Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

MagCache: Fast Video Generation with Magnitude-Aware Cache

Authors: Zehong Ma, Longhui Wei, Feng Wang, Shiliang Zhang, Qi Tian

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Mag Cache achieves 2.10 2.68 speedups on Open-Sora, Cog Video X, Wan 2.1, and Hunyuan Video, while preserving superior visual fidelity. It significantly outperforms existing methods in LPIPS, SSIM, and PSNR, under similar computational budgets.
Researcher Affiliation Collaboration Zehong Ma1,2, , Longhui Wei2, , , Feng Wang2, Shiliang Zhang1, , Qi Tian2 1 State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University 2Huawei Inc.
Pseudocode No The paper describes the Mag Cache method in Section 3.3, detailing error modeling and adaptive caching strategy with equations. Figure 2 provides an overview flowchart, but there are no explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor is the procedure presented in a structured code-like format.
Open Source Code Yes Codes: https://github.com/Zehong-Ma/Mag Cache
Open Datasets Yes We utilize the Vbench benchmark for evaluation, which is open access.
Dataset Splits No In the ablation study, we randomly sample 100 prompts from VBench to conduct our experiments.
Hardware Specification Yes Latency is measured on a single A800 GPU.
Software Dependencies No We enable Flash Attention [85] by default for all experiments.
Experiment Setup Yes For Open-Sora, we set K = 3 and δ = 0.12 for Mag Cache-fast, and K = 1, δ = 0.06 for Mag Cache-slow. For Wan 2.1, Mag Cache-fast uses K = 4 and δ = 0.12, while Mag Cache-slow uses K = 2 and δ = 0.12. For all models, following prior works [86, 87], we keep the first 20% of diffusion steps unchanged, as these initial steps are critical to the overall generation process.