{"id":5269,"date":"2024-12-05T14:41:01","date_gmt":"2024-12-05T06:41:01","guid":{"rendered":"https:\/\/nullthought.net\/?p=5269"},"modified":"2025-02-23T11:08:36","modified_gmt":"2025-02-23T03:08:36","slug":"%e8%a7%86%e8%a7%89%e8%87%aa%e5%9b%9e%e5%bd%92%e5%bb%ba%e6%a8%a1%ef%bc%88visual-autoregressive-var%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5269","title":{"rendered":"\u89c6\u89c9\u81ea\u56de\u5f52\u5efa\u6a21\uff08Visual AutoRegressive, VAR\uff09"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2404.02905\" target=\"_blank\" rel=\"noreferrer noopener\">Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction<\/a><\/strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u5168\u65b0\u7684\u89c6\u89c9\u81ea\u56de\u5f52\u751f\u6210\u6846\u67b6\u2014\u2014<strong>\u89c6\u89c9\u81ea\u56de\u5f52\u5efa\u6a21\uff08VAR\uff09<\/strong>\uff0c\u901a\u8fc7\u591a\u5c3a\u5ea6\u7684\u751f\u6210\u7b56\u7565\uff0c\u6709\u6548\u89e3\u51b3\u4e86\u4f20\u7edf\u81ea\u56de\u5f52\u6a21\u578b\u5728\u56fe\u50cf\u751f\u6210\u4e2d\u7684\u6548\u7387\u4f4e\u4e0b\u548c\u7a7a\u95f4\u7ed3\u6784\u7834\u574f\u7b49\u95ee\u9898\u3002VAR\u6a21\u578b\u9996\u6b21\u5728\u56fe\u50cf\u751f\u6210\u7684\u8d28\u91cf\u3001\u901f\u5ea6\u548c\u6548\u7387\u4e0a\u8d85\u8d8a\u4e86\u4e3b\u6d41\u7684\u6269\u6563\u6a21\u578b\uff0c\u5c55\u793a\u4e86\u5176\u4f5c\u4e3a\u4e0b\u4e00\u4ee3\u56fe\u50cf\u751f\u6210\u6a21\u578b\u7684\u5de8\u5927\u6f5c\u529b\u3002\u6b64\u5916\uff0cVAR\u6a21\u578b\u7684\u6807\u5ea6\u89c4\u5f8b\u548c\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u5176\u6269\u5c55\u6027\u548c\u901a\u7528\u6027\uff0c\u4e3a\u89c6\u89c9\u81ea\u56de\u5f52\u5b66\u4e60\u9886\u57df\u7684\u8fdb\u4e00\u6b65\u63a2\u7d22\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u7406\u8bba\u548c\u5b9e\u8df5\u57fa\u7840\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aKeyu Tian\uff08\u7530\u67ef\u5b87\uff09, Yi Jiang, Zehuan Yuan, Bingyue Peng, Liwei Wang\uff0c\u6765\u81ea\u5317\u4eac\u5927\u5b66\uff08Peking University\uff09\u548c\u5b57\u8282\u8df3\u52a8\uff08Bytedance Inc\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"466\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-5-1024x466.png\" alt=\"\" class=\"wp-image-5270\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-5-1024x466.png 1024w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-5-300x137.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-5-768x350.png 768w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-5.png 1045w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2404.02905\" target=\"_blank\" rel=\"noreferrer