{"id":8112,"date":"2026-05-11T23:45:01","date_gmt":"2026-05-11T15:45:01","guid":{"rendered":"https:\/\/infernews.com\/blog\/?p=8112"},"modified":"2026-05-11T23:46:35","modified_gmt":"2026-05-11T15:46:35","slug":"official-implementation-of-quot-what-matters-for-diffusion-friendly-latent-manif","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/official-implementation-of-quot-what-matters-for-diffusion-friendly-latent-manif\/","title":{"rendered":"PAE\uff1a\u4ee4\u64f4\u6563\u6a21\u578b\u66f4\u5feb\u6536\u6582\u7684\u6f5b\u5728\u7a7a\u9593\u8a2d\u8a08"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/pae_teaser-f9d8136e4860.jpg\" alt=\"pae teaser\"><\/figure>\n<p>PAE \u662f\u4e00\u500b\u70ba\u6f5b\u5728\u64f4\u6563\u6a21\u578b\u800c\u8a2d\u7684\u81ea\u7de8\u78bc\u5668\u6846\u67b6\uff0c\u91cd\u9ede\u4e0d\u53ea\u662f\u628a\u5716\u7247\u58d3\u7e2e\u518d\u9084\u539f\uff0c\u800c\u662f\u5148\u628a\u6f5b\u5728\u7a7a\u9593\u6574\u7406\u6210\u66f4\u9069\u5408\u64f4\u6563\u6a21\u578b\u5b78\u7fd2\u7684\u5f62\u614b\u3002\u7c21\u55ae\u8b1b\uff0c\u5b83\u95dc\u5fc3\u7684\u4e0d\u662f\u300c\u58d3\u5f97\u975a\u5514\u975a\u300d\uff0c\u800c\u662f\u300c\u6a21\u578b\u4e4b\u5f8c\u751f\u5716\u6642\u6703\u5514\u6703\u66f4\u9806\u3001\u66f4\u7a69\u5b9a\u300d\u3002<\/p>\n<p>\u5be6\u969b\u4f7f\u7528\u4e0a\uff0cPAE \u53ef\u7406\u89e3\u70ba\u64f4\u6563\u6a21\u578b\u524d\u9762\u7684 tokenizer \u6216\u5f71\u50cf\u7de8\u78bc\u6a21\u7d44\uff1a\u5148\u628a\u5716\u7247\u8f49\u6210 latent\uff0c\u518d\u4ea4\u7540\u5f8c\u7e8c\u751f\u6210\u6a21\u578b\u8a13\u7df4\u3002\u9019\u7a2e\u505a\u6cd5\u7279\u5225\u9069\u5408\u672c\u8eab\u5df2\u5728\u505a latent diffusion\u3001\u4f46\u89ba\u5f97\u6536\u6582\u6162\u3001\u8a13\u7df4\u6210\u672c\u9ad8\uff0c\u6216\u8005\u751f\u6210\u6548\u679c\u672a\u5920\u7a69\u5b9a\u7684\u7814\u7a76\u8207\u958b\u767c\u6d41\u7a0b\u3002<\/p>\n<p>\u9019\u500b\u5c08\u6848\u8f03\u6709\u610f\u601d\u7684\u5730\u65b9\uff0c\u662f\u5b83\u660e\u78ba\u63d0\u51fa\u4e09\u500b\u300c\u5c0d\u64f4\u6563\u53cb\u5584\u300d\u7684\u6f5b\u5728\u7a7a\u9593\u7279\u8cea\uff1a\u7a7a\u9593\u7d50\u69cb\u4e00\u81f4\u6027\u3001\u5c40\u90e8\u6d41\u5f62\u9023\u7e8c\u6027\uff0c\u4ee5\u53ca\u5168\u57df\u8a9e\u610f\u7d44\u7e54\u3002\u4f5c\u8005\u4e0d\u662f\u5047\u8a2d\u9019\u4e9b\u7279\u8cea\u6703\u81ea\u7136\u51fa\u73fe\uff0c\u800c\u662f\u7528\u4e09\u7a2e prior-alignment regularization \u53bb\u4e3b\u52d5\u7d04\u675f\uff0c\u9019\u6bd4\u55ae\u9760\u91cd\u5efa\u8aa4\u5dee\u7684\u50b3\u7d71\u601d\u8def\u66f4\u9032\u4e00\u6b65\u3002<\/p>\n<p>\u6839\u64da\u5c08\u6848\u63d0\u4f9b\u7684\u7d50\u679c\uff0cPAE \u5728 ImageNet 256\u00d7256 \u4e0a\u505a\u5230 gFID 1.03\uff0c\u4e26\u4e14\u5728\u76f8\u540c LightningDiT \u8a2d\u5b9a\u4e0b\uff0c\u6536\u6582\u901f\u5ea6\u6700\u9ad8\u53ef\u6bd4 RAE \u5feb 13 \u500d\u3002\u5c0d\u975e\u7814\u7a76\u4eba\u54e1\u800c\u8a00\uff0c\u9019\u4ee3\u8868\u540c\u6a23\u8cc7\u6e90\u4e0b\u6709\u6a5f\u6703\u66f4\u5feb\u898b\u5230\u53ef\u7528\u6210\u679c\uff1b\u4e0d\u904e\u9019\u4e9b\u8868\u73fe\u4ecd\u61c9\u8996\u4e4e\u8cc7\u6599\u96c6\u3001\u8a13\u7df4\u8a2d\u5b9a\u8207\u9aa8\u5e79\u6a21\u578b\u800c\u5b9a\u3002<\/p>\n<ul>\n<li><strong>\u6838\u5fc3\u5b9a\u4f4d<\/strong>\uff1a\u70ba