基于扩散模型加速和感知优化的高效姿态驱动人体动作生成技术
打开文本图片集
关键词：视频生成；扩散模型；加速技术；图像超分；视频插帧；数字人；感知质量中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2025）10-009-2964-08doi：10.19734/j.issn.1001-3695.2025.03.0057
Efficient pose-driven human motion generation with diffusion model acceleration and perceptual optimization
Wang Jiasong，Zhou Lei†，Zhang Bo （Schoolof HealthScience&Engineering，UniversityofShanghaiforScience&Technology，Shanghai 2o93，China）
Abstract：Existing pose-conditioneddigital human videogeneration technologyfocusesonimproving thequalityof generated results，emphasizing visualrealismand motionsmoothness.However，this technologyoftenoverlookstheisseofslowgenerationspeed，liigitseiedeplontinal-teapplicais.Tddresstis，isorkpoosdausietionframework fordigital humans（DAF-DH）basedondifusionmodelacclerationand perceptualoptimization.The method tackled thehigh inferencelatencyandcomputationalcostof difusion-model-baseddigitalhumangeneration.Itadopteda thre-stageaccelerationstrategy toenhanceeficiencyandoptimizegenerationquality.Firstly，themethodoptimizedtheinferenceeffciencyofthedifusionmodelusing TensorRT.Secondly，itutilizedtheTensorRT-acelerateddiffusionmodel，combinedwithreducedinputresolutionand framesampling，toquicklygeneratelow-resolution，low-frame-rateinitialvideos. Thirdly，themethodincludedalightweight post-processing module.Thismoduleemployedsuper-resolutionandframe interpolationalgorithms toimprove thevideo’sresolutionandsmothness，enhancing thefinal generationquality.Aditionally，the method introducedasemanticfeatureconsistencylossfunctiontoimprovesubjectivevisual perception.Thiswork alsoconstructeda DH-Motion datasetcontaining1705 motionsequences to provideabenchmark forresearch.Experiments show that thisframeworkachievesa5xspedupcomparedtoMimicMotion.Thegenerationqualityisimproved：theLPIPSmetricdecreasesby0.O33，andtheFVDscorereducesby82.9.TheseresultsdemonstratethatDAF-DHefectivelyreduces inferencelatency，enhances generation quality，and issuitable for real-time digital human video generation applications.
Key words：video generation；difusion model；acceleration technique；image super-resolution；video frame interpolation; digital human；perceptual quality
0 引言
近年来，随着虚拟现实、增强现实以及人工智能技术的快速发展，基于扩散模型的数字人生成技术在多个领域得到了广泛应用[1，2]。（剩余20756字）
