<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>Dynamic Concepts Personalization from Single Videos</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2502.14844">https://arxiv.org/abs/2502.14844</a> </li><li>SIGGRAPH 2025</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1750167701/6B18ED8F-174B-4435-B437-44D36C06D303.png" style="background-color:initial;max-width:min(100%,2510px);max-height:min(1276px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1750167701/6B18ED8F-174B-4435-B437-44D36C06D303.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2510" height="1276"> one-shot进行视频概念学习的方法。文章提出了一个概念,即从视频中学习到的概念不仅仅包含外观信息,还包含时序信息,称之为动态概念。因此学习动态概念需要对两种信息分别进行训练:方法也非常简单,首先对外观信息进行学习,训练LoRA,训练数据为视频帧进行打乱;然后再训练LoRA的第二个子矩阵,训练数据为原视频帧。从而对动态概念进行了学习,并可以使用相应的token进行组合生成。 <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1750167896/306E894B-7184-4415-922D-B0E4ED647D5B.png" style="background-color:initial;max-width:min(100%,2502px);max-height:min(846px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1750167896/306E894B-7184-4415-922D-B0E4ED647D5B.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2502" height="846">