<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>Null-text Inversion for Editing Real Images using Guided Diffusion Models</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2211.09794">https://arxiv.org/abs/2211.09794</a> </li><li>CVPR 2023</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1731314061/E530115A-AA83-4E5D-AF12-286116A65488.png" style="background-color:initial;max-width:min(100%,1188px);max-height:min(1538px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1731314061/E530115A-AA83-4E5D-AF12-286116A65488.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="1188" height="1538"> 这篇文章的目的就是为了将prompt-to-prompt的方法运用到真实图像上。 普通的DDIM反转对于classifier-free引导的文生图模型来说,重建的图像质量很差,因此文章提出了该方法。首先将图像使用DDIM反转到接近高斯分布作为起始点,然后进行去噪时,将classifier-free引导的无条件部分的text替换为可学习的embedding,然后将DDIM反转的原轨迹作为监督训练这些(每个t对应单独的embedding)embedding,从而完成了高质量的文生图重建,从而可以将prompt-to-prompt方法运用于真实图像。 <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1731314761/DF67A5F0-CF32-440C-863C-BB62124FF5C6.png" style="background-color:initial;max-width:min(100%,2362px);max-height:min(1062px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1731314761/DF67A5F0-CF32-440C-863C-BB62124FF5C6.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2362" height="1062"> <ul class="dashed" data-apple-notes-indent-amount="0"><li>开源:<a href="https://github.com/google/prompt-to-prompt">https://github.com/google/prompt-to-prompt</a> </li></ul>