{"id":1184777,"date":"2026-07-14T18:22:00","date_gmt":"2026-07-15T01:22:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1184777"},"modified":"2026-09-01T02:46:37","modified_gmt":"2026-09-01T09:46:37","slug":"icml-%e4%b8%8a%e6%96%b0-%e4%ba%94%e9%a1%b9%e8%a7%86%e8%a7%89%e7%a0%94%e7%a9%b6%ef%bc%8c%e8%ae%a9%e6%a8%a1%e5%9e%8b%e6%9b%b4%e6%8c%81%e4%b9%85%e3%80%81%e6%9b%b4%e8%bd%bb%e9%87%8f%e3%80%81%e6%9b%b4","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/icml-%e4%b8%8a%e6%96%b0-%e4%ba%94%e9%a1%b9%e8%a7%86%e8%a7%89%e7%a0%94%e7%a9%b6%ef%bc%8c%e8%ae%a9%e6%a8%a1%e5%9e%8b%e6%9b%b4%e6%8c%81%e4%b9%85%e3%80%81%e6%9b%b4%e8%bd%bb%e9%87%8f%e3%80%81%e6%9b%b4\/","title":{"rendered":"ICML \u4e0a\u65b0 | \u4e94\u9879\u89c6\u89c9\u7814\u7a76\uff0c\u8ba9\u6a21\u578b\u66f4\u6301\u4e45\u3001\u66f4\u8f7b\u91cf\u3001\u66f4\u7edf\u4e00\u3001\u66f4\u805a\u7126"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><strong>\u672c\u671f\u5185\u5bb9\u901f\u89c8\uff1a<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1.\u8ba9\u865a\u62df\u4e16\u754c\u544a\u522b\u9057\u5fd8\uff1a\u6784\u5efa\u5177\u6709\u6301\u4e45\u4e09\u7ef4\u8bb0\u5fc6\u7684\u4e16\u754c\u6a21\u578b<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2.CoD-Lite\uff1a\u57fa\u4e8e\u6269\u6563\u6a21\u578b\u7684\u5b9e\u65f6\u751f\u6210\u5f0f\u56fe\u50cf\u538b\u7f29<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">3.\u4e00\u4e2a\u5411\u91cf\uff0c\u88c5\u4e0b\u6574\u6bb5\u89c6\u9891\uff1a\u89c6\u89c9\u751f\u6210\u6a21\u578b\u91cc\u7684\u9690\u5f0f\u8bb0\u5fc6\u8868\u5f81<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">4.\u57fa\u4e8e\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u7684\u591a\u6a21\u6001\u9690\u5f0f\u8bed\u8a00\u5efa\u6a21<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">5.ViTL\uff1a\u878d\u5408\u4ea4\u9519\u5f0f\u63a8\u7406\u7684\u957f\u89c6\u9891\u95ee\u7b54<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"1-\u8ba9\u865a\u62df\u4e16\u754c\u544a\u522b\u9057\u5fd8-\u6784\u5efa\u5177\u6709\u6301\u4e45\u4e09\u7ef4\u8bb0\u5fc6\u7684\u4e16\u754c\u6a21\u578b\" class=\"wp-block-heading\">1.\u8ba9\u865a\u62df\u4e16\u754c\u544a\u522b\u9057\u5fd8\uff1a\u6784\u5efa\u5177\u6709\u6301\u4e45\u4e09\u7ef4\u8bb0\u5fc6\u7684\u4e16\u754c\u6a21\u578b<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"811\" height=\"217\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-1.png\" alt=\"title,image\" class=\"wp-image-1184778\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-1.png 811w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-1-300x80.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-1-768x205.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-1-240x64.png 240w\" sizes=\"auto, (max-width: 811px) 100vw, 811px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2603.03482\">https:\/\/arxiv.org\/abs\/2603.03482<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u6709\u7684\u4ea4\u4e92\u5f0f\u4e16\u754c\u6a21\u578b\u5e38\u7f3a\u4e4f\u5bf9\u73af\u5883\u7684\u4e09\u7ef4\u8868\u793a\uff0c\u53ea\u80fd\u4ece\u6570\u636e\u4e2d\u9690\u5f0f\u5730\u5b66\u4e60\u4e09\u7ef4\u4e00\u81f4\u6027\u3002\u800c\u5b83\u4eec\u7684\u7a7a\u95f4\u8bb0\u5fc6\u53d7\u9650\u4e8e\u6709\u9650\u7684\u65f6\u95f4\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u6240\u4ee5\u5728\u8fdb\u884c\u957f\u7a0b\u63a8\u6f14\u65f6\u5bb9\u6613\u4e22\u5931\u7a7a\u95f4\u4fe1\u606f\uff0c\u5bfc\u81f4\u4e00\u8fb9\u751f\u6210\u4e00\u8fb9\u9057\u5fd8\uff0c\u96be\u4ee5\u9002\u914d\u6c89\u6d78\u5f0f\u4ea4\u4e92\u548c\u667a\u80fd\u4f53\u8bad\u7ec3\u7b49\u4efb\u52a1\u3002&nbsp;&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u8fd9\u4e00\u74f6\u9888\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86PERSIST\u6846\u67b6\uff0c\u901a\u8fc7\u9690\u7a7a\u95f4\u4e2d\u7684\u4e09\u7ef4\u73af\u5883\u5e27\uff08world-frame\uff09\u663e\u5f0f\u5efa\u6a21\u73af\u5883\u968f\u65f6\u95f4\u6f14\u5316\u7684\u72b6\u6001\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6a21\u578b\u7ef4\u62a4\u5e76\u9884\u6d4b\u4e00\u4e2a\u4e09\u7ef4\u73af\u5883\u5e27\uff0c\u7528\u4e8e\u4fdd\u5b58\u6301\u4e45\u7684\u7a7a\u95f4\u8bb0\u5fc6\uff1b\u540c\u65f6\u9884\u6d4b\u76f8\u673a\u72b6\u6001\uff0c\u5e76\u6839\u636e\u5f53\u524d\u89c6\u89d2\u5c06\u4e09\u7ef4\u73af\u5883\u5e27\u4e2d\u7684\u4fe1\u606f\u6295\u5f71\u5230\u53ef\u89c2\u6d4b\u7684\u4e8c\u7ef4\u5c4f\u5e55\u7a7a\u95f4\uff1b\u6700\u540e\uff0c\u5229\u7528\u4e00\u4e2aworld-to-pixel\uff08\u4e16\u754c\u72b6\u6001\u5230\u50cf\u7d20\u753b\u9762\uff09\u751f\u6210\u6a21\u5757\uff0c\u6839\u636e\u6295\u5f71\u540e\u7684\u4e09\u7ef4\u73af\u5883\u7279\u5f81\u3001\u52a8\u4f5c\u4ee5\u53ca\u4e8c\u7ef4\u56fe\u50cf\u7279\u5f81\u5408\u6210\u5f53\u524d\u89c6\u9891\u5e27\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8bad\u7ec3\u9636\u6bb5\uff0cPERSIST&nbsp;\u4f7f\u75282D-VAE&nbsp;\u548c&nbsp;3D-VAE&nbsp;\u5206\u522b\u7f16\u7801\u89c6\u9891\u5e27\u548c\u4e09\u7ef4\u73af\u5883\u5e27\uff0c\u5176\u4e2d\u73af\u5883\u5e27\u9884\u6d4b\u5668\u548cworld-to-pixel&nbsp;\u751f\u6210\u5668\u5747\u91c7\u7528&nbsp;Rectified Flow&nbsp;\u6846\u67b6\uff0c\u76f8\u673a\u6a21\u578b\u5219\u901a\u8fc7\u56e0\u679cTransformer&nbsp;\u9884\u6d4b\u4f4d\u7f6e\u3001\u671d\u5411\u548c\u89c6\u573a\u53d8\u5316\u3002\u4e3a\u7f13\u89e3\u81ea\u56de\u5f52\u63a8\u7406\u4e2d\u7684\u8bef\u5dee\u7d2f\u79ef\uff0c\u6a21\u578b\u8fdb\u4e00\u6b65\u5f15\u5165&nbsp;Diffusion Forcing&nbsp;\u548c\u968f\u673a\u566a\u58f0\u589e\u5f3a\uff0c\u4f7f\u5404\u4e2a\u6a21\u5757\u5728\u8bad\u7ec3\u65f6\u80fd\u591f\u9002\u5e94\u63a8\u7406\u9636\u6bb5\u7531\u5176\u4ed6\u6a21\u5757\u751f\u6210\u7684\u9690\u7a7a\u95f4\u7279\u5f81\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"353\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-2.png\" alt=\"diagram\" class=\"wp-image-1184779\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-2.