sapiСистемный анализ и прикладная информатика«System analysis and applied information science»2309-49232414-0481Belarusian National Technical University10.21122/2309-4923-2026-1-49-53sapi-798Research ArticleОбработка информации и принятие решенийData processing and decision–makingСравнительный анализ методов оптимизации нейронных сетейComparative analysis of neural network optimization methodsГаврикД. Н.HaurykD. N.

Гаврик Дмитрий Николаевич – аспирант. 

г. Минск

+375293952030povt@bntu.bydr3952030@icloud.com

Dzmitry N. Hauryk – PhD applicant.

Minsk

+375293952030povt@bntu.bydr3952030@icloud.com

povt@bntu.byФронтПоинтБеларусьBelarusian National Technical University, FrontPointBelarus202610042026014953Copyright © Гаврик Д.Н., 20262026Гаврик Д.Н.Hauryk D.N.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://sapi.bntu.by/jour/article/view/798

Проведено прикладное сравнительное тестирование ускорений вывода Stable Video Diffusion (image-to-video). Для всех методов использован фиксированный вход и параметры (1024×576, 25 кадров), базовый вариант FP16/25 шагов. Описано восемь сравниваемых подходов: стандартный запуск SVD в FP16, INT8 weight-only квантование UNet, torch.compile+TF32, снижение шагов, подстановка дистиллированных весов (AnimateLCM), 2:4 разреженность, LCM-режим, а также генерация ключевых кадров с последующей интерполяцией RIFE. Измерялись время и VRAM, качество / плавность оценивались прокси-метриками CLIP similarity, tSSIM, tLPIPS. Ключевые кадры+RIFE дает наибольшее ускорение при сохранении сильной привязки к исходнику, LCM обеспечивает сбалансированное ~2× ускорение; агрессивное снижение шагов ухудшает динамику.

We benchmark practical ways to accelerate Stable Video Diffusion (SVD) inference for image-to-video. All methods use a fixed setup (1024×576 input, 25 frames) with an FP16 baseline at 25 denoising steps. We compare eight techniques: UNet INT8 weight-only quantization, torch.compile+TF32, step reduction, distilled weights (AnimateLCM), semi-structured 2:4 sparsity, LCM mode/scheduler, and keyframe generation with RIFE interpolation as post-processing. We measure latency and peak VRAM, and track quality/motion via CLIP similarity, tSSIM, and tLPIPS. Keyframes+RIFE achieves the highest speedup while preserving strong conditioning to the input. LCM provides a balanced ~2× speedup, whereas aggressive step cuts (and untuned 2:4) can degrade motion.

Stable Video Diffusion (SVD)диффузионные моделиimage-to-videoускорение выводауменьшение числа шагов диффузииINT8 weight-only квантование (UNet)torch.compileTF32дистиллированные весаполуструктурная разреженность 2:4LCM-режим (scheduler)ключевые кадрыинтерполяция кадров RIFECLIP similaritytSSIMtLPIPSStable Video Diffusion (SVD)diffusion modelsimage-to-videoinference accelerationdenoising step reduction (num_inference_steps)UNet INT8 weight-only quantizationtorch.compileTF32distilled weights (AnimateLCM)semi-structured 2:4 sparsityLCM mode/schedulerkeyframesRIFE frame interpolationCLIP similaritytSSIMtLPIPSReferencesHo J., Jain A., Abbeel P. Denoising diffusion probabilistic models. arXiv:2006.11239; 2020. https://doi.org/10.48550/arXiv.2006.11239.Ho J., Jain A., Abbeel P. Denoising diffusion probabilistic models. arXiv:2006.11239; 2020. https://doi.org/10.48550/arXiv.2006.11239.Blattmann A., Dockhorn T., Kulal S., Mendelevitch D., Kilian M., Lorenz D., et al. Stable video diffusion: scaling latent video diffusion models to large datasets. arXiv:2311.15127; 2023. https://doi.org/10.48550/arXiv.2311.15127.Blattmann A., Dockhorn T., Kulal S., Mendelevitch D., Kilian M., Lorenz D., et al. Stable video diffusion: scaling latent video diffusion models to large datasets. arXiv:2311.15127; 2023. https://doi.org/10.48550/arXiv.2311.15127.Radford A., Kim J.W., Hallacy C., Ramesh A., Goh G., Agarwal S., et al. Learning transferable visual models from natural language supervision. arXiv:2103.00020; 2021. https://doi.org/10.48550/arXiv.2103.00020.Radford A., Kim J.W., Hallacy C., Ramesh A., Goh G., Agarwal S., et al. Learning transferable visual models from natural language supervision. arXiv:2103.00020; 2021. https://doi.org/10.48550/arXiv.2103.00020.Wang Z., Bovik A.C., Sheikh H.R., Simoncelli E.P. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP). 2004;13(4):600–612. https://doi.org/10.1109/TIP.2003.819861.Wang Z., Bovik A.C., Sheikh H.R., Simoncelli E.P. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP). 2004;13(4):600–612. https://doi.org/10.1109/TIP.2003.819861.Zhang R., Isola P., Efros A.A., Shechtman E., Wang O. The Unreasonable effectiveness of deep features as a perceptual metric. Computer Vision and Pattern Recognition (CVPR); 2018. https://doi.org/10.48550/arXiv.1801.03924.Zhang R., Isola P., Efros A.A., Shechtman E., Wang O. The Unreasonable effectiveness of deep features as a perceptual metric. Computer Vision and Pattern Recognition (CVPR); 2018. https://doi.org/10.48550/arXiv.1801.03924.PyTorch. TorchAO Quantization API. [Documentation]. Available at: https://docs.pytorch.org/docs/stable/quantization-support.html (accessed 19 January 2026).PyTorch. TorchAO Quantization API. [Documentation]. Available at: https://docs.pytorch.org/docs/stable/quantization-support.html (accessed 19 January 2026).PyTorch. torch.compile. [Documentation]. Available at: https://docs.pytorch.org/docs/stable/generated/torch.compile.html (accessed 19 January 2026).PyTorch. torch.compile. [Documentation]. Available at: https://docs.pytorch.org/docs/stable/generated/torch.compile.html (accessed 19 January 2026).Luo S., Tan Y., Huang L., Li J., Zhao H. Latent consistency models: synthesizing high-resolution images with few-step inference. arXiv:2310.04378; 2023. https://doi.org/10.48550/arXiv.2310.04378.Luo S., Tan Y., Huang L., Li J., Zhao H. Latent consistency models: synthesizing high-resolution images with few-step inference. arXiv:2310.04378; 2023. https://doi.org/10.48550/arXiv.2310.04378.NVIDIA. cuSPARSELt: A high-performance CUDA library for sparse matrix-matrix multiplication. [Documentation]. Available at: https://docs.nvidia.com/cuda/cusparselt/ (accessed 19 January 2026).NVIDIA. cuSPARSELt: A high-performance CUDA library for sparse matrix-matrix multiplication. [Documentation]. Available at: https://docs.nvidia.com/cuda/cusparselt/ (accessed 19 January 2026).Huang Z., Zhang T., Heng W., Shi B. Zhou S. RIFE: Real-Time intermediate flow estimation for video frame interpolation. Computer Vision and Pattern Recognition (CVPR); 2022. https://doi.org/10.48550/arXiv.2011.06294.Huang Z., Zhang T., Heng W., Shi B. Zhou S. RIFE: Real-Time intermediate flow estimation for video frame interpolation. Computer Vision and Pattern Recognition (CVPR); 2022. https://doi.org/10.48550/arXiv.2011.06294.

The authors declare that there are no conflicts of interest present.