<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangSC-Regular'">文章标题:</span>Continual Diffusion: Continual Customization of Text-to-Image Diffusion with C-LoRA</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2304.06027">https://arxiv.org/abs/2304.06027</a> </li><li>TMLR 2024</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1726071528/30A1F930-536E-48A3-A5D5-1CFA9CEE307A.png" style="background-color:initial;max-width:min(100%,2334px);max-height:min(1184px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1726071528/30A1F930-536E-48A3-A5D5-1CFA9CEE307A.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2334" height="1184"> <span style="font-family: '.PingFangSC-Regular'"> 文章提出了一个新的问题,即通过多对象的图片序列微调定制化一个模型,从而使该模型同时具有生成多个对象的能力。</span> <span style="font-family: '.PingFangSC-Regular'"> 当前模型针对这种序列化定制的任务往往表现得不好,会出现灾难性遗忘的情况。为了防止遗忘,作者提出了一种新的方法</span>C-LoRA,由SD中的交叉注意力层的持续性自正则化LoRA组成。 <span style="font-family: '.PingFangSC-Regular'"> 具体来说,对于每个新的对象,学习一套新的</span>LoRA权重,模型最后的权重为每套LoRA相加(持续性),并引入自正则化机制防止遗忘。<span style="font-family: '.PingFangSC-Regular'">此外,每个对象的</span>embedding随机初始化,在推理时,将名词替换为特定对象的词。 <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1726071528/4EE43E6D-DF38-415B-8486-6F39CC1769A3.png" style="background-color:initial;max-width:min(100%,2368px);max-height:min(1122px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1726071528/4EE43E6D-DF38-415B-8486-6F39CC1769A3.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2368" height="1122"> <ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangSC-Regular'">数据:人脸(</span>Celeb-A HQ);地标建筑(Google Landmarks dataset v2)</li><li><span style="font-family: '.PingFangSC-Regular'">指标:图像对齐(</span>CLIP);两图片集的分布差异(MMD);</li><li><span style="font-family: '.PingFangSC-Regular'">硬件:</span>2 A100</li><li><span style="font-family: '.PingFangSC-Regular'">开源:未开源</span></li></ul>