised base to 5 step distilled model

MovieRoll training and inference architecture combining AnnealCFG base training with 5-step CFG free distillation. Credit: Peking University & Tencent Hunyuan (arXiv:2611.21339).

In standard sliding window rollouts, hard boundaries between windows create visible temporal pops every few seconds.

MovieRoll's joint window denoising overlaps transition latents during both supervised training and 5-step distillation so camera pans and character walks cross window boundaries invisibly.

Try WaveForcing Live in the Browser

You can explore WaveForcing directly in the embedded Hugging Face Space below (akhaliq-wave-forcing.hf.space), which loads without login.

Side by side qualitative comparisons across 60 second continuous takes confirm that `WaveForcing-1.3B` avoids the progressive darkening and background collapse seen in earlier causal forcing baselines.

Because `WaveForcing-1.3B` uses standard `Wan` `model.safetensors` formatting under the **Apache 2.0** license, studios can drop the `4step/` or `5step/` weights straight into existing local `Wan 1.3B` pipelines.

WaveForcing 1.3B qualitative long video comparison showing stable color and facial identity over time

Long horizon frame stability in WaveForcing-1.3B compared against baseline causal rollouts. Credit: SJTU DENG Lab (Apache 2.0).

Checkpoints, Serving Backends & License Matrix

Model Variant Hugging Face ID Denoising Steps & Guidance Resolution & Serving Engine License
WaveForcing-1.3B (4-Step) SJTU-DENG-Lab/WaveForcing-1.3B (4step/) 4 steps (--rf-step 4, BF16 EMA) 480p (832x480), single GPU or pipelined multi GPU Apache 2.0
WaveForcing-1.3B (5-Step) SJTU-DENG-Lab/WaveForcing-1.3B (5step/) 5 steps (--rf-step 5, 5+2 schedule) 480p (832x480), 60s+ continuous long take streaming Apache 2.0
MovieRoll-Distilled (T2V) yunyangge/MovieRoll (distilled/t2v) 5 steps (No CFG, CFG=1.0) 480x864 / 864x480, SGLang Diffusion & vLLM-Omni Tencent Hunyuan Community
MovieRoll-Base (T2V & I2V) yunyangge/MovieRoll (base/) 50 steps (AnnealCFG) 480p arbitrary aspect ratios, T2V & I2V rollouts Tencent Hunyuan Community

Running WaveForcing-1.3B and MovieRoll Locally

To run WaveForcing-1.3B under Apache 2.0, clone SJTU-DENG-Lab/WaveForcing and download either the 4step/ or 5step/ native Wan checkpoint from Hugging Face.

git clone https://github.com/SJTU-DENG-Lab/WaveForcing.git
cd WaveForcing
pip install -r requirements.txt

# Download 4-step and 5-step WaveForcing-1.3B checkpoints
huggingface-cli download SJTU-DENG-Lab/WaveForcing-1.3B --local-dir ./ckpts/WaveForcing-1.3B

# Run 5-step streaming long-take video generation
python generate_waveforcing.py \
  --gen-ckpt ckpts/WaveForcing-1.3B/5step/model.safetensors \
  --rf-step 5 \
  --prompt "A continuous tracking shot following a cyclist riding along a coastal cliff road at golden hour, steady camera." \
  --num_frames 1440 \
  --output outputs/coastal_60s_take.mp4

To run MovieRoll-Distilled (5 steps, no CFG) or MovieRoll-Base with SGLang Diffusion or vLLM-Omni, clone Tencent-Hunyuan/MovieRoll and pull yunyangge/MovieRoll.

git clone https://github.com/Tencent-Hunyuan/MovieRoll.git
cd MovieRoll
pip install -r requirements.txt

huggingface-cli download yunyangge/MovieRoll --local-dir ./ckpts/MovieRoll

# Launch 5-step CFG-free long-form T2V rollout via SGLang Diffusion backend
python -m sglang_movieroll.infer \
  --model_path ./ckpts/MovieRoll/distilled/t2v \
  --prompt "An unbroken Steadicam shot moving through a bustling night market under paper lanterns, 864x480 widescreen." \
  --height 480 \
  --width 864 \
  --steps 5 \
  --output outputs/night_market_rollout.mp4

Production Playbook for Filmmakers

  1. Use WaveForcing-1.3B for commercial Apache 2.0 previsualization. Because WaveForcing-1.3B carries an unrestricted Apache 2.0 license and runs on a compact 1.3B Wan backbone, indie studios can generate 60 second blocking and pacing tests locally, then upscale selected cuts to 1080p or 4K using SeedVR2 or Real-ESRGAN Video.
  2. Pair MovieRoll-Base I2V with MovieRoll-Distilled T2V. When a scene must start from a specific concept art plate or storyboard frame, initialize the rollout with MovieRoll-Base in image-to-video mode (AnnealCFG), or use the 5-step MovieRoll-Distilled checkpoint for fast widescreen (864x480) or vertical (480x864) previsualization.
  3. Compare across long-video architectures. Explore our companion deep dives on History-Forcing and LightX2V long video inference, SkyReels V2 DF 1.3B Diffusion Forcing, LongCat Video 15-minute extended duration generation, and Apple FastFrames 2-step interactive video diffusion.

To generate extended 30 second shots with WAN 3.0 or chain Seedance 2.5 Video Extend and MiniMax H3 Video Extend in the cloud without local setup, open the Text to Video page or connect sequential video extension and upscaling nodes visually in the Nodes Graph Editor.


Sources