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-2-300x122.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-2-768x313.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-2-240x98.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1a\u4ee5\u5355\u4e2a\u89c6\u9891\u5e27\u521d\u59cb\u5316\u540e\uff0cPERSIST\u00a0\u4f1a\u6839\u636e\u7528\u6237\u52a8\u4f5c\u5728\u81ea\u56de\u5f52\u5faa\u73af\u4e2d\u6301\u7eed\u8fed\u4ee3\u4e16\u754c\u72b6\u6001\u3002\u6a21\u578b\u9996\u5148\u9884\u6d4b\u667a\u80fd\u4f53\u5468\u56f4\u7684\u4e09\u7ef4\u73af\u5883\u5e27\uff0c\u5e76\u7528\u4e00\u4e2a\u524d\u9988Transformer\u00a0\u9884\u6d4b\u76f8\u673a\u53c2\u6570\uff0c\u968f\u540e\u6839\u636e\u9884\u6d4b\u7684\u76f8\u673a\u53c2\u6570\u5c06\u4e09\u7ef4\u4e16\u754c\u6295\u5f71\u5230\u76f8\u673a\u5e73\u9762\uff0c\u5f97\u5230\u4e00\u7c07\u6309\u6df1\u5ea6\u6392\u5e8f\u7684\u4e16\u754c\u9690\u7279\u5f81\uff1b\u6700\u540e\uff0c\u6a21\u578b\u5229\u7528\u8fd9\u4e9b\u4e0e\u50cf\u7d20\u5bf9\u9f50\u7684\u4e09\u7ef4\u4fe1\u606f\uff0c\u5bf9\u50cf\u7d20\u9690\u7279\u5f81\u8fdb\u884c\u53bb\u566a\u5e76\u5408\u6210\u4e0b\u4e00\u89c6\u9891\u5e27\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u00a0Luanti\u00a0\u73af\u5883\u4e2d\u7684\u5b9e\u9a8c\u8868\u660e\uff0cPERSIST\u00a0\u5728\u957f\u7a0b\u4ea4\u4e92\u5f0f\u751f\u6210\u4e2d\u660e\u663e\u4f18\u4e8e\u00a0Oasis\u3001WorldMem\u00a0\u7b49\u57fa\u7ebf\u65b9\u6cd5\uff1a600\u00a0\u5e27\u751f\u6210\u4efb\u52a1\u4e2d\uff0cFVD\u00a0\u4ece\u00a0Oasis\u00a0\u7684\u00a0875\u00a0\u964d\u81f3148\uff1b\u7528\u6237\u8bc4\u4f30\u4e2d\uff0c\u4e09\u7ef4\u7a7a\u95f4\u4e00\u81f4\u6027\u548c\u65f6\u95f4\u7a33\u5b9a\u6027\u4e5f\u663e\u8457\u63d0\u5347\u3002\u5b9a\u6027\u7ed3\u679c\u663e\u793a\uff0cPERSIST\u00a0\u80fd\u5728\u6570\u5343\u5e27\u4ea4\u4e92\u63a8\u6f14\u4e2d\u4fdd\u6301\u573a\u666f\u7ed3\u6784\u8fde\u8d2f\uff0c\u5e76\u80fd\u5efa\u6a21\u89c6\u91ce\u5916\u73af\u5883\u7684\u6301\u7eed\u53d8\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"2-cod-lite-\u57fa\u4e8e\u6269\u6563\u6a21\u578b\u7684\u5b9e\u65f6\u751f\u6210\u5f0f\u56fe\u50cf\u538b\u7f29\" class=\"wp-block-heading\">2.CoD-Lite\uff1a\u57fa\u4e8e\u6269\u6563\u6a21\u578b\u7684\u5b9e\u65f6\u751f\u6210\u5f0f\u56fe\u50cf\u538b\u7f29<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"882\" height=\"239\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-3.png\" alt=\"title,image\" class=\"wp-image-1184780\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-3.png 882w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-3-300x81.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-3-768x208.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-3-240x65.png 240w\" sizes=\"auto, (max-width: 882px) 100vw, 882px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2604.12525\">https:\/\/arxiv.org\/abs\/2604.12525<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u9ad8\u6027\u80fd\u751f\u6210\u5f0f\u538b\u7f29\u9886\u57df\uff0c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u901a\u5e38\u4f9d\u9760\u5927\u5e45\u6269\u5145\u53c2\u6570\u91cf\u6765\u63d0\u5347\u89c6\u89c9\u91cd\u5efa\u6548\u679c\u3002\u8fd9\u4f1a\u5e26\u6765\u6781\u9ad8\u7684\u63a8\u7406\u5ef6\u8fdf\uff0c\u5f88\u96be\u6ee1\u8db3\u5404\u7c7b\u5b9e\u65f6\u843d\u5730\u573a\u666f\u3002\u5982\u4f55\u5e73\u8861\u6a21\u578b\u89c4\u6a21\u4e0e\u8fd0\u884c\u6548\u7387\uff0c\u8bbe\u8ba1\u8f7b\u91cf\u5316\u6269\u6563\u67b6\u6784\u662f\u5f53\u4e0b\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u7a81\u7834\u8fd9\u4e00\u5c40\u9650\uff0c\u7814\u7a76\u5458\u4eec\u5f00\u5c55\u4e86\u591a\u7ec4\u5b9e\u9a8c\u5206\u6790\uff0c\u901a\u8fc7\u5bf9\u6bd4\u4e0d\u540c\u53c2\u6570\u91cf\u7684\u6269\u6563\u6a21\u578b\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u7684\u4f18\u52bf\u3002\u5728\u8f7b\u91cf\u5316\u6a21\u578b\u4e0a\uff0c\u76f8\u6bd4\u901a\u7528\u751f\u6210\u6269\u6563\u9884\u8bad\u7ec3\u65b9\u6848\uff0c\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u80fd\u751f\u6210\u4fe1\u606f\u5bc6\u5ea6\u66f4\u9ad8\u7684\u56fe\u50cf\u8868\u5f81\uff0c\u53ef\u5f25\u8865\u5c0f\u6a21\u578b\u8868\u5f81\u80fd\u529b\u4e0d\u8db3\u7684\u7f3a\u9677\u3002\u7814\u7a76\u5458\u4eec\u540c\u65f6\u53d1\u73b0\uff0c\u56fe\u50cf\u538b\u7f29\u4efb\u52a1\u5e76\u4e0d\u9700\u8981\u5168\u5c40\u6ce8\u610f\u529b\uff0c\u8f7b\u91cf\u5316\u5377\u79ef\u7f51\u7edc\u642d\u914d\u84b8\u998f\u5b66\u4e60\uff0c\u5373\u53ef\u8fd8\u539f\u5b8c\u6574\u7684\u7cbe\u7ec6\u7eb9\u7406\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u4e8e\u8fd9\u4e9b\u53d1\u73b0\uff0c\u7814\u7a76\u5458\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u4e00\u6b65\u5f0f\u5377\u79ef\u6269\u6563\u56fe\u50cf\u7f16\u89e3\u7801\u5668\u3002\u8be5\u65b9\u6cd5\u91c7\u7528\u7d27\u51d1\u7684\u5377\u79ef\u4e3b\u5e72\u7f51\u7edc\uff0c\u901a\u8fc7\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u548c\u7edf\u4e00\u7684\u84b8\u998f\u4e0e\u5bf9\u6297\u8bad\u7ec3\u6846\u67b6\uff0c\u6210\u529f\u5b9e\u73b0\u4e86\u5728A100 GPU\u4e0a1080p\u5206\u8fa8\u7387\u4e0b60 FPS\u7684\u7f16\u7801\u901f\u5ea6\u4e0e42 FPS\u7684\u89e3\u7801\u901f\u5ea6\u3002\u540c\u65f6\uff0c\u8be5\u65b9\u6cd5\u5728\u4fdd\u6301\u4e0eMS-ILLM\u7b49GAN\u7c7b\u7f16\u89e3\u7801\u5668\u76f8\u5f53\u7684FID\u6307\u6807\u7684\u540c\u65f6\uff0c\u53ef\u8282\u770185%\u6bd4\u7279\u7387\uff0c\u5728\u63d0\u4f9b\u9ad8\u8d28\u91cf\u89c6\u89c9\u91cd\u5efa\u7684\u540c\u65f6\u6ee1\u8db3\u4e86\u4f4e\u5ef6\u8fdf\u7684\u5b9e\u65f6\u9700\u6c42\uff0c\u6709\u6548\u642d\u5efa\u4e86\u751f\u6210\u5f0f\u538b\u7f29\u4e0e\u5b9e\u9645\u90e8\u7f72\u573a\u666f\u4e4b\u95f4\u7684\u6865\u6881\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"586\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-4.jpg\" alt=\"diagram\" class=\"wp-image-1184781\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-4.jpg 