noopener\">Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">1. \u81ea\u56de\u5f52\u6a21\u578b\u4e0e\u8bed\u8a00\u751f\u6210\u7684\u6210\u529f<\/h5>\n\n\n\n<p>\u81ea\u56de\u5f52\uff08Autoregressive, AR\uff09\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u751f\u6210\u4e2d\u7684\u6210\u529f\uff0c\u7279\u522b\u662f\u50cfGPT\u7cfb\u5217\u6a21\u578b\uff08\u5982GPT-2\u3001GPT-3\uff09\u7684\u51fa\u73b0\uff0c\u663e\u8457\u63a8\u52a8\u4e86\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u3002\u8fd9\u4e9b\u6a21\u578b\u901a\u8fc7\u7b80\u5355\u7684\u201c\u4e0b\u4e00\u4e2atoken\u9884\u6d4b\u201d\u7b56\u7565\uff0c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5230\u5e7f\u6cdb\u7684\u8bed\u8a00\u6a21\u5f0f\uff0c\u5e76\u5c55\u793a\u4e86\u51fa\u8272\u7684\u6269\u5c55\u6027\u548c\u96f6\u6837\u672c\u5b66\u4e60\u80fd\u529b\u3002\u81ea\u56de\u5f52\u6a21\u578b\u7684\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u5c06\u590d\u6742\u7684\u751f\u6210\u95ee\u9898\u5206\u89e3\u4e3a\u4e00\u7cfb\u5217\u987a\u5e8f\u7684\u751f\u6210\u4efb\u52a1\uff0c\u8fd9\u79cd\u65b9\u6cd5\u7684\u6210\u529f\u542f\u53d1\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\uff0c\u5e0c\u671b\u5728\u89c6\u89c9\u751f\u6210\u4e2d\u4e5f\u80fd\u5b9e\u73b0\u7c7b\u4f3c\u7684\u7a81\u7834\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2. \u89c6\u89c9\u751f\u6210\u7684\u6311\u6218<\/h5>\n\n\n\n<p>\u4e0e\u8bed\u8a00\u6570\u636e\u76f8\u6bd4\uff0c\u89c6\u89c9\u6570\u636e\u5177\u6709\u4e8c\u7ef4\u7ed3\u6784\uff0c\u800c\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u6a21\u578b\u4e3b\u8981\u5e94\u7528\u4e8e\u4e00\u7ef4\u5e8f\u5217\u6570\u636e\uff0c\u56e0\u6b64\u5728\u5e94\u7528\u4e8e\u56fe\u50cf\u751f\u6210\u65f6\u5b58\u5728\u663e\u8457\u7684\u6311\u6218\u3002\u65e9\u671f\u7684\u89c6\u89c9\u81ea\u56de\u5f52\u6a21\u578b\u901a\u5e38\u91c7\u7528\u201c\u6805\u683c\u626b\u63cf\u201d\uff08raster scan\uff09\u7684\u65b9\u5f0f\u6765\u5904\u7406\u56fe\u50cf\uff0c\u5373\u5c06\u4e8c\u7ef4\u56fe\u50cf\u5206\u89e3\u4e3a\u4e00\u7ef4token\u5e8f\u5217\uff0c\u6309\u7167\u4ece\u5de6\u5230\u53f3\u3001\u81ea\u4e0a\u800c\u4e0b\u7684\u987a\u5e8f\u9010\u4e2a\u751f\u6210\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u5728\u6269\u5c55\u6027\u548c\u6548\u7387\u4e0a\u5747\u9762\u4e34\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6548\u7387\u4f4e\u4e0b<\/strong>\uff1a\u751f\u6210\u56fe\u50cf\u9700\u8981\u9010\u50cf\u7d20\u8fdb\u884c\uff0c\u5e76\u4e14\u6bcf\u4e2a\u50cf\u7d20\u7684\u751f\u6210\u90fd\u4f9d\u8d56\u4e8e\u4e4b\u524d\u6240\u6709\u50cf\u7d20\uff0c\u5bfc\u81f4\u63a8\u7406\u901f\u5ea6\u975e\u5e38\u6162\uff0c\u5c24\u5176\u662f\u5728\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u751f\u6210\u4e2d\uff0c\u8ba1\u7b97\u590d\u6742\u5ea6\u5448\u6307\u6570\u7ea7\u589e\u957f\u3002<\/li>\n\n\n\n<li><strong>\u7ed3\u6784\u4fe1\u606f\u7684\u4e22\u5931<\/strong>\uff1a\u56fe\u50cf\u5177\u6709\u5929\u7136\u7684\u7a7a\u95f4\u7ed3\u6784\uff0c\u800c\u5c06\u5176\u8f6c\u5316\u4e3a\u4e00\u7ef4\u5e8f\u5217\u540e\uff0c\u8fd9\u79cd\u7a7a\u95f4\u7ed3\u6784\u4fe1\u606f\u88ab\u7834\u574f\uff0c\u96be\u4ee5\u6709\u6548\u6355