<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>Imagic: Text-Based Real Image Editing with Diffusion Models</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2210.09276">https://arxiv.org/abs/2210.09276</a> </li><li>CVPR 2023</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1734440618/2F57F633-E702-42CD-B462-E3E3C44C771C.png" style="background-color:initial;max-width:min(100%,2366px);max-height:min(910px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1734440618/2F57F633-E702-42CD-B462-E3E3C44C771C.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2366" height="910"> 文章目标是实现对单张真实图像的non-rigid(可改变布局)编辑。 方法非常简单,总共分为三步:1、将目标prompt作为初始值,微调prompt embedding,使得该prompt引导生成原真实图像。2、固定优化后的embedding,微调整个扩散模型,使得更好还原真实图像。3、用优化后的embedding与目标prompt embedding进行线性组合,得到编辑后的图像。 此外,还开源了一个基于文本进行真实图像编辑的benchmark用于评测对比。 <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1734441028/109A935B-9E66-4389-BA08-1E757891BD94.png" style="background-color:initial;max-width:min(100%,2342px);max-height:min(1194px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1734441028/109A935B-9E66-4389-BA08-1E757891BD94.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2342" height="1194"> <ul class="dashed" data-apple-notes-indent-amount="0"><li>数据:单张真实图像微调</li><li>指标:User Study</li><li>硬件:1 A100</li><li>开源:benchmark开源</li></ul>