
{"id":9367,"date":"2026-06-18T15:40:50","date_gmt":"2026-06-18T07:40:50","guid":{"rendered":"https:\/\/infernews.com\/blog\/icml-2026-the-official-implementation-of-paper-quot-unified-multimodal-autoregre\/"},"modified":"2026-06-18T15:42:28","modified_gmt":"2026-06-18T07:42:28","slug":"icml-2026-the-official-implementation-of-paper-quot-unified-multimodal-autoregre","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/icml-2026-the-official-implementation-of-paper-quot-unified-multimodal-autoregre\/","title":{"rendered":"UniAR \u7528\u4e00\u500b Transformer \u5305\u8fa6\u770b\u5716\u3001\u4f5c\u5716\u3001\u6539\u5716"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/teaser.jpg\" alt=\"teaser\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 Unified Multimodal Models\uff08UMMs\uff09\u591a\u6578\u6703\u628a\u5f71\u50cf\u7406\u89e3\u548c\u5f71\u50cf\u751f\u6210\u5206\u958b\u8655\u7406\uff0c\u5e38\u898b\u505a\u6cd5\u662f\u7528\u5169\u5957 visual tokenizers\u3002\u4f5c\u8005\u8a8d\u70ba\u9019\u7a2e fixed paradigm \u6703\u628a\u8868\u793a\u7a7a\u9593\u62c6\u958b\uff0c\u6a21\u578b\u751f\u6210\u5b8c\u5716\u7247\u5f8c\uff0c\u9084\u8981\u518d\u91cd\u65b0\u7de8\u78bc\u624d\u80fd\u7406\u89e3\u81ea\u5df1\u525b\u7522\u751f\u7684\u5167\u5bb9\uff0cshared context \u4e5f\u5c31\u96e3\u4ee5\u771f\u6b63\u6210\u7acb\uff1bUniAR \u56e0\u6b64\u63d0\u51fa\u4e00\u500b unified autoregressive framework\uff0c\u7528\u55ae\u4e00 discrete visual tokenizer \u9023\u63a5\u7406\u89e3\u3001\u751f\u6210\u8207\u7de8\u8f2f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u5c6c\u65bc<strong>\u591a\u6a21\u614b\u6a21\u578b<\/strong>\uff0c\u76ee\u6a19\u662f\u7528\u540c\u4e00\u500b Transformer \u89e3\u6c7a image understanding\u3001image generation \u548c image editing \u4e4b\u9593\u4f86\u56de\u5207\u63db\u7684\u6210\u672c\u3002\u5b83\u7684\u6838\u5fc3\u5224\u65b7\u5f88\u76f4\u63a5\uff1a\u82e5\u6a21\u578b\u770b\u5716\u8207\u4f5c\u5716\u5171\u7528\u540c\u4e00\u5957\u8996\u89ba token\uff0c\u6d41\u7a0b\u5c31\u4e0d\u9700\u8981\u984d\u5916 re-encoding\uff0c\u7cfb\u7d71\u7d50\u69cb\u6703\u66f4\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6280\u8853\u4e0a\uff0cUniAR \u6709\u5e7e\u500b\u8fa8\u8b58\u5ea6\u5f88\u9ad8\u7684\u8a2d\u8a08\u3002Multi-level BSQ tokenizer \u628a\u9ad8\u5c64\u8a9e\u610f\u8207\u4f4e\u5c64\u7d30\u7bc0\u4e00\u4f75\u4fdd\u7559\uff0c\u4e26\u900f\u904e Binary Spherical Quantization \u64f4\u5927\u6709\u6548 vocabulary\uff1bparallel bitwise prediction \u5247\u628a\u8996\u89ba\u78bc\u4ee5\u5206\u7d44\u65b9\u5f0f\u4e00\u9f4a\u9810\u6e2c\uff0c\u4ee4 autoregressive \u9577\u5e8f\u5217\u58d3\u77ed\uff0c\u8ad6\u6587\u63d0\u5230 1024&#215;1024 \u5f71\u50cf\u53ea\u9700 256 \u500b AR tokens\uff0c\u5c0d\u61c9 32x visual compression ratio\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u55ae\u4e00 discrete visual tokenizer \u53d6\u4ee3\u96d9 tokenizer \u67b6\u69cb<\/li>\n\n\n\n<li>\u652f\u63f4 image understanding\u3001image generation\u3001image editing \u540c\u6a21\u904b\u4f5c<\/li>\n\n\n\n<li>Multi-level BSQ tokenizer \u540c\u6642\u9867\u53ca\u8a9e\u610f\u8207\u7d30\u7bc0<\/li>\n\n\n\n<li>parallel bitwise prediction \u58d3\u7e2e\u8996\u89ba\u5e8f\u5217\uff0c\u52a0\u5feb autoregressive \u751f\u6210<\/li>\n\n\n\n<li>DiT-based visual decoder \u4ee5 discrete visual tokens \u91cd\u5efa\u9ad8\u4fdd\u771f\u5f71\u50cf<\/li>\n\n\n\n<li>\u9700\u6c42\uff1aPython 3.12\u3001CUDA 12.1+\u3001\u63a8\u7406\u7684 GPU \u8a18\u61b6\u9ad4 >= 24 