\u6349\u90bb\u8fd1\u50cf\u7d20\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/li>\n\n\n\n<li><strong>\u53cc\u5411\u76f8\u5173\u6027\u95ee\u9898<\/strong>\uff1a\u56fe\u50cf\u7684\u91cf\u5316\u7f16\u7801\u8fc7\u7a0b\u4e2d\uff0ctoken\u4e4b\u95f4\u5b58\u5728\u53cc\u5411\u7684\u76f8\u4e92\u4f9d\u8d56\uff0c\u4f46\u81ea\u56de\u5f52\u6a21\u578b\u5047\u8bbetoken\u4ec5\u4f9d\u8d56\u4e8e\u5176\u5148\u524d\u7684\u5e8f\u5217\uff0c\u8fd9\u79cd\u5047\u8bbe\u5728\u56fe\u50cf\u751f\u6210\u4e2d\u4e0d\u5b8c\u5168\u6210\u7acb\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">3. \u7814\u7a76\u52a8\u673a<\/h5>\n\n\n\n<p>\u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u89c6\u89c9\u81ea\u56de\u5f52\u5efa\u6a21\u65b9\u6cd5\u2014\u2014<strong>\u89c6\u89c9\u81ea\u56de\u5f52\u5efa\u6a21\uff08Visual AutoRegressive, VAR\uff09<\/strong>\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u201c\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\u9884\u6d4b\u201d\u6765\u5b9e\u73b0\u56fe\u50cf\u7684\u9010\u6b65\u751f\u6210\uff0c\u4ece\u4f4e\u5206\u8fa8\u7387\u5230\u9ad8\u5206\u8fa8\u7387\u9010\u6b65\u9012\u8fdb\uff0c\u4ece\u800c\u66f4\u597d\u5730\u4fdd\u7559\u56fe\u50cf\u7684\u7a7a\u95f4\u7ed3\u6784\u4fe1\u606f\uff0c\u540c\u65f6\u663e\u8457\u63d0\u9ad8\u751f\u6210\u6548\u7387\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e8c\u3001VAR\u6a21\u578b\u7684\u521b\u65b0\u65b9\u6cd5<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">1. \u591a\u5c3a\u5ea6\u751f\u6210\u4e0e\u201c\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\u9884\u6d4b\u201d<\/h5>\n\n\n\n<p>VAR\u7684\u6838\u5fc3\u5728\u4e8e\u5c06\u4f20\u7edf\u7684\u201c\u4e0b\u4e00\u4e2atoken\u9884\u6d4b\u201d\u91cd\u6784\u4e3a\u201c\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\u9884\u6d4b\u201d\u3002\u5177\u4f53\u6765\u8bf4\uff0cVAR\u901a\u8fc7\u4ee5\u4e0b\u6b65\u9aa4\u6765\u5b9e\u73b0\u56fe\u50cf\u751f\u6210\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u5c3a\u5ea6\u91cf\u5316\u7f16\u7801<\/strong>\uff1a\u9996\u5148\uff0c\u56fe\u50cf\u901a\u8fc7\u4e00\u4e2a\u591a\u5c3a\u5ea6\u91cf\u5316\u81ea\u52a8\u7f16\u7801\u5668\uff08VQVAE\uff09\u88ab\u7f16\u7801\u4e3a\u591a\u4e2a\u4e0d\u540c\u5206\u8fa8\u7387\u7684token map\u3002\u8fd9\u4e9btoken map\u4ece\u4f4e\u5206\u8fa8\u7387\u5230\u9ad8\u5206\u8fa8\u7387\u9010\u7ea7\u7ec6\u5316\uff0c\u6700\u7ec8\u8fbe\u5230\u539f\u56fe\u7684\u5206\u8fa8\u7387\u3002<\/li>\n\n\n\n<li><strong>\u4ece\u7c97\u5230\u7ec6\u7684\u81ea\u56de\u5f52\u751f\u6210<\/strong>\uff1a\u751f\u6210\u8fc7\u7a0b\u4ece\u6700\u7c97\u7cd9\u76841\u00d71 token map\u5f00\u59cb\uff0c\u9010\u6b65\u5411\u66f4\u9ad8\u5206\u8fa8\u7387\u6269\u5c55\u3002\u5728\u6bcf\u4e00\u4e2a\u751f\u6210\u6b65\u9aa4\u4e2d\uff0c\u5f53\u524d\u5c3a\u5ea6\u7684token map\u88ab\u7528\u6765\u6761\u4ef6\u751f\u6210\u4e0b\u4e00\u5c3a\u5ea6\u7684token