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-4-300x203.jpg 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-4-768x520.jpg 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-4-240x163.jpg 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1a\u751f\u6210\u5f0f\u56fe\u50cf\u7f16\u89e3\u7801\u5668\u7684\u53d1\u5c55\u4e3b\u8981\u4f9d\u8d56\u4e8e\u6269\u5927\u6a21\u578b\u89c4\u6a21\uff0c\u8fd9\u5bfc\u81f4\u89e3\u7801\u5ef6\u8fdf\u663e\u8457\u589e\u52a0\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cCoD-Lite\u5728\u611f\u77e5\u8d28\u91cf\u4e0e\u89e3\u7801\u901f\u5ea6\u4e4b\u95f4\u5b9e\u73b0\u4e86\u66f4\u4f18\u7684\u5e73\u8861\uff1a\u80fd\u591f\u5728A100 GPU\u4e0a\u5b9e\u73b0\u5b9e\u65f61080p\u89e3\u7801\uff0c\u5e76\u540c\u65f6\u8fbe\u5230\u63a5\u8fd1\u5f53\u524d\u6700\u5148\u8fdb\u7684FID\u503c\u3002\u89e3\u7801\u5668\u53c2\u6570\u4e0e\u89e3\u7801\u65f6\u95f4\u5982\u56fe\u6240\u793a\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"3-\u4e00\u4e2a\u5411\u91cf-\u88c5\u4e0b\u6574\u6bb5\u89c6\u9891-\u89c6\u89c9\u751f\u6210\u6a21\u578b\u91cc\u7684\u9690\u5f0f\u8bb0\u5fc6\u8868\u5f81\" class=\"wp-block-heading\">3.\u4e00\u4e2a\u5411\u91cf\uff0c\u88c5\u4e0b\u6574\u6bb5\u89c6\u9891\uff1a\u89c6\u89c9\u751f\u6210\u6a21\u578b\u91cc\u7684\u9690\u5f0f\u8bb0\u5fc6\u8868\u5f81<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"841\" height=\"236\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-5.png\" alt=\"title,image\" class=\"wp-image-1184782\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-5.png 841w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-5-300x84.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-5-768x216.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-5-240x67.png 240w\" sizes=\"auto, (max-width: 841px) 100vw, 841px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2603.07615\">https:\/\/arxiv.org\/abs\/2603.07615<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3\u89c6\u89c9\u751f\u6210\u5927\u6a21\u578b\u7ecf\u8fc7\u6d77\u91cf\u6570\u636e\u8bad\u7ec3\uff0c\u6c89\u6dc0\u4e86\u4e30\u5bcc\u7684\u56fe\u50cf\u3001\u89c6\u9891\u5148\u9a8c\u77e5\u8bc6\u3002\u4f46\u50cf\u7d20\u3001\u9690\u53d8\u91cf\u3001\u56fe\u50cf\u6807\u8bb0\u8fd9\u7c7b\u4f20\u7edf\u89c6\u89c9\u8868\u5f81\u90fd\u72ec\u7acb\u4e8e\u751f\u6210\u6a21\u578b\uff0c\u65e0\u6cd5\u76f4\u63a5\u590d\u7528\u6a21\u578b\u5185\u7f6e\u77e5\u8bc6\uff0c\u5f88\u96be\u5b9e\u73b0\u8d85\u9ad8\u538b\u7f29\u500d\u7387\u4e0b\u7684\u9ad8\u6548\u5b58\u50a8\u4e0e\u5185\u5bb9\u590d\u7528\uff0c\u4f20\u7edf\u7f16\u7801\u65b9\u6848\u5728\u6781\u4f4e\u7801\u7387\u573a\u666f\u753b\u8d28\u8870\u51cf\u4e25\u91cd\u3002\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u9690\u5f0f\u89c6\u89c9\u8868\u5f81\u6846\u67b6\u3002\u8be5\u65b9\u6848\u4e0d\u518d\u5355\u72ec\u7f16\u7801\u753b\u9762\u50cf\u7d20\uff0c\u800c\u662f\u628a\u89c6\u89c9\u4fe1\u53f7\u5efa\u6a21\u4e3a\u751f\u6210\u51fd\u6570\uff0c\u4f9d\u9760\u51bb\u7ed3\u6269\u6563\u6a21\u578b\u914d\u5957LoRA\u53c2\u6570\u5b8c\u6210\u8868\u5f81\u5efa\u6a21\u3002\u50cf81\u5e27\u957f\u89c6\u9891\u8fd9\u7c7b\u590d\u6742\u89c6\u89c9\u5185\u5bb9\uff0c\u90fd\u80fd\u8fdb\u4e00\u6b65\u54c8\u5e0c\u538b\u7f29\u4e3a\u5355\u4e2a\u7d27\u51d1\u5411\u91cf\uff0c\u5728\u6781\u4f4e\u6bd4\u7279\u6761\u4ef6\u4e0b\u4f9d\u65e7\u4fdd\u6301\u4f18\u79c0\u7684\u89c6\u89c9\u91cd\u5efa\u6548\u679c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"744\" height=\"344\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-6.png\" alt=\"diagram\" class=\"wp-image-1184783\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-6.png 744w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-6-300x139.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-6-240x111.png 240w\" sizes=\"auto, (max-width: 744px) 100vw, 744px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1a\u57fa\u4e8e\u6269\u6563\u751f\u6210\u6a21\u578b\u7684\u5355\u5411\u91cf\u89c6\u89c9\u538b\u7f29\uff08\u5de6\uff09\u548c\u751f\u6210\uff08\u53f3\uff09\u6846\u67b6\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u6bd4\u4f20\u7edf\u89c6\u9891\u7f16\u7801\u3001\u751f\u6210\u5f0f\u538b\u7f29\u57fa\u7ebf\u7684\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8fd9\u5957\u9690\u5f0f\u8868\u5f81\u65b9\u6848\u91cd\u5efa\u753b\u8d28\u7684\u4f18\u52bf\u660e\u663e\u3002\u5f97\u76ca\u4e8e\u51fd\u6570\u5316\u7684\u8868\u5f81\u8bbe\u8ba1\uff0c\u63a8\u7406\u9636\u6bb5\u53ef\u7075\u6d3b\u62d3\u5c55\u4f18\u5316\u7b56\u7565\uff0c\u901a\u8fc7\u63a8\u7406\u65f6\u591a\u8def\u5f84\u91c7\u6837\u7ec6\u5316\u8fd8\u539f\u7ec6\u8282\uff0c\u6301\u7eed\u63d0\u5347\u753b\u9762\u89c2\u611f\u3002\u53e6\u5916\uff0c\u8fd9\u79cd\u9690\u5f0f\u8868\u5f81\u53ef\u4ee5\u4f5c\u4e3a\u89c6\u89c9\u751f\u6210\u6a21\u578b\u7684\u8bb0\u5fc6\uff0c\u4f7f\u5f97\u751f\u6210\u76f8\u4f3c\u5185\u5bb9\u65f6\uff0c\u53ef\u4ee5\u5145\u5206\u4fdd\u7559\u8bb0\u5fc6\u4e2d\u7684\u4fe1\u606f\uff0c\u4f8b\u5982\u52a8\u4f5c\u3001\u4eba\u7269\u3001\u80cc\u666f\u7b49\u3002\u8be5\u601d\u8def\u6253\u901a\u4e86\u89c6\u89c9\u538b\u7f29\u4e0e\u751f\u6210\u6a21\u578b\u7684\u8fb9\u754c\uff0c\u5f62\u6210\u4e86\u7edf\u4e00\u7684\u6280\u672f\u6846\u67b6\uff0c\u4e3a\u9ad8\u6548\u89c6\u89c9\u5b58\u50a8\u3001\u53ef\u63a7\u89c6\u89c9\u8bb0\u5fc6\u63d0\u4f9b\u4e86\u5168\u65b0\u7684\u5b9e\u73b0\u601d\u8def\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"4-\u57fa\u4e8e\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u7684\u591a\u6a21\u6001\u9690\u5f0f\u8bed\u8a00\u5efa\u6a21\" class=\"wp-block-heading\">4.