
{"id":13277,"date":"2026-10-13T05:18:48","date_gmt":"2026-10-12T21:18:48","guid":{"rendered":"https:\/\/infernews.com\/blog\/representation-guidance\/"},"modified":"2026-10-13T05:18:48","modified_gmt":"2026-10-12T21:18:48","slug":"representation-guidance","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/representation-guidance\/","title":{"rendered":"Representation Guidance\uff08RG\uff09\uff1a\u4ee5\u4f4e\u7dad\u5f15\u5c0e Pixel Diffusion Transformer \u751f\u6210\u5f71\u50cf"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/Representation-Guidance-RG-Pixel-Diffusion-Transformer.jpg\" alt=\"Samples generated by RG\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u50cf\u7d20\u76f4\u63a5\u751f\u6210\u9ad8\u8cea\u7d20\u5f71\u50cf\uff0c\u5f80\u5f80\u8981\u5728\u8a9e\u7fa9\u7406\u89e3\u3001\u7a7a\u9593\u7d30\u7bc0\u8207\u6a21\u578b\u8a08\u7b97\u91cf\u4e4b\u9593\u53d6\u6368\u3002Representation Guidance (RG) \u5c6c\u65bc\u5f71\u50cf\u751f\u6210\u6a21\u578b\u65b9\u6cd5\uff0c\u91dd\u5c0d pixel diffusion transformers \u52a0\u5165\u53ef\u88ab\u751f\u6210\u6d41\u7a0b\u4e3b\u52d5\u5229\u7528\u7684\u5167\u90e8\u8868\u793a\uff0c\u540c\u6642\u4fdd\u7559 plain Transformer \u67b6\u69cb\uff0c\u8655\u7406\u8868\u793a\u5b78\u7fd2\u8207\u50cf\u7d20\u5408\u6210\u672a\u80fd\u6709\u6548\u929c\u63a5\u7684\u554f\u984c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">RG \u5148\u7528 low-dimensional bottleneck \u5c07\u8207\u751f\u6210\u76f8\u95dc\u7684\u8cc7\u8a0a\u96c6\u4e2d\u5230\u7dca\u6e4a\u8868\u793a\uff0c\u518d\u6309\u7167\u8a72\u8868\u793a\u5e36\u4f86\u7684 prediction offset\uff0c\u4fee\u6539 flow-matching target\uff0c\u76f4\u63a5\u5f15\u5c0e decoder \u751f\u6210\u3002\u5b83\u6c92\u6709\u4f9d\u8cf4\u5916\u90e8 autoencoder \u9810\u5148\u5efa\u7acb representation space\uff0c\u800c\u662f\u7531 raw pixels \u7aef\u5230\u7aef\u5b78\u7fd2\uff1b\u76f8\u8f03\u52a0\u5165\u5c08\u7528 pixel decoder \u7684\u67b6\u69cb\uff0cRG \u5617\u8a66\u907f\u514d\u53ea\u91cd\u65b0\u5206\u914d\u8a9e\u7fa9\u53ca\u7a7a\u9593\u5efa\u6a21\u5de5\u4f5c\uff0c\u4e26\u4ee5\u8f03\u7c21\u55ae\u7684 Transformer \u7d50\u69cb\u64f4\u5c55\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 ImageNet 256\u00d7256 \u6e2c\u8a66\u4e2d\uff0cRG-L\/16 \u4f7f\u7528 459M \u53c3\u6578\uff0c\u5728 600 epochs \u9054\u5230 1.79 FID\uff1bRG-XL\/16 \u4ee5 724.9M \u53c3\u6578\u53d6\u5f97 1.55 FID\u3001293.8 IS\u3002\u9805\u76ee\u4ea6\u5831\u544a 512\u00d7512 \u7684 1.74 FID\uff0c\u4e26\u5ef6\u4f38\u81f3 text-to-image\uff0c\u5728 GenEval \u53d6\u5f97 0.89\u3001DPG-Bench \u53d6\u5f97 87.1\uff1b\u9019\u4e9b\u7d50\u679c\u986f\u793a\u65b9\u6cd5\u9069\u5408\u7814\u7a76\u9ad8\u89e3\u6790\u5ea6\u50cf\u7d20\u751f\u6210\u53ca\u6587\u5b57\u751f\u5716\uff0c\u4f46\u8f03\u5927\u578b\u6a21\u578b\u4ecd\u9700\u8981\u76f8\u7576\u8a08\u7b97\u8cc7\u6e90\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u5f9e\u4ee5\u4e0b\u5e7e\u9ede\u7406\u89e3\u5b83\u7684\u5be6\u969b\u50f9\u503c\uff1a<br \/>\n&#8211; \u4f4e\u7dad bottleneck \u5c07\u751f\u6210\u76f8\u95dc\u8cc7\u8a0a\u96c6\u4e2d\uff0c\u6e1b\u5c11\u8868\u793a\u8207\u751f\u6210\u6d41\u7a0b\u5404\u81ea\u5b78\u7fd2\u7684\u843d\u5dee\u3002<br \/>\n&#8211; RG-L\/16 \u4ee5\u8f03\u5c11\u53c3\u6578\u8d85\u8d8a\u66f4\u5927\u578b\u7684 pixel diffusion transformers\u3002<br \/>\n&#8211; plain Transformer \u65b9\u4fbf\u7814\u7a76\u5167\u90e8\u8868\u793a\u5982\u4f55\u5f62\u6210\u53ca\u5f71\u97ff\u50cf\u7d20\u5408\u6210\u3002<br \/>\n&#8211; \u9805\u76ee\u8cc7\u6599\u63d0\u4f9b\u7d50\u679c\u8207\u6a21\u578b\u898f\u6a21\uff0c\u4f46\u672a\u4ea4\u4ee3\u5b8c\u6574\u5b89\u88dd\u3001\u4e0b\u8f09\u6216\u63a8\u7406\u6d41\u7a0b\uff0c\u6e2c\u8a66\u524d\u4ecd\u9700\u67e5\u770b Project Page \u6216\u5f8c\u7e8c\u7a0b\u5f0f\u78bc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f03\u9069\u5408\u7814\u7a76 pixel diffusion\u3001\u751f\u6210\u5f0f\u8868\u793a\u5b78\u7fd2\uff0c\u4ee5\u53ca\u5e0c\u671b\u6bd4\u8f03 flow matching \u67b6\u69cb\u7684\u5718\u968a\u3002\u5b83\u7684\u50f9\u503c\u4e0d\u5728\u65bc\u63d0\u4f9b\u5373\u6642\u53ef\u7528\u7684\u751f\u5716\u5de5\u5177\uff0c\u800c\u5728\u65bc\u628a\u300c\u5b78\u5230\u8868\u793a\u300d\u8207\u300c\u5229\u7528\u8868\u793a\u751f\u6210\u300d\u653e\u5165\u540c\u4e00\u500b\u8a13\u7df4\u6a5f\u5236\uff1b\u6a21\u578b\u53c3\u6578\u548c GFLOPs \u4ecd\u7136\u504f\u9ad8\uff0c\u672a\u5fc5\u9069\u5408\u8cc7\u6e90\u6709\u9650\u7684\u500b\u4eba\u96fb\u8166\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/pan-zihao.github.io\/Representation-Guidance\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Pan-Zihao\/Representation-Guidance\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Representation Guidance \u628a\u751f\u6210\u6240\u9700\u8cc7\u8a0a\u58d3\u7e2e\u5230\u4f4e\u7dad\u8868\u793a\uff0c\u518d\u76f4\u63a5\u5f15\u5c0e\u50cf\u7d20\u751f\u6210\uff0c\u70ba\u5927\u578b pixel diffusion Transformer \u63d0\u4f9b\u66f4\u6709\u6548\u7387\u7684\u6539\u826f\u65b9\u5411\u3002<\/p>\n","protected":false},"author":8,"featured_media":13276,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,169,171,76,127,30,210,129,163,38],"tags":[],"class_list":["post-13277","post","type-post","status-publish","format-standard","hentry","category-133","category-169","category-171","category-76","category-127","category-image","category-worldactionmodel","category-txt2img","category-163","category-38"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13277","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=13277"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13277\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13276"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13277"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13277"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13277"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}