map\uff0c\u8fd9\u79cd\u65b9\u5f0f\u4fdd\u8bc1\u4e86\u751f\u6210\u8fc7\u7a0b\u4e2d\u7a7a\u95f4\u7ed3\u6784\u4fe1\u606f\u7684\u4fdd\u7559\u3002<\/li>\n\n\n\n<li><strong>\u5e76\u884c\u751f\u6210<\/strong>\uff1a\u5728\u6bcf\u4e2a\u5c3a\u5ea6\u5185\uff0c\u6240\u6709token\u662f\u5e76\u884c\u751f\u6210\u7684\uff0c\u8fd9\u4f7f\u5f97\u751f\u6210\u8fc7\u7a0b\u66f4\u9ad8\u6548\uff0c\u76f8\u8f83\u4e8e\u9010\u50cf\u7d20\u7684\u751f\u6210\u65b9\u5f0f\u5927\u5e45\u51cf\u5c11\u4e86\u65f6\u95f4\u590d\u6742\u5ea6\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u79cd\u591a\u5c3a\u5ea6\u751f\u6210\u7b56\u7565\uff0cVAR\u80fd\u591f\u66f4\u6709\u6548\u5730\u5b66\u4e60\u5230\u56fe\u50cf\u7684\u591a\u5c42\u6b21\u7ed3\u6784\u4fe1\u606f\uff0c\u514b\u670d\u4e86\u4f20\u7edf\u65b9\u6cd5\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u9762\u4e34\u7684\u591a\u79cd\u9650\u5236\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2. \u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\u4e0e\u5c42\u6b21\u7f16\u7801\u7b56\u7565<\/h5>\n\n\n\n<p>VAR\u91c7\u7528\u4e86GPT-2\u98ce\u683c\u7684\u89e3\u7801\u5668\u67b6\u6784\uff0c\u4f46\u5728\u5b9e\u73b0\u4e0a\u505a\u4e86\u51e0\u9879\u91cd\u8981\u6539\u8fdb\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\uff08Adaptive Layer Normalization, AdaLN\uff09<\/strong>\uff1a\u4e3a\u4e86\u589e\u5f3a\u6a21\u578b\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u751f\u6210\u8d28\u91cf\uff0cVAR\u4f7f\u7528\u4e86\u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\u6280\u672f\u3002\u76f8\u6bd4\u6807\u51c6\u7684\u5c42\u5f52\u4e00\u5316\uff0cAdaLN\u80fd\u591f\u6839\u636e\u6761\u4ef6\u4fe1\u606f\uff08\u4f8b\u5982\u7c7b\u522b\u6807\u7b7e\uff09\u8fdb\u884c\u5f52\u4e00\u5316\u8c03\u6574\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u5904\u7406\u4e0d\u540c\u7c7b\u578b\u7684\u8f93\u5165\u3002<\/li>\n\n\n\n<li><strong>\u5171\u4eab\u7801\u672c\u7684\u591a\u5c3a\u5ea6\u91cf\u5316<\/strong>\uff1aVAR\u5728\u4e0d\u540c\u5c3a\u5ea6\u95f4\u4f7f\u7528\u4e86\u5171\u4eab\u7684\u91cf\u5316\u7801\u672c\uff08codebook\uff09\uff0c\u8fd9\u6709\u52a9\u4e8e\u51cf\u5c11\u5c3a\u5ea6\u8f6c\u6362\u65f6\u7684\u4fe1\u606f\u4e22\u5931\uff0c\u4f7f\u5f97\u5404\u4e2a\u5c3a\u5ea6\u4e4b\u95f4\u7684\u4fe1\u606f\u80fd\u591f\u66f4\u81ea\u7136\u5730\u4f20\u9012\u548c\u878d\u5408\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">3. \u751f\u6210\u8fc7\u7a0b\u7684\u590d\u6742\u5ea6\u5206\u6790<\/h5>\n\n\n\n<p>VAR\u76f8\u6bd4\u4e8e\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u6a21\u578b\uff0c\u6781\u5927\u5730\u964d\u4f4e\u4e86\u751f\u6210\u8fc7\u7a0b\u7684\u590d\u6742\u5ea6\u3002\u5728\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u6a21\u578b\u4e2d\uff0c\u751f\u6210n\u00d7n\u5206\u8fa8\u7387\u7684\u56fe\u50cf\u9700\u8981O(n^6)\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u800cVAR\u901a\u8fc7\u201c\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