{"id":858,"date":"2026-01-09T02:52:20","date_gmt":"2026-01-08T18:52:20","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=858"},"modified":"2026-01-12T10:20:26","modified_gmt":"2026-01-12T02:20:26","slug":"visual-autoregressive-modeling-scalable-image-generation-via-next-scale-prediction","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=858","title":{"rendered":"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\"><strong>1. \u7814\u7a76\u52a8\u673a\u4e0e\u80cc\u666f<\/strong><\/h1>\n\n\n\n<p>\u4f20\u7edf\u7684\u89c6\u89c9 <strong>\u81ea\u56de\u5f52\uff08Autoregressive\uff0cAR\uff09\u751f\u6210\u65b9\u6cd5<\/strong>\u901a\u5e38\u6309\u5149\u6805\u626b\u63cf\uff08raster-scan\uff09\u7684\u987a\u5e8f\u9010\u50cf\u7d20\u6216\u9010 patch \u9884\u6d4b\u4e0b\u4e00\u4e2a token\uff0c\u800c\u8fd9\u79cd\u7b56\u7565\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5ffd\u7565\u56fe\u50cf\u7684\u4e8c\u7ef4\u7ed3\u6784\uff1b<\/li>\n\n\n\n<li>\u96be\u4ee5\u5e76\u884c\u8ba1\u7b97\uff1b<\/li>\n\n\n\n<li>\u5728\u751f\u6210\u8d28\u91cf\u548c\u751f\u6210\u901f\u5ea6\u4e0a\u843d\u540e\u4e8e\u6269\u6563\u6a21\u578b\uff08\u5982 Stable Diffusion \/ DiT\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u81ea\u56de\u5f52\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\uff08\u4f8b\u5982 GPT \u7cfb\u5217\uff09\u5c55\u73b0\u51fa<strong>\u53ef\u6269\u5c55\u6027\uff08scaling laws\uff09\u548c\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b<\/strong>\u7b49\u4f18\u52bf\uff0c\u4f46\u8fd9\u4e9b\u7279\u6027\u5c1a\u672a\u5728\u89c6\u89c9\u751f\u6210\u9886\u57df\u5f97\u5230\u5145\u5206\u5229\u7528\u3002<\/p>\n\n\n\n<p>\u672c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u5168\u65b0\u89c6\u89c9\u751f\u6210\u8303\u5f0f <strong>Visual AutoRegressive Modeling\uff08VAR\uff09<\/strong>\uff0c\u901a\u8fc7\u4ece\u7c97\u5230\u7ec6\u7684**\u5c3a\u5ea6\u9884\u6d4b\uff08next-scale prediction\uff09**\u66ff\u4ee3\u4f20\u7edf\u5149\u6805\u9884\u6d4b\uff0c\u4f7f\u81ea\u56de\u5f52\u6a21\u578b\u80fd\u5728\u56fe\u50cf\u751f\u6210\u4e2d\u540c\u65f6\u517c\u987e\u8d28\u91cf\u548c\u9ad8\u6548\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"518\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-8-1024x518.png\"  class=\"wp-image-859\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-8-1024x518.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-8-300x152.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-8-768x389.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-8.png 1249w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe\" \/><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"666\" height=\"638\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-9.png\"  class=\"wp-image-860\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-9.png 666w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-9-300x287.png 300w\" sizes=\"auto, (max-width: 666px) 100vw, 666px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe1\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>2. \u6838\u5fc3\u65b9\u6cd5\uff1aVAR\uff08Visual AutoRegressive Modeling\uff09<\/strong><\/h1>\n\n\n\n<p>VAR \u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u5c06\u56fe\u50cf\u751f\u6210\u4efb\u52a1\u770b\u6210\u4e00\u7cfb\u5217\u201c<strong>\u4ece\u4f4e\u5206\u8fa8\u7387\u5411\u9ad8\u5206\u8fa8\u7387\u7684\u9884\u6d4b<\/strong>\u201d\uff0c\u800c\u4e0d\u662f\u4e00\u4e2a\u957f\u5e8f\u5217 token \u7684\u9010\u70b9\u751f\u6210\u529f\u80fd\u3002<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">2.1 <strong>\u4ece\u7c97\u5230\u7ec6\u7684 Next-Scale \u9884\u6d4b<\/strong><\/h2>\n\n\n\n<p>VAR \u5c06\u56fe\u50cf\u5212\u5206\u6210\u591a\u4e2a\u5c3a\u5ea6\uff08scale\uff09\u8868\u793a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a\u5c3a\u5ea6\u5bf9\u5e94\u56fe\u50cf\u7684\u4e0d\u540c\u5206\u8fa8\u7387 token map\uff1b<\/li>\n\n\n\n<li>\u6a21\u578b\u6309\u7167\u5c3a\u5ea6\u987a\u5e8f\uff08\u4f8b\u5982 1\u00d71 \u2192 4\u00d74 \u2192 16\u00d716 \u2192 