SCOPE

Not every frame deserves full computation.
Training-free tri-modal acceleration for autoregressive video diffusion.

Overview · Method · Quick Start · Backends · Citation · License

Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
Hanshuai Cui, Zhiqing Tang, Zhi Yao, Fanshuai Meng, Weijia Jia, and Wei Zhao
Accepted at ACM Multimedia 2026 · arXiv:2604.02979

Overview

SCOPE is a training-free inference accelerator for autoregressive video diffusion. Instead of making a binary choice between reusing a stale result and running the full model, SCOPE introduces a third PREDICT path. It combines temporal skipping, lightweight output extrapolation, and active-frame selective computation to reduce redundant denoising work without retraining the model.

This release contains a reusable method core and reviewable adapters for MAGI-1 4.5B-distill and SkyReels-V2 DF-1.3B. Model weights, datasets, generated videos, and complete upstream repositories are not redistributed.

Method

At every frame or chunk denoising step, SCOPE accumulates a lightweight feature difference and selects one of three execution paths:

risk < cache_threshold                         -> CACHE
risk < predict_threshold and history is ready -> PREDICT
otherwise                                      -> RECOMPUTE

PREDICT extrapolates the velocity or noise output from recent full evaluations in scheduler sigma space. Prediction decay, a maximum consecutive-skip cap, threshold-boundary guards, periodic correction, and a reset after recomputation limit drift. For SkyReels-V2, selective computation also restricts each model call to the active frame interval derived from the diffusion-forcing step_matrix.

Quick Start

1. Install the method core

git clone https://github.com/hsc113/SCOPE.git
cd SCOPE
python3 -m venv .venv
source .venv/bin/activate
pip install -e '.[test]'
pytest

2. Install and validate an upstream backend

Install the selected model environment separately, then verify its checkout before a GPU run:

scope-check skyreels-v2 /path/to/SkyReels-V2
scope-check magi-1 /path/to/MAGI-1

Reviewed upstream revisions and sentinel hashes are recorded in provenance.json. A matching sentinel permits compatible later MAGI revisions; an upstream API change should be reviewed instead of bypassing the check.

Backend Integration

SkyReels-V2 DF-1.3B

Clone the reviewed upstream revision and install its requirements:

git clone https://github.com/SkyworkAI/SkyReels-V2.git
git -C SkyReels-V2 checkout 9351d13152207cc04de780e055346b08ade0b851
export SKYREELS_V2_ROOT=$PWD/SkyReels-V2
scope-check skyreels-v2 "$SKYREELS_V2_ROOT"

Minimal smoke command (downloads or loads the upstream model and requires a CUDA GPU):

python integrations/skyreels_v2/run_scope.py \
  --scope --model_id Skywork/SkyReels-V2-DF-1.3B-540P \
  --resolution 540P --num_frames 49 --base_num_frames 49 \
  --inference_steps 30 --seed 0 \
  --policy_profile manual \
  --cache_threshold 0.18 --predict_threshold 0.42 \
  --taylor_order 2 --history_length 3 --warmup_steps 2 \
  --gradient_clip 2.0 --use_ret_steps \
  --predict_decay 0.75 --max_skip_steps 5 \
  --enable_selective_compute --selective_compute_max_interval_ratio 1.0 \
  --enable_sensitivity_guard --sensitivity_guard_band 0.03 \
  --sensitivity_ema_decay 0.80 \
  --prompt "A paper boat drifting across a quiet pond." \
  --outdir outputs/skyreels_smoke

MAGI-1 4.5B-distill

Use the MAGI checkout and model/runtime JSON described by MAGI-1. The research source record is 6ff822e74ded; runtime compatibility is verified against 0fcefdef8ce2df37a3b8890979433c06eb003328 or a matching sentinel.

export MAGI1_ROOT=/path/to/MAGI-1
export PYTHONPATH="$MAGI1_ROOT:$PWD/integrations/magi1:$PYTHONPATH"
scope-check magi-1 "$MAGI1_ROOT"
python integrations/magi1/run_scope.py \
  --config_file /path/to/4.5B_distill_runtime.json \
  --additional_config configs/magi1_paper.yaml \
  --mode t2v \
  --prompt "A paper boat drifting across a quiet pond." \
  --output_path outputs/magi_smoke.mp4

Repository Layout

src/scope_accel/             reusable predictor, policies, guards, stats
integrations/skyreels_v2/    pipeline-level SkyReels adapter and runner
integrations/magi1/          SampleTransport/chunk adapter and runner
configs/                     paper defaults and a lightweight smoke profile
tests/                       CPU unit and static integration tests

Citation

Citation metadata is also provided in CITATION.cff.

@misc{cui2026framesdeservecomputationaccelerating,
  title={Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation},
  author={Hanshuai Cui and Zhiqing Tang and Zhi Yao and Fanshuai Meng and Weijia Jia and Wei Zhao},
  year={2026},
  eprint={2604.02979},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.02979}
}

License

Author-owned SCOPE core code is released under Apache-2.0. Third-party adapter exceptions and reviewed upstream revisions are documented in NOTICE, THIRD_PARTY.md, and each integration directory.

The root license does not relicense SkyReels-V2 integration code, upstream repositories, or model weights. Review the adjacent Skywork Community License and every upstream model license before redistribution or use.