\u9884\u6d4b\u201d\u7b56\u7565\uff0c\u5c06\u5176\u590d\u6742\u5ea6\u964d\u4f4e\u81f3O(n^4)\u3002\u5177\u4f53\u800c\u8a00\uff0cVAR\u5728\u6bcf\u4e2a\u5c3a\u5ea6\u4e0a\u5e76\u884c\u751f\u6210token\uff0c\u800c\u4e0d\u662f\u9010\u4e2a\u751f\u6210\uff0c\u8fd9\u4f7f\u5f97\u6574\u4f53\u751f\u6210\u8fc7\u7a0b\u66f4\u52a0\u9ad8\u6548\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u591a\u5c3a\u5ea6\u5e76\u884c\u751f\u6210\uff0cVAR\u5728\u6bcf\u4e2a\u751f\u6210\u6b65\u9aa4\u4e2d\u90fd\u4fdd\u6301\u4e86\u56fe\u50cf\u7684\u7a7a\u95f4\u7ed3\u6784\u5b8c\u6574\u6027\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u751f\u6210\u8d28\u91cf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e09\u3001\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5bf9\u6bd4\u5206\u6790<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">1. \u4e0e\u5176\u4ed6\u751f\u6210\u6a21\u578b\u7684\u6bd4\u8f83<\/h5>\n\n\n\n<p>VAR\u6a21\u578b\u5728ImageNet 256\u00d7256\u548c512\u00d7512\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u7279\u522b\u662f\u5728\u751f\u6210\u8d28\u91cf\u3001\u6570\u636e\u6548\u7387\u548c\u63a8\u7406\u901f\u5ea6\u7b49\u65b9\u9762\uff0c\u76f8\u8f83\u4e8e\u5176\u4ed6\u751f\u6210\u6a21\u578b\uff08\u5305\u62ecGAN\u3001\u6269\u6563\u6a21\u578b\u3001BERT\u98ce\u683c\u7684\u63a9\u7801\u9884\u6d4b\u6a21\u578b\u548c\u4f20\u7edf\u81ea\u56de\u5f52\u6a21\u578b\uff09\u5177\u6709\u663e\u8457\u4f18\u52bf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u751f\u6210\u8d28\u91cf<\/strong>\uff1aVAR\u5728ImageNet 256\u00d7256\u6570\u636e\u96c6\u4e0a\u7684Fr\u00e9chet Inception Distance\uff08FID\uff09\u8fbe\u52301.73\uff0c\u8fd9\u4e2a\u7ed3\u679c\u8d85\u8fc7\u4e86\u8bb8\u591a\u9886\u5148\u7684\u6269\u6563\u6a21\u578b\uff08\u4f8b\u5982DiT-L\/2\u3001L-DiT-3B\u7b49\uff09\uff0c\u5e76\u9996\u6b21\u4f7f\u5f97\u81ea\u56de\u5f52\u6a21\u578b\u5728\u751f\u6210\u8d28\u91cf\u4e0a\u8d85\u8d8a\u4e86\u6269\u6563\u6a21\u578b\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u901f\u5ea6<\/strong>\uff1aVAR\u7684\u63a8\u7406\u901f\u5ea6\u6bd4\u4f20\u7edf\u7684\u81ea\u56de\u5f52\u6a21\u578b\uff08\u5982VQGAN\u3001ViT-VQGAN\uff09\u5feb\u7ea620\u500d\uff0c\u8fbe\u5230\u4e86\u4e0e\u9ad8\u6548GAN\u6a21\u578b\u76f8\u5f53\u7684\u901f\u5ea6\u3002\u8fd9\u79cd\u901f\u5ea6\u63d0\u5347\u5f97\u76ca\u4e8e\u591a\u5c3a\u5ea6\u5e76\u884c\u751f\u6210\u7684\u7b56\u7565\uff0c\u4f7f\u5f97\u6bcf\u6b21\u751f\u6210\u7684\u8ba1\u7b97\u91cf\u663e\u8457\u51cf\u5c11\u3002<\/li>\n\n\n\n<li><strong>\u6570\u636e\u6548\u7387<\/strong>\uff1a\u4e0e\u6269\u6563\u6a21\u578b\uff08\u5982DiT\u7cfb\u5217\uff09\u76f8\u6bd4\uff0cVAR\u6a21\u578b\u53ea\u9700\u8f83\u5c11\u7684\u8bad\u7ec3\u8f6e\u6570\uff08350\u8f6e\u76f8\u8f83\u4e8eDiT-XL\/2\u76841400\u8f6e\uff09\u5c31\u80fd\u8fbe\u5230\u66f4\u597d\u7684\u751f\u6210\u8d28\u91cf\uff0c\u4f53\u73b0\u4e86\u66f4\u9ad8\u7684\u6570\u636e\u5229\u7528\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">2. \u6807\u5ea6\u89c4\u5f8b\uff08Scaling