\u57fa\u4e8e\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u7684\u591a\u6a21\u6001\u9690\u5f0f\u8bed\u8a00\u5efa\u6a21<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"321\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-7.png\" alt=\"text\" class=\"wp-image-1184784\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-7.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-7-300x111.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-7-768x285.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-7-240x89.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2412.08635\">https:\/\/arxiv.org\/abs\/2412.08635<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u591a\u6a21\u6001\u751f\u6210\u957f\u671f\u9762\u4e34\u5f02\u6784\u6570\u636e\u96be\u4ee5\u7edf\u4e00\u5efa\u6a21\u7684\u95ee\u9898\u3002\u6587\u672c\u548c\u4ee3\u7801\u901a\u5e38\u8868\u793a\u4e3a\u79bb\u6563\u8bcd\u5143\uff0c\u800c\u56fe\u50cf\u3001\u97f3\u9891\u3001\u89c6\u9891\u53ca\u673a\u5668\u4eba\u52a8\u4f5c\u7b49\u6a21\u6001\u5219\u5177\u6709\u8fde\u7eed\u6027\uff0c\u4e8c\u8005\u96be\u4ee5\u5728\u540c\u4e00\u6846\u67b6\u4e0b\u8fdb\u884c\u7edf\u4e00\u5904\u7406\u3002\u73b0\u6709\u65b9\u6848\u5927\u591a\u662f\u62fc\u63a5\u72ec\u7acb\u6a21\u5757\u800c\u642d\u5efa\u7684\u6d41\u6c34\u7ebf\u6846\u67b6\uff0c\u4e0d\u4ec5\u96be\u4ee5\u5b8c\u6210\u7aef\u5230\u7aef\u7684\u8054\u5408\u4f18\u5316\uff0c\u8fd8\u53ef\u80fd\u5728\u6a21\u5757\u7684\u4ea4\u4e92\u8fc7\u7a0b\u4e2d\u4ea7\u751f\u4fe1\u606f\u635f\u8017\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u6f5c\u8bed\u8a00\u5efa\u6a21LatentLM\u3002\u8be5\u6846\u67b6\u91c7\u7528\u56e0\u679cTransformer\u4f5c\u4e3a\u5171\u4eab\u4e3b\u5e72\uff0c\u901a\u8fc7\u53d8\u5206\u81ea\u7f16\u7801\u5668\uff08VAE\uff09\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u8fde\u7eed\u6570\u636e\u538b\u7f29\u4e3a\u6f5c\u5411\u91cf\uff0c\u5e76\u5f15\u5165\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u673a\u5236\u3002\u5bf9\u4e8e\u8fde\u7eed\u6a21\u6001\uff0c\u6a21\u578b\u4ee5Transformer\u7684\u9690\u72b6\u6001\u4e3a\u6761\u4ef6\uff0c\u901a\u8fc7\u8f7b\u91cf\u7ea7\u6269\u6563\u5934\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u9010\u4e2a\u751f\u6210\u6f5c\u5411\u91cf\uff1b\u5bf9\u4e8e\u6587\u672c\u7b49\u79bb\u6563\u6a21\u6001\uff0c\u5219\u6cbf\u7528\u6807\u51c6\u7684\u4e0b\u4e00\u8bcd\u5143\u9884\u6d4b\u5206\u652f\u3002\u6b64\u5916\uff0c\u8be5\u6846\u67b6\u914d\u5957\u8bbe\u8ba1\u4e86\u03c3-VAE\uff0c\u4ee5\u7f13\u89e3\u6f5c\u7a7a\u95f4\u4e2d\u7684\u65b9\u5dee\u574d\u584c\u95ee\u9898\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"360\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-8.png\" alt=\"diagram\" class=\"wp-image-1184785\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-8.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-8-300x125.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-8-768x320.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-8-240x100.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1a\u6f5c\u8bed\u8a00\u6a21\u578bLatentLM\u4ee5\u56e0\u679cTransformer\u4e3a\u5171\u4eab\u4e3b\u5e72\uff0c\u7edf\u4e00\u5904\u7406\u56fe\u50cf\u3001\u97f3\u9891\u548c\u89c6\u9891\u7b49\u8fde\u7eed\u6570\u636e\uff0c\u4ee5\u53ca\u6587\u672c\u548c\u4ee3\u7801\u7b49\u79bb\u6563\u6570\u636e\u3002\u8be5\u65b9\u6cd5\u5f15\u5165\u4e0b\u4e00\u8bcd\u6269\u6563\u673a\u5236\uff0c\u901a\u8fc7\u81ea\u56de\u5f52\u65b9\u5f0f\u9010\u4e2a\u751f\u6210\u6f5c\u5728\u5411\u91cf\u3002\u8be5\u65b9\u6cd5\u63d0\u4f9b\u4e86\u4e00\u4e2a\u901a\u7528\u63a5\u53e3\uff0c\u5b9e\u73b0\u4e86\u591a\u6a21\u6001\u751f\u6210\u4e0e\u7406\u89e3\u7684\u7edf\u4e00\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u8868\u660e\uff0c\u5728ImageNet\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u4e2d\uff0c4.79\u4ebf\u53c2\u6570\u7248\u672c\u7684LatentLM\u53d6\u5f97\u4e862.24\u7684FID\u6307\u6807\uff0c\u4ee5\u66f4\u5c11\u7684\u53c2\u6570\u8d85\u8d8a\u4e866.75\u4ebf\u53c2\u6570\u7684DiT-XL\/2\uff08FID 2.27\uff09\u3002\u968f\u7740\u6a21\u578b\u89c4\u6a21\u6269\u5927\uff0cLatentLM\u8868\u73b0\u51fa\u826f\u597d\u7684\u53ef\u6269\u5c55\u6027\uff0c\u63a8\u7406\u541e\u5410\u91cf\u6700\u9ad8\u53ef\u63d0\u53472.84\u500d\u3002\u5c06\u8be5\u6846\u67b6\u878d\u5165\u591a\u6a21\u6001\u5927\u6a21\u578b\u540e\uff0c\u5176\u5728\u8bed\u8a00\u5efa\u6a21\u56f0\u60d1\u5ea6\u3001\u6587\u751f\u56feFID\u548c\u89c6\u89c9\u95ee\u7b54\u7b49\u6307\u6807\u4e0a\u5747\u4f18\u4e8eTransfusion\u53ca\u5411\u91cf\u91cf\u5316\u57fa\u7ebf\uff0c\u5e76\u4e14\u968f\u7740\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u6269\u5927\uff0c\u6027\u80fd\u4f18\u52bf\u8fdb\u4e00\u6b65\u589e\u5f3a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8bed\u97f3\u5408\u6210\u4efb\u52a1\u4e2d\uff0c\u03c3-VAE\u53ef\u5b9e\u73b01600\u500d\u8d85\u9ad8\u538b\u7f29\u6bd4\uff0c\u89e3\u7801\u6b65\u6570\u4ec5\u4e3aVALL-E 2\u7684\u5341\u5206\u4e4b\u4e00\uff0c\u5e76\u5728\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u548c\u97f3\u9891\u9c81\u68d2\u6027\u65b9\u9762\u53d6\u5f97\u66f4\u4f18\u8868\u73b0\u3002\u8be5\u6846\u67b6\u8fd8\u53ef\u590d\u7528\u5927\u6a21\u578b\u73b0\u6709\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u4f53\u7cfb\uff0c\u4e3a\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u63d0\u4f9b\u9ad8\u6548\u4e14\u6613\u4e8e\u6269\u5c55\u7684\u6280\u672f\u8def\u5f84\u3002<\/p>\n\n\n\n<h2 id=\"5-vitl-\u878d\u5408\u4ea4\u9519\u5f0f\u63a8\u7406\u7684\u957f\u89c6\u9891\u95ee\u7b54\" class=\"wp-block-heading\">5.ViTL\uff1a\u878d\u5408\u4ea4\u9519\u5f0f\u63a8\u7406\u7684\u957f\u89c6\u9891\u95ee\u7b54<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"276\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-9.png\" alt=\"title,image\" class=\"wp-image-1184786\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-9.