<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题：</span>ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址：</span><a href="https://arxiv.org/abs/2502.04320">https://arxiv.org/abs/2502.04320</a> </li><li>ICML2025</li></ul> <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_da2aed87-0115-4877-b0b4-c2153acd9947/public" style="background-color:initial;max-width:min(100%,2312px);max-height:min(878px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_da2aed87-0115-4877-b0b4-c2153acd9947/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2312" height="878"> 作者提出了ConceptAttention，通过在推理时引入额外的概念token，并在attention计算中进行设计，完成了概念的显著图的提取，无需任何训练过程，可以完成图像分割等任务。 <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_21e7f68e-ca9e-4444-8cec-eac3e53e19fc/public" style="background-color:initial;max-width:min(100%,1126px);max-height:min(1276px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_21e7f68e-ca9e-4444-8cec-eac3e53e19fc/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="1126" height="1276">