latent diffusion \u6e96\u5099\u66f4\u6613\u5b78\u7684\u6f5b\u5728\u8868\u793a<\/li>\n<li><strong>\u4e3b\u8981\u5275\u65b0<\/strong>\uff1a\u628a\u300c\u64f4\u6563\u53cb\u5584\u300d\u62c6\u6210\u4e09\u500b\u53ef\u512a\u5316\u7684\u6027\u8cea\u4f86\u8a13\u7df4<\/li>\n<li><strong>\u5be6\u969b\u50f9\u503c<\/strong>\uff1a\u6709\u6a5f\u6703\u7e2e\u77ed\u8a13\u7df4\u9031\u671f\uff0c\u63d0\u5347\u751f\u6210\u8cea\u7d20\u8207\u5c11\u6b65\u6578\u63a1\u6a23\u8868\u73fe<\/li>\n<li><strong>\u9069\u5408\u5834\u666f<\/strong>\uff1a\u5f71\u50cf\u751f\u6210\u7814\u7a76\u3001\u9700\u8981\u9ad8\u6548\u7387\u8a13\u7df4\u7684\u751f\u6210\u7cfb\u7d71\u3001\u6bd4\u8f03\u4e0d\u540c tokenizer \u8a2d\u8a08<\/li>\n<li><strong>\u53ef\u914d\u9aa8\u5e79<\/strong>\uff1a\u652f\u63f4\u591a\u7a2e\u7de8\u78bc\u5668\u65b9\u5411\uff0c\u5305\u62ec DINOv2\u3001SigLIP2\u3001DINOv3\u3001MAE<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u95dc\u5fc3\u7684\u662f\u300c\u5982\u4f55\u4ee4\u64f4\u6563\u6a21\u578b\u5b78\u5f97\u66f4\u5feb\uff0c\u800c\u5514\u4fc2\u53ea\u63db\u66f4\u5927\u6a21\u578b\u300d\uff0cPAE \u7684\u5207\u5165\u9ede\u76f8\u7576\u503c\u5f97\u53c3\u8003\u3002\u5b83\u6700\u9069\u5408\u6709\u4e00\u5b9a\u751f\u6210\u6a21\u578b\u6d41\u7a0b\u7684\u4eba\u63a1\u7528\uff1b\u5c0d\u4e00\u822c\u7528\u5bb6\u800c\u8a00\uff0c\u672a\u5fc5\u662f\u5373\u88dd\u5373\u7528\u5de5\u5177\uff0c\u4f46\u4f5c\u70ba\u4e0b\u4e00\u4ee3 latent tokenizer \u7684\u8a2d\u8a08\u601d\u8def\uff0c\u542b\u91d1\u91cf\u76f8\u7576\u9ad8\u3002<\/p>\n<p><strong>Source:<\/strong> <a href=\"https:\/\/github.com\/ZhengrongYue\/PAE\" rel=\"noopener noreferrer\">https:\/\/github.com\/ZhengrongYue\/PAE<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>PAE\u4e0d\u662f\u53ea\u8ffd\u6c42\u91cd\u5efa\u6e05\u6670\u5ea6\uff0c\u800c\u662f\u91cd\u65b0\u6574\u7406\u6f5b\u5728\u7a7a\u9593\u672c\u8eab\u3002\u5c0d\u60f3\u63d0\u5347\u751f\u6210\u8cea\u7d20\u8207\u8a13\u7df4\u6548\u7387\u7684\u4eba\uff0c\u9019\u500b\u65b9\u5411\u503c\u5f97\u7559\u610f\u3002<\/p>\n","protected":false},"author":8,"featured_media":8111,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[133,171,157,76,127,149,189],"tags":[],"class_list":["post-8112","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-157","category-76","category-127","category-149","category-189"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8112","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=8112"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8112\/revisions"}],"predecessor-version":[{"id":8114,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8112\/revisions\/8114"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8111"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8112"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8112"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8112"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}