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-9-300x96.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-9-768x245.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-9-240x77.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2510.04022\">https:\/\/arxiv.org\/abs\/2510.04022<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u957f\u89c6\u9891\u95ee\u7b54\u957f\u671f\u53d7\u56f0\u4e8e\u201c\u56fa\u5b9a\u7b97\u529b\u9884\u7b97\u201d\u4e0e\u201c\u7b54\u6848\u5f80\u5f80\u85cf\u5728\u77ed\u6682\u7247\u6bb5\u91cc\u201d\u7684\u77db\u76fe\u4e2d\u3002\u7531\u4e8e\u6a21\u578b\u7684token\u3001\u5e27\u6570\u5b58\u5728\u56fa\u5b9a\u4e0a\u9650\uff0c\u73b0\u6709\u6a21\u578b\u591a\u91c7\u7528\u5747\u5300\u6216\u542f\u53d1\u5f0f\u91c7\u6837\uff0c\u5c06\u5927\u90e8\u5206\u5bb9\u91cf\u6d88\u8017\u5728\u65e0\u5173\u7684\u80cc\u666f\u753b\u9762\u4e0a\uff0c\u8fd9\u6070\u6070\u6f0f\u6389\u4e86\u627f\u8f7d\u7b54\u6848\u7684\u5173\u952e\u7247\u6bb5\u3002\u73b0\u6709\u7684\u516c\u5f00\u8bad\u7ec3\u96c6\u5f88\u5c11\u5c06\u95ee\u9898\u4e0e\u771f\u5b9e\u65f6\u95f4\u533a\u95f4\u7ed1\u5b9a\uff0c\u6a21\u578b\u53ea\u80fd\u76f4\u63a5\u8f93\u51fa\u7b54\u6848\uff0c\u65e0\u6cd5\u5b9a\u4f4d\u5bf9\u5e94\u7684\u63a8\u7406\u4f9d\u636e\uff0c\u8bc4\u6d4b\u53ef\u8ffd\u6eaf\u6027\u8f83\u5dee\u3002\u800c\u65f6\u5e8f\u5b9a\u4f4d\u4e0e\u95ee\u7b54\u6a21\u5757\u5206\u5f00\u4f18\u5316\uff0c\u5373\u4fbf\u7247\u6bb5\u5b9a\u4f4d\u7cbe\u5ea6\u63d0\u5347\uff0c\u6548\u679c\u589e\u76ca\u4e5f\u5f88\u96be\u8f6c\u5316\u4e3a\u7b54\u9898\u51c6\u786e\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u8fd9\u4e00\u74f6\u9888\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86ViTL\uff08Video-in-the-Loop\uff09\u4e24\u9636\u6bb5\u6846\u67b6\u3002\u7b2c\u4e00\u9636\u6bb5\uff0c\u6a21\u578b\u5148\u7528\u4f4e\u5e27\u7387\u5feb\u901f\u6d4f\u89c8\u5b8c\u6574\u89c6\u9891\uff0c\u5b9a\u4f4d\u4e00\u4e2a\u6216\u591a\u4e2a\u4e0e\u95ee\u9898\u76f8\u5173\u7684\u65f6\u95f4\u6bb5\uff1b\u7b2c\u4e8c\u9636\u6bb5\u518d\u653e\u5927\u5bf9\u5e94\u7247\u6bb5\uff0c\u4ee5\u66f4\u9ad8\u6709\u6548\u5e27\u7387\u7f16\u7801\u753b\u9762\u5e76\u5b8c\u6210\u4f5c\u7b54\u3002\u4e24\u9636\u6bb5\u5171\u4eab\u56fa\u5b9a\u7684token\u603b\u91cf\uff0c\u5c06\u89c6\u89c9\u7b97\u529b\u91cd\u65b0\u5206\u914d\u7ed9\u5173\u952e\u8bc1\u636e\u7247\u6bb5\uff0c\u800c\u975e\u5e73\u5747\u5206\u914d\u81f3\u5168\u89c6\u9891\u753b\u9762\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u91c7\u7528\u7684\u662f\u4ea4\u9519\u5f0f\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\uff0c\u8026\u5408\u65f6\u5e8fIoU\u5b9a\u4f4d\u6307\u6807\u4e0e\u7b54\u9898\u5956\u52b1\uff0c\u8ba9\u5956\u52b1\u4fe1\u53f7\u53ef\u4ee5\u53cd\u5411\u4f20\u5bfc\u81f3\u5b9a\u4f4d\u5206\u652f\uff0c\u5f15\u5bfc\u6a21\u578b\u4f18\u5148\u5b66\u4e60\u5229\u4e8e\u6b63\u786e\u4f5c\u7b54\u7684\u89c6\u9891\u7247\u6bb5\u3002\u4e3a\u8865\u9f50\u65f6\u5e8f\u5b9a\u4f4d\u6240\u9700\u7684\u76d1\u7763\u4fe1\u53f7\uff0c\u7814\u7a76\u5458\u4eec\u642d\u5efa\u4e86VGrounding-QA\u6570\u636e\u96c6\uff0c\u4f9d\u6258GPT-4o\u5bf9\u89c6\u9891\u505a\u7ec6\u7c92\u5ea6\u7ed3\u6784\u5316\u89e3\u6790\uff0c\u63d0\u53d6\u5b9e\u4f53\u5e76\u6784\u5efa\u4e8b\u4ef6\u77e5\u8bc6\u56fe\uff0c\u518d\u57fa\u4e8e\u56fe\u8c31\u751f\u6210\u7ed1\u5b9a\u771f\u5b9e\u65f6\u95f4\u533a\u95f4\u7684\u591a\u65f6\u6bb5\u3001\u591a\u9009\u95ee\u7b54\u6837\u672c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"461\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-10.png\" alt=\"diagram\" class=\"wp-image-1184787\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-10.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-10-300x160.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-10-768x409.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/8-10-240x128.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1aViTL\u4e0eVGrounding-QA\u6846\u67b6\u603b\u89c8\u3002\u7ed9\u5b9a\u957f\u89c6\u9891\u4e0e\u95ee\u9898\uff0c\u9636\u6bb51\u4ece\u95ee\u53e5\u4e2d\u63d0\u53d6\u5b9a\u4f4d\u67e5\u8be2\uff0c\u9884\u6d4b\u76f8\u5173\u65f6\u95f4\u533a\u95f4\uff1b\u9636\u6bb52\u4ec5\u5bf9\u7b5b\u9009\u7247\u6bb5\u505a\u9ad8\u4fdd\u771f\u91cd\u7f16\u7801\uff0c\u5b8c\u6210\u591a\u9009\u95ee\u7b54\u3002\u6574\u5957\u6d41\u7a0b\u91c7\u7528R1\u7c7b\u5f3a\u5316\u5b66\u4e60\u5faa\u73af\u8bad\u7ec3\u65b9\u6848\uff0c\u8054\u5408\u4f18\u5316\u65f6\u5e8fIoU\u4e0e\u95ee\u7b54\u4e24\u5927\u76ee\u6807\uff0c\u4f18\u5148\u4f18\u5316\u6709\u6548\u8bc1\u636e\u7247\u6bb5\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u4e0e\u6d88\u878d\u7ed3\u679c\u663e\u793a\uff0cViTL\u4ec5\u4ee5128\u5e27\u8f93\u5165\u5373\u5728LVBench\u4e0a\u8fbe\u523047.4%\uff0c\u8d85\u8fc7256\u5e27\u5747\u5300\u91c7\u6837\u7684Qwen2.5-VL 7B\u57fa\u7ebf\uff0843.7%\uff09\uff0c\u5e76\u4ee550%\u7684\u5e27\u8f93\u5165\u91cf\u53d6\u5f97\u6700\u9ad88.6%\u7684\u6027\u80fd\u63d0\u5347\u3002\u6d88\u878d\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u8868\u660e\uff0c\u57fa\u4e8e\u7247\u6bb5\u611f\u77e5\u7684token\u5206\u914d\u59cb\u7ec8\u4f18\u4e8e\u5168\u5c40\u5747\u5300\u91c7\u6837\u3002ViTL\u9a8c\u8bc1\u4e86\u56fa\u5b9a\u7b97\u529b\u4e0b\u805a\u7126\u8bc1\u636e\u7247\u6bb5\u7684\u8bbe\u8ba1\u601d\u8def\uff0c\u540c\u65f6\u517c\u987e\u7b54\u9898\u7cbe\u5ea6\u4e0e\u65f6\u5e8f\u53ef\u8ffd\u6eaf\u6027\uff0c\u4e3a\u9ad8\u6548\u3001\u53ef\u89e3\u91ca\u957f\u89c6\u9891\u5927\u6a21\u578b\u63a8\u7406\u63d0\u4f9b\u4e86\u5168\u65b0\u53ef\u7684\u884c\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u671f\u5185\u5bb9\u901f\u89c8\uff1a 1.\u8ba9\u865a\u62df\u4e16\u754c\u544a\u522b\u9057\u5fd8\uff1a\u6784\u5efa\u5177\u6709\u6301\u4e45\u4e09\u7ef4\u8bb0\u5fc6\u7684\u4e16\u754c\u6a21\u578b 2.CoD-Lite\uff1a\u57fa\u4e8e\u6269\u6563\u6a21\u578b\u7684\u5b9e\u65f6\u751f\u6210\u5f0f\u56fe\u50cf\u538b\u7f29 3.