基于稳定扩散与自适应增强技术的服装模特图像生成方法中

打开文本图片集

关键词：稳定扩散；图像生成；自适应增强；模型微调；多模态评估 中图分类号：TP181 文献标志码：A 文章编号：1001-3695（2025）08-003-2267-07 doi：10.19734/j.issn.1001-3695.2025.01.0008

Method for generating clothing model images based on stable diffusion and adaptive enhancement techniques

Liu Dawei 1，2 ，Yu Bihui 1，2‡ ，Shi Jiawei1，2，Wei Jingxuan 1，2 ，Shi Huiyang2，³，Jin Hexuan ， Sun Linzhuang1，2 （1.Shenyanguteofomputinghlg，CnseAcdefSiecs，hyag8，ha;2.UniersitofCinsedf Sciences，Bjing4，hin;3Sholofompuercience&Techog，UniersityfCneseAdeyfSiences，Beiin10 China）

Abstract：Withtheadvancementofcomputervisionand generativemodels，，image generationtechnologyhas made significant strides，particularlyine-commerceproductdisplays，enhancinguserinteraction.Realisticclothing modelgenerationhas becomeaninnovativeapplication，deeplyintegratinggenerativetechnologywithe-commerce.However，challngesremain，especiallyingenerating high-quality，realisticclothing imagesthatcapturedetails，texture，andconsistency.Currentmodelsoftenstrugglewithaccuratelyrepresentingthefactualconsistencyofclothingandmaintaining naturalnessandcoherencecompared toreal images.To improvetheperformanceofclothing model generation technologyine-commerceapplications，this studypresentedLoRA-DAE，animprovedstabledifusion generative modelthatintegratedLoRAforoptimizedweightadjustment in atentionand convolutionlayers.Additionall，itaddedanadaptiveenhancement module tothegenerationprocess， dynamicallyadjusting textureanddetail distribution，adressing issues liketexture blurrngandedge distortion.Experments show thatLoRA-DAEoutperforms mainstream methods ontheFashion Mannequindataset，achieving notableimprovements in perceived quality（user evaluation），quantitative metrics（FID，IS，PSNR，SSIM），and multi-modal large model VQA evaluation.

Key words：stable difusion;image generation；adaptive enhancement；model fine-tuning；multimodal evaluation

0 引言

随着数字化技术的快速发展，时尚行业正迎来深刻的变革，传统的服装设计、展示和销售方式逐步被人工智能等新兴技术所革新[1]。（剩余18415字）