Laws\uff09\u9a8c\u8bc1<\/h5>\n\n\n\n<p>VAR\u6a21\u578b\u5c55\u793a\u4e86\u4e0e\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7c7b\u4f3c\u7684\u6807\u5ea6\u89c4\u5f8b\uff0c\u5373\u6a21\u578b\u6027\u80fd\u968f\u7740\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\u5448\u73b0\u51fa\u5e42\u5f8b\u4e0b\u964d\u8d8b\u52bf\u3002\u5b9e\u9a8c\u4e2d\uff0c\u4f5c\u8005\u8bad\u7ec3\u4e86\u4e0d\u540c\u53c2\u6570\u91cf\u7684VAR\u6a21\u578b\uff08\u4ece18M\u52302B\u53c2\u6570\uff09\uff0c\u5e76\u89c2\u5bdf\u5176\u5728ImageNet\u9a8c\u8bc1\u96c6\u4e0a\u7684\u6d4b\u8bd5\u635f\u5931\u3002\u7ed3\u679c\u8868\u660e\uff0cVAR\u6a21\u578b\u7684\u6d4b\u8bd5\u635f\u5931\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\u800c\u5e73\u6ed1\u4e0b\u964d\uff0c\u4e14\u8fd9\u79cd\u4e0b\u964d\u8d8b\u52bf\u7b26\u5408\u5e42\u5f8b\u5173\u7cfb\uff0c\u8bf4\u660eVAR\u6a21\u578b\u5177\u5907\u826f\u597d\u7684\u53ef\u6269\u5c55\u6027\u3002\u8fd9\u79cd\u6807\u5ea6\u89c4\u5f8b\u7684\u5b58\u5728\uff0c\u4f7f\u5f97\u53ef\u4ee5\u901a\u8fc7\u5c0f\u6a21\u578b\u7684\u8bad\u7ec3\u7ed3\u679c\u6765\u9884\u6d4b\u5927\u6a21\u578b\u7684\u6027\u80fd\uff0c\u4ece\u800c\u66f4\u6709\u6548\u5730\u8fdb\u884c\u8d44\u6e90\u914d\u7f6e\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3. \u96f6\u6837\u672c\u4efb\u52a1\u6cdb\u5316\u80fd\u529b<\/h5>\n\n\n\n<p>VAR\u6a21\u578b\u5728\u56fe\u50cf\u8865\u5168\uff08in-painting\uff09\u3001\u6269\u5c55\u7ed8\u5236\uff08out-painting\uff09\u548c\u56fe\u50cf\u7f16\u8f91\u7b49\u4efb\u52a1\u4e0a\u5c55\u793a\u4e86\u826f\u597d\u7684\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\u3002\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e2d\uff0cVAR\u80fd\u591f\u5728\u4e0d\u9700\u8981\u989d\u5916\u8c03\u4f18\u7684\u60c5\u51b5\u4e0b\uff0c\u76f4\u63a5\u751f\u6210\u7b26\u5408\u4e0a\u4e0b\u6587\u7684\u5185\u5bb9\uff0c\u8bf4\u660e\u6a21\u578b\u5177\u5907\u8f83\u5f3a\u7684\u901a\u7528\u6027\u548c\u9002\u5e94\u6027\u3002\u4f8b\u5982\uff0cVAR\u5728\u56fe\u50cf\u7f16\u8f91\u4efb\u52a1\u4e2d\u80fd\u591f\u6839\u636e\u7ed9\u5b9a\u7684\u7c7b\u522b\u6807\u7b7e\u751f\u6210\u7b26\u5408\u8bed\u4e49\u7684\u5185\u5bb9\uff0c\u5e76\u4e14\u4e0e\u5468\u56f4\u672a\u7f16\u8f91\u90e8\u5206\u7684\u5185\u5bb9\u81ea\u7136\u878d\u5408\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u56db\u3001\u6a21\u578b\u8bbe\u8ba1\u4e0e\u8bad\u7ec3\u7ec6\u8282<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">1. VAR\u91cf\u5316\u5668\uff08Tokenizer\uff09<\/h5>\n\n\n\n<p>VAR\u6a21\u578b\u91c7\u7528\u4e86\u4e00\u79cd\u6539\u8fdb\u7684\u591a\u5c3a\u5ea6\u91cf\u5316\u81ea\u52a8\u7f16\u7801\u5668\uff08VQVAE\uff09\u6765\u5bf9\u56fe\u50cf\u8fdb\u884c\u591a\u5c3a\u5ea6\u7f16\u7801\u3002\u5177\u4f53\u6765\u8bf4\uff0cVAR\u91cf\u5316\u5668\u9996\u5148\u901a\u8fc7\u7f16\u7801\u5668\u5c06\u56fe\u50cf\u7279\u5f81\u6620\u5c04\u5230\u591a\u5c3a\u5ea6token