摘要
当前Stable diffusion等人工智能绘画模型在绘画时难以直接控制图像风格,同时风格模型训练仅针对单种风格。针对该问题,提出了一种基于美学梯度法的人工智能风格化绘画系统,以实现多种图像风格的控制和融合,并提供更加便捷的图像创作体验。收集并分析网络用户数据,结合问卷得到用户对图像风格的感性需求;根据感性需求收集各风格图像数据得到对应的风格图像训练集。使用梯度下降算法计算风格化文本编码器的权重,实现生成图像风格化的效果。通过可用性测试对比用户对该系统与传统人工智能绘画系统产出图像的风格满意程度,结果表明:人工智能风格化绘画系统的平均满意度相较传统人工智能绘画系统提升23%,表明人工智能风格化绘画系统在图像风格生成上具有更好的效果,可满足用户对图像风格的需求。该人工智能风格化绘画系统可以更便捷地实现图像风格调整,允许用户直观选择不同风格的权重,便捷使用一种或多种风格,能够有效满足用户对图像风格设计的需求。
At present it is difficult for artificial intelligence painting models such as Stable diffusion to directly control image style in painting.At the same time,current style model training is focused on a single style.To address this issue,an artificial intelligence stylized painting system based on the aesthetic gradient method was proposed.It aimed to achieve control and integration of multiple image styles,and to provide a more convenient image creation experience.It collected and analyzed network user data and employed a questionnaire-based approach to obtain the user′s perceptual needs for image style.Furthermore,it collected the data of each style image according to the perceptual requirements to obtain the corresponding style image training set.It also used the gradient descent algorithm to calculate the weights of the stylized text encoder to achieve the effect of generating image stylization.A usability test was conducted to compare user satisfaction with the image styles produced by the traditional artificial intelligence painting system and the artificial intelligence stylized painting system.The results show that the average satisfaction of the latter is 23%higher than that of the former,indicating that artificial intelligence stylized painting system has better effects in image style generation and can effectively meet users′needs for image styles.This artificial intelligence stylized painting system can realize image style adjustment more easily,allow users to intuitively choose the weight of different styles and easily use one or more styles,and can effectively meet users′needs for image style design.
作者
钟梓锐
梁玲琳
ZHONG Zirui;LIANG Lingin(School of Art and Design,Zhejiang Sci-Tech University,Hangzhou 310018,China)
出处
《浙江理工大学学报(自然科学版)》
2024年第4期537-547,共11页
Journal of Zhejiang Sci-Tech University(Natural Sciences)
基金
国家社会科学基金青年项目(22CXW024)。