GB<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u60f3\u8a66\u9019\u500b\u9805\u76ee\uff0c\u8f03\u5408\u7406\u7684\u5207\u5165\u9ede\u4e0d\u662f\u76f4\u63a5\u62ff\u4f86\u7576\u65e5\u5e38\u5de5\u5177\uff0c\u800c\u662f\u5148\u770b\u5b83\u516c\u958b\u7684\u6a21\u578b\u6b0a\u91cd\u8207\u9805\u76ee\u9801\uff0c\u5206\u958b\u6e2c\u7406\u89e3\u3001\u751f\u6210\u3001\u7de8\u8f2f\u4e09\u985e\u8f38\u51fa\u662f\u5426\u4e00\u81f4\u3002\u5b83\u8f03\u9069\u5408\u7814\u7a76\u591a\u6a21\u614b\u7d71\u4e00\u67b6\u69cb\u7684\u4eba\u3001\u95dc\u6ce8 Qwen \u751f\u614b\u7684\u958b\u767c\u8005\uff0c\u4ee5\u53ca\u60f3\u6bd4\u8f03 autoregressive \u8207 diffusion \u6df7\u5408\u8def\u7dda\u7684\u8b80\u8005\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6027\u80fd\u65b9\u9762\uff0c\u539f\u6587\u8072\u7a31 UniAR \u7d93 large-scale pre-training\u3001supervised fine-tuning \u548c reinforcement learning \u5f8c\uff0c\u5728 image generation \u8207 image editing \u9054\u5230 state-of-the-art\uff0c\u540c\u6642\u5728\u591a\u6a21\u614b\u7406\u89e3 benchmark \u4fdd\u6301\u7af6\u722d\u529b\u3002\u4e0d\u904e\u76ee\u524d\u516c\u958b\u8cc7\u8a0a\u8f03\u50cf\u7814\u7a76\u6210\u679c\u5c55\u793a\uff0cvisual decoder training code \u4ecd\u672a\u5b8c\u6574\u653e\u51fa\uff0c\u56e0\u6b64\u66f4\u9069\u5408\u62ff\u4f86\u7406\u89e3\u65b9\u6cd5\u8ad6\uff0c\u800c\u4e0d\u662f\u7acb\u5373\u8a55\u4f30\u6210\u6210\u719f\u751f\u7522\u5de5\u5177\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u95dc\u6a21\u578b\u8207\u7d44\u4ef6\u5305\u62ec SD3-medium visual decoder\u3001Qwen Team \u80cc\u666f\u4e0b\u7684\u591a\u6a21\u614b\u6a21\u578b\u8def\u7dda\uff0c\u4ee5\u53ca\u8ad6\u6587\u805a\u7126\u7684 Unified Multimodal Models\uff08UMMs\uff09\u3002\u82e5\u4f60\u5728\u610f\u7684\u4e0d\u662f\u55ae\u6b21\u751f\u6210\u6548\u679c\uff0c\u800c\u662f\u6a21\u578b\u80fd\u5426\u300c\u7406\u89e3\u81ea\u5df1\u751f\u6210\u7684\u5167\u5bb9\u300d\uff0cUniAR \u7684 shared context \u8a2d\u8a08\u78ba\u5be6\u63d0\u51fa\u4e86\u4e00\u500b\u6709\u610f\u601d\u800c\u4e14\u76f8\u7576\u5177\u9ad4\u7684\u7b54\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/ShareLab-SII\/UniAR\" rel=\"noopener noreferrer\">https:\/\/github.com\/ShareLab-SII\/UniAR<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.18249\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.18249<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>UniAR \u91dd\u5c0d\u591a\u6a21\u614b\u6a21\u578b\u5e38\u898b\u7684\u96d9 tokenizer \u5206\u88c2\u554f\u984c\uff0c\u6539\u7528\u55ae\u4e00\u8996\u89ba tokenizer\u3002\u9019\u500b\u65b9\u5411\u540c\u6642\u7784\u6e96\u7406\u89e3\u3001\u751f\u6210\u8207\u7de8\u8f2f\u4e09\u7a2e\u5f71\u50cf\u4efb\u52d9\u3002<\/p>\n","protected":false},"author":8,"featured_media":9366,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,171,185,33,30,38,162,159,119,157,120,76,197],"tags":[],"class_list":["post-9367","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-qwen","category-stablediffusion","category-image","category-38","category-ai-productions","category-vibe-coding","category-119","category-157","category-120","category-76","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9367","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=9367"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9367\/revisions"}],"predecessor-version":[{"id":9369,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9367\/revisions\/9369"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9366"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9367"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9367"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9367"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}