map\u4e0a\uff0c\u6bcf\u4e2a\u5c3a\u5ea6\u7684token map\u90fd\u4f7f\u7528\u76f8\u540c\u7684\u5171\u4eab\u7801\u672c\u8fdb\u884c\u91cf\u5316\uff0c\u8fd9\u786e\u4fdd\u4e86\u4e0d\u540c\u5c3a\u5ea6\u95f4\u7684\u4e00\u81f4\u6027\u548c\u4fe1\u606f\u4f20\u9012\u7684\u6709\u6548\u6027\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u51cf\u5c11\u4fe1\u606f\u5728\u5c3a\u5ea6\u8f6c\u6362\u4e2d\u7684\u4e22\u5931\uff0c\u4f5c\u8005\u5f15\u5165\u4e86\u6b8b\u5dee\u98ce\u683c\u7684\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u5728\u6bcf\u4e2a\u5c3a\u5ea6\u4e0a\u91cf\u5316\u540e\u7684\u7279\u5f81\u53ef\u4ee5\u66f4\u597d\u5730\u4fdd\u6301\u4e0e\u539f\u59cb\u56fe\u50cf\u7684\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2. VAR Transformer\u67b6\u6784<\/h5>\n\n\n\n<p>VAR\u6a21\u578b\u7684\u81ea\u56de\u5f52\u751f\u6210\u90e8\u5206\u91c7\u7528\u4e86\u6807\u51c6\u7684GPT-2\u98ce\u683c\u7684\u89e3\u7801\u5668\u67b6\u6784\uff0c\u5e76\u7ed3\u5408\u4e86\u4ee5\u4e0b\u6280\u672f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\uff08AdaLN\uff09<\/strong>\uff1a\u901a\u8fc7\u5728\u6bcf\u4e00\u5c42\u4f7f\u7528\u81ea\u9002\u5e94\u5c42\u5f52\u4e00\u5316\uff0cVAR\u80fd\u591f\u6839\u636e\u8f93\u5165\u7684\u7c7b\u522b\u4fe1\u606f\u8fdb\u884c\u5f52\u4e00\u5316\u8c03\u6574\uff0c\u4ece\u800c\u63d0\u9ad8\u751f\u6210\u8fc7\u7a0b\u7684\u7a33\u5b9a\u6027\u548c\u751f\u6210\u56fe\u50cf\u7684\u591a\u6837\u6027\u3002<\/li>\n\n\n\n<li><strong>\u67e5\u8be2\u548c\u952e\u7684\u5f52\u4e00\u5316<\/strong>\uff1a\u4e3a\u4e86\u589e\u5f3a\u6a21\u578b\u5728\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u7684\u7a33\u5b9a\u6027\uff0cVAR\u5bf9\u67e5\u8be2\u548c\u952e\u8fdb\u884c\u4e86\u5355\u4f4d\u5411\u91cf\u5f52\u4e00\u5316\uff0c\u8fd9\u79cd\u64cd\u4f5c\u4f7f\u5f97\u6a21\u578b\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u80fd\u591f\u66f4\u52a0\u7a33\u5b9a\u5730\u6355\u6349\u7279\u5f81\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e94\u3001\u5b9e\u9a8c\u7ed3\u679c\u7684\u53ef\u89c6\u5316\u4e0e\u5206\u6790<\/h4>\n\n\n\n<p>\u4f5c\u8005\u901a\u8fc7\u5bf9\u4e0d\u540c\u5927\u5c0f\u7684VAR\u6a21\u578b\uff08\u4ece\u6df1\u5ea66\u523030\uff09\u5728\u4e0d\u540c\u8bad\u7ec3\u9636\u6bb5\uff0820%\u300160%\u3001100%\u8bad\u7ec3\u5b8c\u6210\uff09\u751f\u6210\u7684\u6837\u672c\u8fdb\u884c\u53ef\u89c6\u5316\uff0c\u5c55\u793a\u4e86\u6a21\u578b\u89c4\u6a21\u548c\u8bad\u7ec3\u91cf\u5bf9\u751f\u6210\u8d28\u91cf\u7684\u5f71\u54cd\u3002\u968f\u7740\u6a21\u578b\u53c2\u6570\u91cf\u7684\u589e\u52a0\u548c\u8bad\u7ec3\u91cf\u7684\u589e\u52a0\uff0c\u751f\u6210\u56fe\u50cf\u7684\u89c6\u89c9\u4fdd\u771f\u5ea6\u548c\u5185\u5bb9\u4e00\u81f4\u6027\u663e\u8457\u63d0\u9ad8\u3002\u8fd9\u4e9b\u89c2\u5bdf\u4e0e\u6807\u5ea6\u89c4\u5f8b\u7684\u7ed3\u8bba\u4e00\u81f4\uff0c\u8868\u660e\u66