\u4e00\u4e2a\u5411\u91cf\uff0c\u88c5\u4e0b\u6574\u6bb5\u89c6\u9891\uff1a\u89c6\u89c9\u751f\u6210\u6a21\u578b\u91cc\u7684\u9690\u5f0f\u8bb0\u5fc6\u8868\u5f81 4.\u57fa\u4e8e\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u7684\u591a\u6a21\u6001\u9690\u5f0f\u8bed\u8a00\u5efa\u6a21 5.ViTL\uff1a\u878d\u5408\u4ea4\u9519\u5f0f\u63a8\u7406\u7684\u957f\u89c6\u9891\u95ee\u7b54 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2603.03482 (opens in new tab) \u73b0\u6709\u7684\u4ea4\u4e92\u5f0f\u4e16\u754c\u6a21\u578b\u5e38\u7f3a\u4e4f\u5bf9\u73af\u5883\u7684\u4e09\u7ef4\u8868\u793a\uff0c\u53ea\u80fd\u4ece\u6570\u636e\u4e2d\u9690\u5f0f\u5730\u5b66\u4e60\u4e09\u7ef4\u4e00\u81f4\u6027\u3002\u800c\u5b83\u4eec\u7684\u7a7a\u95f4\u8bb0\u5fc6\u53d7\u9650\u4e8e\u6709\u9650\u7684\u65f6\u95f4\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u6240\u4ee5\u5728\u8fdb\u884c\u957f\u7a0b\u63a8\u6f14\u65f6\u5bb9\u6613\u4e22\u5931\u7a7a\u95f4\u4fe1\u606f\uff0c\u5bfc\u81f4\u4e00\u8fb9\u751f\u6210\u4e00\u8fb9\u9057\u5fd8\uff0c\u96be\u4ee5\u9002\u914d\u6c89\u6d78\u5f0f\u4ea4\u4e92\u548c\u667a\u80fd\u4f53\u8bad\u7ec3\u7b49\u4efb\u52a1\u3002&nbsp;&nbsp; \u9488\u5bf9\u8fd9\u4e00\u74f6\u9888\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86PERSIST\u6846\u67b6\uff0c\u901a\u8fc7\u9690\u7a7a\u95f4\u4e2d\u7684\u4e09\u7ef4\u73af\u5883\u5e27\uff08world-frame\uff09\u663e\u5f0f\u5efa\u6a21\u73af\u5883\u968f\u65f6\u95f4\u6f14\u5316\u7684\u72b6\u6001\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6a21\u578b\u7ef4\u62a4\u5e76\u9884\u6d4b\u4e00\u4e2a\u4e09\u7ef4\u73af\u5883\u5e27\uff0c\u7528\u4e8e\u4fdd\u5b58\u6301\u4e45\u7684\u7a7a\u95f4\u8bb0\u5fc6\uff1b\u540c\u65f6\u9884\u6d4b\u76f8\u673a\u72b6\u6001\uff0c\u5e76\u6839\u636e\u5f53\u524d\u89c6\u89d2\u5c06\u4e09\u7ef4\u73af\u5883\u5e27\u4e2d\u7684\u4fe1\u606f\u6295\u5f71\u5230\u53ef\u89c2\u6d4b\u7684\u4e8c\u7ef4\u5c4f\u5e55\u7a7a\u95f4\uff1b\u6700\u540e\uff0c\u5229\u7528\u4e00\u4e2aworld-to-pixel\uff08\u4e16\u754c\u72b6\u6001\u5230\u50cf\u7d20\u753b\u9762\uff09\u751f\u6210\u6a21\u5757\uff0c\u6839\u636e\u6295\u5f71\u540e\u7684\u4e09\u7ef4\u73af\u5883\u7279\u5f81\u3001\u52a8\u4f5c\u4ee5\u53ca\u4e8c\u7ef4\u56fe\u50cf\u7279\u5f81\u5408\u6210\u5f53\u524d\u89c6\u9891\u5e27\u3002 \u5728\u8bad\u7ec3\u9636\u6bb5\uff0cPERSIST&nbsp;\u4f7f\u75282D-VAE&nbsp;\u548c&nbsp;3D-VAE&nbsp;\u5206\u522b\u7f16\u7801\u89c6\u9891\u5e27\u548c\u4e09\u7ef4\u73af\u5883\u5e27\uff0c\u5176\u4e2d\u73af\u5883\u5e27\u9884\u6d4b\u5668\u548cworld-to-pixel&nbsp;\u751f\u6210\u5668\u5747\u91c7\u7528&nbsp;Rectified Flow&nbsp;\u6846\u67b6\uff0c\u76f8\u673a\u6a21\u578b\u5219\u901a\u8fc7\u56e0\u679cTransformer&nbsp;\u9884\u6d4b\u4f4d\u7f6e\u3001\u671d\u5411\u548c\u89c6\u573a\u53d8\u5316\u3002\u4e3a\u7f13\u89e3\u81ea\u56de\u5f52\u63a8\u7406\u4e2d\u7684\u8bef\u5dee\u7d2f\u79ef\uff0c\u6a21\u578b\u8fdb\u4e00\u6b65\u5f15\u5165&nbsp;Diffusion Forcing&nbsp;\u548c\u968f\u673a\u566a\u58f0\u589e\u5f3a\uff0c\u4f7f\u5404\u4e2a\u6a21\u5757\u5728\u8bad\u7ec3\u65f6\u80fd\u591f\u9002\u5e94\u63a8\u7406\u9636\u6bb5\u7531\u5176\u4ed6\u6a21\u5757\u751f\u6210\u7684\u9690\u7a7a\u95f4\u7279\u5f81\u3002 \u5728\u00a0Luanti\u00a0\u73af\u5883\u4e2d\u7684\u5b9e\u9a8c\u8868\u660e\uff0cPERSIST\u00a0\u5728\u957f\u7a0b\u4ea4\u4e92\u5f0f\u751f\u6210\u4e2d\u660e\u663e\u4f18\u4e8e\u00a0Oasis\u3001WorldMem\u00a0\u7b49\u57fa\u7ebf\u65b9\u6cd5\uff1a600\u00a0\u5e27\u751f\u6210\u4efb\u52a1\u4e2d\uff0cFVD\u00a0\u4ece\u00a0Oasis\u00a0\u7684\u00a0875\u00a0\u964d\u81f3148\uff1b\u7528\u6237\u8bc4\u4f30\u4e2d\uff0c\u4e09\u7ef4\u7a7a\u95f4\u4e00\u81f4\u6027\u548c\u65f6\u95f4\u7a33\u5b9a\u6027\u4e5f\u663e\u8457\u63d0\u5347\u3002\u5b9a\u6027\u7ed3\u679c\u663e\u793a\uff0cPERSIST\u00a0\u80fd\u5728\u6570\u5343\u5e27\u4ea4\u4e92\u63a8\u6f14\u4e2d\u4fdd\u6301\u573a\u666f\u7ed3\u6784\u8fde\u8d2f\uff0c\u5e76\u80fd\u5efa\u6a21\u89c6\u91ce\u5916\u73af\u5883\u7684\u6301\u7eed\u53d8\u5316\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2604.12525 (opens in new tab) \u5728\u9ad8\u6027\u80fd\u751f\u6210\u5f0f\u538b\u7f29\u9886\u57df\uff0c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u901a\u5e38\u4f9d\u9760\u5927\u5e45\u6269\u5145\u53c2\u6570\u91cf\u6765\u63d0\u5347\u89c6\u89c9\u91cd\u5efa\u6548\u679c\u3002\u8fd9\u4f1a\u5e26\u6765\u6781\u9ad8\u7684\u63a8\u7406\u5ef6\u8fdf\uff0c\u5f88\u96be\u6ee1\u8db3\u5404\u7c7b\u5b9e\u65f6\u843d\u5730\u573a\u666f\u3002\u5982\u4f55\u5e73\u8861\u6a21\u578b\u89c4\u6a21\u4e0e\u8fd0\u884c\u6548\u7387\uff0c\u8bbe\u8ba1\u8f7b\u91cf\u5316\u6269\u6563\u67b6\u6784\u662f\u5f53\u4e0b\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002 \u4e3a\u7a81\u7834\u8fd9\u4e00\u5c40\u9650\uff0c\u7814\u7a76\u5458\u4eec\u5f00\u5c55\u4e86\u591a\u7ec4\u5b9e\u9a8c\u5206\u6790\uff0c\u901a\u8fc7\u5bf9\u6bd4\u4e0d\u540c\u53c2\u6570\u91cf\u7684\u6269\u6563\u6a21\u578b\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u7684\u4f18\u52bf\u3002\u5728\u8f7b\u91cf\u5316\u6a21\u578b\u4e0a\uff0c\u76f8\u6bd4\u901a\u7528\u751f\u6210\u6269\u6563\u9884\u8bad\u7ec3\u65b9\u6848\uff0c\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u80fd\u751f\u6210\u4fe1\u606f\u5bc6\u5ea6\u66f4\u9ad8\u7684\u56fe\u50cf\u8868\u5f81\uff0c\u53ef\u5f25\u8865\u5c0f\u6a21\u578b\u8868\u5f81\u80fd\u529b\u4e0d\u8db3\u7684\u7f3a\u9677\u3002\u7814\u7a76\u5458\u4eec\u540c\u65f6\u53d1\u73b0\uff0c\u56fe\u50cf\u538b\u7f29\u4efb\u52a1\u5e76\u4e0d\u9700\u8981\u5168\u5c40\u6ce8\u610f\u529b\uff0c\u8f7b\u91cf\u5316\u5377\u79ef\u7f51\u7edc\u642d\u914d\u84b8\u998f\u5b66\u4e60\uff0c\u5373\u53ef\u8fd8\u539f\u5b8c\u6574\u7684\u7cbe\u7ec6\u7eb9\u7406\u3002 \u57fa\u4e8e\u8fd9\u4e9b\u53d1\u73b0\uff0c\u7814\u7a76\u5458\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u4e00\u6b65\u5f0f\u5377\u79ef\u6269\u6563\u56fe\u50cf\u7f16\u89e3\u7801\u5668\u3002\u8be5\u65b9\u6cd5\u91c7\u7528\u7d27\u51d1\u7684\u5377\u79ef\u4e3b\u5e72\u7f51\u7edc\uff0c\u901a\u8fc7\u538b\u7f29\u5bfc\u5411\u6269\u6563\u9884\u8bad\u7ec3\u548c\u7edf\u4e00\u7684\u84b8\u998f\u4e0e\u5bf9\u6297\u8bad\u7ec3\u6846\u67b6\uff0c\u6210\u529f\u5b9e\u73b0\u4e86\u5728A100 GPU\u4e0a1080p\u5206\u8fa8\u7387\u4e0b60 FPS\u7684\u7f16\u7801\u901f\u5ea6\u4e0e42 FPS\u7684\u89e3\u7801\u901f\u5ea6\u3002\u540c\u65f6\uff0c\u8be5\u65b9\u6cd5\u5728\u4fdd\u6301\u4e0eMS-ILLM\u7b49GAN\u7c7b\u7f16\u89e3\u7801\u5668\u76f8\u5f53\u7684FID\u6307\u6807\u7684\u540c\u65f6\uff0c\u53ef\u8282\u770185%\u6bd4\u7279\u7387\uff0c\u5728\u63d0\u4f9b\u9ad8\u8d28\u91cf\u89c6\u89c9\u91cd\u5efa\u7684\u540c\u65f6\u6ee1\u8db3\u4e86\u4f4e\u5ef6\u8fdf\u7684\u5b9e\u65f6\u9700\u6c42\uff0c\u6709\u6548\u642d\u5efa\u4e86\u751f\u6210\u5f0f\u538b\u7f29\u4e0e\u5b9e\u9645\u90e8\u7f72\u573a\u666f\u4e4b\u95f4\u7684\u6865\u6881\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2603.07615 (opens in new tab) \u73b0\u4ee3\u89c6\u89c9\u751f\u6210\u5927\u6a21\u578b\u7ecf\u8fc7\u6d77\u91cf\u6570\u636e\u8bad\u7ec3\uff0c\u6c89\u6dc0\u4e86\u4e30\u5bcc\u7684\u56fe\u50cf\u3001\u89c6\u9891\u5148\u9a8c\u77e5\u8bc6\u3002\u4f46\u50cf\u7d20\u3001\u9690\u53d8\u91cf\u3001\u56fe\u50cf\u6807\u8bb0\u8fd9\u7c7b\u4f20\u7edf\u89c6\u89c9\u8868\u5f81\u90fd\u72ec\u7acb\u4e8e\u751f\u6210\u6a21\u578b\uff0c\u65e0\u6cd5\u76f4\u63a5\u590d\u7528\u6a21\u578b\u5185\u7f6e\u77e5\u8bc6\uff0c\u5f88\u96be\u5b9e\u73b0\u8d85\u9ad8\u538b\u7f29\u500d\u7387\u4e0b\u7684\u9ad8\u6548\u5b58\u50a8\u4e0e\u5185\u5bb9\u590d\u7528\uff0c\u4f20\u7edf\u7f16\u7801\u65b9\u6848\u5728\u6781\u4f4e\u7801\u7387\u573a\u666f\u753b\u8d28\u8870\u51cf\u4e25\u91cd\u3002\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u9690\u5f0f\u89c6\u89c9\u8868\u5f81\u6846\u67b6\u3002\u8be5\u65b9\u6848\u4e0d\u518d\u5355\u72ec\u7f16\u7801\u753b\u9762\u50cf\u7d20\uff0c\u800c\u662f\u628a\u89c6\u89c9\u4fe1\u53f7\u5efa\u6a21\u4e3a\u751f\u6210\u51fd\u6570\uff0c\u4f9d\u9760\u51bb\u7ed3\u6269\u6563\u6a21\u578b\u914d\u5957LoRA\u53c2\u6570\u5b8c\u6210\u8868\u5f81\u5efa\u6a21\u3002\u50cf81\u5e27\u957f\u89c6\u9891\u8fd9\u7c7b\u590d\u6742\u89c6\u89c9\u5185\u5bb9\uff0c\u90fd\u80fd\u8fdb\u4e00\u6b65\u54c8\u5e0c\u538b\u7f29\u4e3a\u5355\u4e2a\u7d27\u51d1\u5411\u91cf\uff0c\u5728\u6781\u4f4e\u6bd4\u7279\u6761\u4ef6\u4e0b\u4f9d\u65e7\u4fdd\u6301\u4f18\u79c0\u7684\u89c6\u89c9\u91cd\u5efa\u6548\u679c\u3002 \u5bf9\u6bd4\u4f20\u7edf\u89c6\u9891\u7f16\u7801\u3001\u751f\u6210\u5f0f\u538b\u7f29\u57fa\u7ebf\u7684\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8fd9\u5957\u9690\u5f0f\u8868\u5f81\u65b9\u6848\u91cd\u5efa\u753b\u8d28\u7684\u4f18\u52bf\u660e\u663e\u3002\u5f97\u76ca\u4e8e\u51fd\u6570\u5316\u7684\u8868\u5f81\u8bbe\u8ba1\uff0c\u63a8\u7406\u9636\u6bb5\u53ef\u7075\u6d3b\u62d3\u5c55\u4f18\u5316\u7b56\u7565\uff0c\u901a\u8fc7\u63a8\u7406\u65f6\u591a\u8def\u5f84\u91c7\u6837\u7ec6\u5316\u8fd8\u539f\u7ec6\u8282\uff0c\u6301\u7eed\u63d0\u5347\u753b\u9762\u89c2\u611f\u3002\u53e6\u5916\uff0c\u8fd9\u79cd\u9690\u5f0f\u8868\u5f81\u53ef\u4ee5\u4f5c\u4e3a\u89c6\u89c9\u751f\u6210\u6a21\u578b\u7684\u8bb0\u5fc6\uff0c\u4f7f\u5f97\u751f\u6210\u76f8\u4f3c\u5185\u5bb9\u65f6\uff0c\u53ef\u4ee5\u5145\u5206\u4fdd\u7559\u8bb0\u5fc6\u4e2d\u7684\u4fe1\u606f\uff0c\u4f8b\u5982\u52a8\u4f5c\u3001\u4eba\u7269\u3001\u80cc\u666f\u7b49\u3002\u8be5\u601d\u8def\u6253\u901a\u4e86\u89c6\u89c9\u538b\u7f29\u4e0e\u751f\u6210\u6a21\u578b\u7684\u8fb9\u754c\uff0c\u5f62\u6210\u4e86\u7edf\u4e00\u7684\u6280\u672f\u6846\u67b6\uff0c\u4e3a\u9ad8\u6548\u89c6\u89c9\u5b58\u50a8\u3001\u53ef\u63a7\u89c6\u89c9\u8bb0\u5fc6\u63d0\u4f9b\u4e86\u5168\u65b0\u7684\u5b9e\u73b0\u601d\u8def\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2412.08635 (opens in new tab) \u591a\u6a21\u6001\u751f\u6210\u957f\u671f\u9762\u4e34\u5f02\u6784\u6570\u636e\u96be\u4ee5\u7edf\u4e00\u5efa\u6a21\u7684\u95ee\u9898\u3002\u6587\u672c\u548c\u4ee3\u7801\u901a\u5e38\u8868\u793a\u4e3a\u79bb\u6563\u8bcd\u5143\uff0c\u800c\u56fe\u50cf\u3001\u97f3\u9891\u3001\u89c6\u9891\u53ca\u673a\u5668\u4eba\u52a8\u4f5c\u7b49\u6a21\u6001\u5219\u5177\u6709\u8fde\u7eed\u6027\uff0c\u4e8c\u8005\u96be\u4ee5\u5728\u540c\u4e00\u6846\u67b6\u4e0b\u8fdb\u884c\u7edf\u4e00\u5904\u7406\u3002\u73b0\u6709\u65b9\u6848\u5927\u591a\u662f\u62fc\u63a5\u72ec\u7acb\u6a21\u5757\u800c\u642d\u5efa\u7684\u6d41\u6c34\u7ebf\u6846\u67b6\uff0c\u4e0d\u4ec5\u96be\u4ee5\u5b8c\u6210\u7aef\u5230\u7aef\u7684\u8054\u5408\u4f18\u5316\uff0c\u8fd8\u53ef\u80fd\u5728\u6a21\u5757\u7684\u4ea4\u4e92\u8fc7\u7a0b\u4e2d\u4ea7\u751f\u4fe1\u606f\u635f\u8017\u3002 \u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u6f5c\u8bed\u8a00\u5efa\u6a21LatentLM\u3002\u8be5\u6846\u67b6\u91c7\u7528\u56e0\u679cTransformer\u4f5c\u4e3a\u5171\u4eab\u4e3b\u5e72\uff0c\u901a\u8fc7\u53d8\u5206\u81ea\u7f16\u7801\u5668\uff08VAE\uff09\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u8fde\u7eed\u6570\u636e\u538b\u7f29\u4e3a\u6f5c\u5411\u91cf\uff0c\u5e76\u5f15\u5165\u4e0b\u4e00\u8bcd\u5143\u6269\u6563\u673a\u5236\u3002\u5bf9\u4e8e\u8fde\u7eed\u6a21\u6001\uff0c\u6a21\u578b\u4ee5Transformer\u7684\u9690\u72b6\u6001\u4e3a\u6761\u4ef6\uff0c\u901a\u8fc7\u8f7b\u91cf\u7ea7\u6269\u6563\u5934\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u9010\u4e2a\u751f\u6210\u6f5c\u5411\u91cf\uff1b\u5bf9\u4e8e\u6587\u672c\u7b49\u79bb\u6563\u6a21\u6001\uff0c\u5219\u6cbf\u7528\u6807\u51c6\u7684\u4e0b\u4e00\u8bcd\u5143\u9884\u6d4b\u5206\u652f\u3002\u6b64\u5916\uff0c\u8be5\u6846\u67b6\u914d\u5957\u8bbe\u8ba1\u4e86\u03c3-VAE\uff0c\u4ee5\u7f13\u89e3\u6f5c\u7a7a\u95f4\u4e2d\u7684\u65b9\u5dee\u574d\u584c\u95ee\u9898\u3002 \u5b9e\u9a8c\u8868\u660e\uff0c\u5728ImageNet\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u4e2d\uff0c4.79\u4ebf\u53c2\u6570\u7248\u672c\u7684LatentLM\u53d6\u5f97\u4e862.24\u7684FID\u6307\u6807\uff0c\u4ee5\u66f4\u5c11\u7684\u53c2\u6570\u8d85\u8d8a\u4e866.75\u4ebf\u53c2\u6570\u7684DiT-XL\/2\uff08FID 2.27\uff09\u3002\u968f\u7740\u6a21\u578b\u89c4\u6a21\u6269\u5927\uff0cLatentLM\u8868\u73b0\u51fa\u826f\u597d\u7684\u53ef\u6269\u5c55\u6027\uff0c\u63a8\u7406\u541e\u5410\u91cf\u6700\u9ad8\u53ef\u63d0\u53472.84\u500d\u3002\u5c06\u8be5\u6846\u67b6\u878d\u5165\u591a\u6a21\u6001\u5927\u6a21\u578b\u540e\uff0c\u5176\u5728\u8bed\u8a00\u5efa\u6a21\u56f0\u60d1\u5ea6\u3001\u6587\u751f\u56feFID\u548c\u89c6\u89c9\u95ee\u7b54\u7b49\u6307\u6807\u4e0a\u5747\u4f18\u4e8eTransfusion\u53ca\u5411\u91cf\u91cf\u5316\u57fa\u7ebf\uff0c\u5e76\u4e14\u968f\u7740\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u6269\u5927\uff0c\u6027\u80fd\u4f18\u52bf\u8fdb\u4e00\u6b65\u589e\u5f3a\u3002 \u5728\u8bed\u97f3\u5408\u6210\u4efb\u52a1\u4e2d\uff0c\u03c3-VAE\u53ef\u5b9e\u73b01600\u500d\u8d85\u9ad8\u538b\u7f29\u6bd4\uff0c\u89e3\u7801\u6b65\u6570\u4ec5\u4e3aVALL-E 2\u7684\u5341\u5206\u4e4b\u4e00\uff0c\u5e76\u5728\u8bf4\u8bdd\u4eba\u76f8\u4f3c\u5ea6\u548c\u97f3\u9891\u9c81\u68d2\u6027\u65b9\u9762\u53d6\u5f97\u66f4\u4f18\u8868\u73b0\u3002\u8be5\u6846\u67b6\u8fd8\u53ef\u590d\u7528\u5927\u6a21\u578b\u73b0\u6709\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u4f53\u7cfb\uff0c\u4e3a\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u63d0\u4f9b\u9ad8\u6548\u4e14\u6613\u4e8e\u6269\u5c55\u7684\u6280\u672f\u8def\u5f84\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2510.04022 (opens in new tab) \u957f\u89c6\u9891\u95ee\u7b54\u957f\u671f\u53d7\u56f0\u4e8e\u201c\u56fa\u5b9a\u7b97\u529b\u9884\u7b97\u201d\u4e0e\u201c\u7b54\u6848\u5f80\u5f80\u85cf\u5728\u77ed\u6682\u7247\u6bb5\u91cc\u201d\u7684\u77db\u76fe\u4e2d\u3002\u7531\u4e8e\u6a21\u578b\u7684token\u3001\u5e27\u6570\u5b58\u5728\u56fa\u5b9a\u4e0a\u9650\uff0c\u73b0\u6709\u6a21\u578b\u591a\u91c7\u7528\u5747\u5300\u6216\u542f\u53d1\u5f0f\u91c7\u6837\uff0c\u5c06\u5927\u90e8\u5206\u5bb9\u91cf\u6d88\u8017\u5728\u65e0\u5173\u7684\u80cc\u666f\u753b\u9762\u4e0a\uff0c\u8fd9\u6070\u6070\u6f0f\u6389\u4e86\u627f\u8f7d\u7b54\u6848\u7684\u5173\u952e\u7247\u6bb5\u3002\u73b0\u6709\u7684\u516c\u5f00\u8bad\u7ec3\u96c6\u5f88\u5c11\u5c06\u95ee\u9898\u4e0e\u771f\u5b9e\u65f6\u95f4\u533a\u95f4\u7ed1\u5b9a\uff0c\u6a21\u578b\u53ea\u80fd\u76f4\u63a5\u8f93\u51fa\u7b54\u6848\uff0c\u65e0\u6cd5\u5b9a\u4f4d\u5bf9\u5e94\u7684\u63a8\u7406\u4f9d\u636e\uff0c\u8bc4\u6d4b\u53ef\u8ffd\u6eaf\u6027\u8f83\u5dee\u3002\u800c\u65f6\u5e8f\u5b9a\u4f4d\u4e0e\u95ee\u7b54\u6a21\u5757\u5206\u5f00\u4f18\u5316\uff0c\u5373\u4fbf\u7247\u6bb5\u5b9a\u4f4d\u7cbe\u5ea6\u63d0\u5347\uff0c\u6548\u679c\u589e\u76ca\u4e5f\u5f88\u96be\u8f6c\u5316\u4e3a\u7b54\u9898\u51c6\u786e\u7387\u3002 \u9488\u5bf9\u8fd9\u4e00\u74f6\u9888\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86ViTL\uff08Video-in-the-Loop\uff09\u4e24\u9636\u6bb5\u6846\u67b6\u3002\u7b2c\u4e00\u9636\u6bb5\uff0c\u6a21\u578b\u5148\u7528\u4f4e\u5e27\u7387\u5feb\u901f\u6d4f\u89c8\u5b8c\u6574\u89c6\u9891\uff0c\u5b9a\u4f4d\u4e00\u4e2a\u6216\u591a\u4e2a\u4e0e\u95ee\u9898\u76f8\u5173\u7684\u65f6\u95f4\u6bb5\uff1b\u7b2c\u4e8c\u9636\u6bb5\u518d\u653e\u5927\u5bf9\u5e94\u7247\u6bb5\uff0c\u4ee5\u66f4\u9ad8\u6709\u6548\u5e27\u7387\u7f16\u7801\u753b\u9762\u5e76\u5b8c\u6210\u4f5c\u7b54\u3002\u4e24\u9636\u6bb5\u5171\u4eab\u56fa\u5b9a\u7684token\u603b\u91cf\uff0c\u5c06\u89c6\u89c9\u7b97\u529b\u91cd\u65b0\u5206\u914d\u7ed9\u5173\u952e\u8bc1\u636e\u7247\u6bb5\uff0c\u800c\u975e\u5e73\u5747\u5206\u914d\u81f3\u5168\u89c6\u9891\u753b\u9762\u3002 \u8bad\u7ec3\u91c7\u7528\u7684\u662f\u4ea4\u9519\u5f0f\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\uff0c\u8026\u5408\u65f6\u5e8fIoU\u5b9a\u4f4d\u6307\u6807\u4e0e\u7b54\u9898\u5956\u52b1\uff0c\u8ba9\u5956\u52b1\u4fe1\u53f7\u53ef\u4ee5\u53cd\u5411\u4f20\u5bfc\u81f3\u5b9a\u4f4d\u5206\u652f\uff0c\u5f15\u5bfc\u6a21\u578b\u4f18\u5148\u5b66\u4e60\u5229\u4e8e\u6b63\u786e\u4f5c\u7b54\u7684\u89c6\u9891\u7247\u6bb5\u3002\u4e3a\u8865\u9f50\u65f6\u5e8f\u5b9a\u4f4d\u6240\u9700\u7684\u76d1\u7763\u4fe1\u53f7\uff0c\u7814\u7a76\u5458\u4eec\u642d\u5efa\u4e86VGrounding-QA\u6570\u636e\u96c6\uff0c\u4f9d\u6258GPT-4o\u5bf9\u89c6\u9891\u505a\u7ec6\u7c92\u5ea6\u7ed3\u6784\u5316\u89e3\u6790\uff0c\u63d0\u53d6\u5b9e\u4f53\u5e76\u6784\u5efa\u4e8b\u4ef6\u77e5\u8bc6\u56fe\uff0c\u518d\u57fa\u4e8e\u56fe\u8c31\u751f\u6210\u7ed1\u5b9a\u771f\u5b9e\u65f6\u95f4\u533a\u95f4\u7684\u591a\u65f6\u6bb5\u3001\u591a\u9009\u95ee\u7b54\u6837\u672c\u3002 \u5b9e\u9a8c\u4e0e\u6d88\u878d\u7ed3\u679c\u663e\u793a\uff0cViTL\u4ec5\u4ee5128\u5e27\u8f93\u5165\u5373\u5728LVBench\u4e0a\u8fbe\u523047.4%\uff0c\u8d85\u8fc7256\u5e27\u5747\u5300\u91c7\u6837\u7684Qwen2.5-VL [&hellip;]<\/p>\n","protected":false},"author":44093,"featured_media":1184788,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[13556,13562,13551],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1184777","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-research-area-computer-vision","msr-research-area-graphics-and-multimedia","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184777","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/44093"}],"version-history":[{"count":1,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184777\/revisions"}],"predecessor-version":[{"id":1184789,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184777\/revisions\/1184789"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1184788"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1184777"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1184777"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1184777"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1184777"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}