f4\u5927\u7684\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u5b66\u4e60\u5230\u590d\u6742\u7684\u56fe\u50cf\u5206\u5e03\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u516d\u3001\u9650\u5236\u4e0e\u672a\u6765\u7814\u7a76\u65b9\u5411<\/h4>\n\n\n\n<p>\u5c3d\u7ba1VAR\u6a21\u578b\u5728\u56fe\u50cf\u751f\u6210\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u4f5c\u8005\u4e5f\u6307\u51fa\u4e86\u4e00\u4e9b\u9700\u8981\u6539\u8fdb\u7684\u5730\u65b9\u548c\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6539\u8fdb\u91cf\u5316\u5668<\/strong>\uff1a\u5f53\u524dVAR\u6a21\u578b\u91c7\u7528\u7684\u662f\u4e0eVQGAN\u7c7b\u4f3c\u7684VQVAE\u7ed3\u6784\uff0c\u672a\u6765\u53ef\u4ee5\u5c1d\u8bd5\u4f7f\u7528\u66f4\u5148\u8fdb\u7684\u91cf\u5316\u5668\u6765\u8fdb\u4e00\u6b65\u63d0\u5347\u751f\u6210\u8d28\u91cf\u3002<\/li>\n\n\n\n<li><strong>\u6587\u672c\u63d0\u793a\u751f\u6210<\/strong>\uff1a\u76ee\u524d\u7684\u7814\u7a76\u4e3b\u8981\u96c6\u4e2d\u5728\u56fe\u50cf\u751f\u6210\u4e0a\uff0c\u800c\u901a\u8fc7\u4e0e\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u7ed3\u5408\uff0cVAR\u6a21\u578b\u53ef\u4ee5\u6269\u5c55\u5230\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u4e2d\uff0c\u5b9e\u73b0\u66f4\u590d\u6742\u7684\u591a\u6a21\u6001\u751f\u6210\u3002<\/li>\n\n\n\n<li><strong>\u89c6\u9891\u751f\u6210<\/strong>\uff1aVAR\u7684\u591a\u5c3a\u5ea6\u751f\u6210\u65b9\u6cd5\u4e5f\u53ef\u4ee5\u81ea\u7136\u5730\u6269\u5c55\u5230\u89c6\u9891\u751f\u6210\u4e2d\uff0c\u901a\u8fc7\u8003\u8651\u89c6\u9891\u7684\u65f6\u95f4\u7ef4\u5ea6\uff0c\u4f7f\u7528\u7c7b\u4f3c\u7684\u201c3D\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\u9884\u6d4b\u201d\u65b9\u6cd5\uff0c\u53ef\u4ee5\u751f\u6210\u65f6\u7a7a\u4e00\u81f4\u6027\u8f83\u5f3a\u7684\u89c6\u9891\u5185\u5bb9\u3002\u8fd9\u5728\u957f\u65f6\u5e8f\u4f9d\u8d56\u7684\u89c6\u9891\u751f\u6210\u4e2d\u6709\u8f83\u5927\u6f5c\u529b\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Visual Autoregressive Modeling: Scalable Image Genera [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,63,95,90,80],"class_list":{"0":"post-5269","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-tech","7":"tag-ai","9":"tag-transformer","10":"tag-multimodal","11":"tag-cv"},"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Visual Autoregressive Modeling: Scalable Image Genera&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5269","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5269"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5269\/revisions"}],"predecessor-version":[{"id":5271,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5269\/revisions\/5271"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5269"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5269"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5269"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}