64\u00d764\uff09\u81ea\u56de\u5f52\u5730\u9884\u6d4b\u4e0b\u4e00\u7ea7\u66f4\u9ad8\u5206\u8fa8\u7387\u7684 token map\uff1b<\/li>\n\n\n\n<li>\u6a21\u578b\u4e0d\u518d\u6309\u50cf\u7d20\u987a\u5e8f\u9884\u6d4b\u5355\u4e2a token\uff0c\u800c\u662f\u6309<strong>\u6574\u4e2a\u5c3a\u5ea6 token map<\/strong>\u4f5c\u4e3a\u4e00\u4e2a\u81ea\u56de\u5f52\u5355\u5143\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5982\u6570\u5b66\u8868\u8fbe\uff1a<math display=\"block\"><semantics><mrow><mi>p<\/mi><mo stretchy=\"false\">(<\/mo><msub><mi>r<\/mi><mn>1<\/mn><\/msub><mo separator=\"true\">,<\/mo><msub><mi>r<\/mi><mn>2<\/mn><\/msub><mo separator=\"true\">,<\/mo><mo>\u2026<\/mo><mo separator=\"true\">,<\/mo><msub><mi>r<\/mi><mi>K<\/mi><\/msub><mo stretchy=\"false\">)<\/mo><mo>=<\/mo><munderover><mo>\u220f<\/mo><mrow><mi>k<\/mi><mo>=<\/mo><mn>1<\/mn><\/mrow><mi>K<\/mi><\/munderover><mi>p<\/mi><mo stretchy=\"false\">(<\/mo><msub><mi>r<\/mi><mi>k<\/mi><\/msub><mo>\u2223<\/mo><msub><mi>r<\/mi><mn>1<\/mn><\/msub><mo separator=\"true\">,<\/mo><msub><mi>r<\/mi><mn>2<\/mn><\/msub><mo separator=\"true\">,<\/mo><mo>\u2026<\/mo><mo separator=\"true\">,<\/mo><msub><mi>r<\/mi><mrow><mi>k<\/mi><mo>\u2212<\/mo><mn>1<\/mn><\/mrow><\/msub><mo stretchy=\"false\">)<\/mo><mo separator=\"true\">,<\/mo><\/mrow><annotation encoding=\"application\/x-tex\">p(r_1, r_2, \\ldots, r_K) = \\prod_{k=1}^K p(r_k \\mid r_1, r_2, \\ldots, r_{k-1}),<\/annotation><\/semantics><\/math>p(r1\u200b,r2\u200b,\u2026,rK\u200b)=k=1\u220fK\u200bp(rk\u200b\u2223r1\u200b,r2\u200b,\u2026,rk\u22121\u200b),<\/p>\n\n\n\n<p>\u5176\u4e2d <math><semantics><mrow><msub><mi>r<\/mi><mi>k<\/mi><\/msub><\/mrow><annotation encoding=\"application\/x-tex\">r_k<\/annotation><\/semantics><\/math>rk\u200b \u662f\u7b2c k \u4e2a\u5c3a\u5ea6\u7684 token map\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h2 class=\"wp-block-heading\">2.2 <strong>\u591a\u5c3a\u5ea6\u7f16\u7801\uff08Multi-Scale Quantization\uff09<\/strong><\/h2>\n\n\n\n<p>\u4e3a\u4e86\u652f\u6301\u4e0a\u8ff0\u673a\u5236\uff0cVAR \u5f15\u5165\u4e86\u4e00\u4e2a\u591a\u5c3a\u5ea6\u91cf\u5316\u7f16\u7801\u6a21\u5757\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528\u7c7b\u4f3c\u4e8e VQ-VAE \u7684\u673a\u5236\uff0c\u5c06\u539f\u56fe\u7f16\u7801\u6210\u5206\u5c42 token maps\uff1b<\/li>\n\n\n\n<li>\u4e0d\u540c\u5c3a\u5ea6\u7684 token maps \u5171\u4eab\u4e00\u4e2a codebook\uff0c\u4f7f\u5f97\u8de8\u5c3a\u5ea6\u4fdd\u6301\u4e00\u81f4\u7684 token vocabulary\uff1b<\/li>\n\n\n\n<li>\u7f16\u7801\u5668\u8d1f\u8d23\u5c06\u56fe\u50cf\u6620\u5c04\u4e3a\u4e00\u7cfb\u5217\u5c3a\u5ea6\u7684 token map\uff0c\u800c\u89e3\u7801\u5668\u8d1f\u8d23\u5c06\u6700\u540e\u9884\u6d4b\u7684 token maps \u590d\u539f\u4e3a\u56fe\u50cf\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u5206\u5c42\u8868\u793a\u65b9\u5f0f\u6709\u5229\u4e8e\u6355\u6349\u5168\u5c40\u7ed3\u6784\uff08\u5728\u4f4e\u5c3a\u5ea6\uff09\u548c\u7ec6\u8282\u4fe1\u606f\uff08\u5728\u9ad8\u5c3a\u5ea6\uff09\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u6709\u6548\u6a21\u62df\u81ea\u7136\u56fe\u50cf\u7684\u751f\u6210\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"329\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-11-1024x329.png\"  class=\"wp-image-862\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-11-1024x329.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-11-300x96.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-11-768x247.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-11.png 1126w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe2\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h2 class=\"wp-block-heading\">2.3 <strong>VAR Transformer \u67b6\u6784<\/strong><\/h2>\n\n\n\n<p>\u5728\u7f16\u7801\u5b8c\u6210\u540e\uff0c\u7528\u4e00\u4e2a<strong>\u81ea\u56de\u5f52 Transformer \u6a21\u578b<\/strong>\u9884\u6d4b\u4e0b\u4e00\u4e2a\u5c3a\u5ea6\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\uff1a\u5f53\u524d\u5df2\u6709\u7684\u4f4e\u5c3a\u5ea6 token maps\uff1b<\/li>\n\n\n\n<li>\u8f93\u51fa\uff1a\u4e0b\u4e00\u7ea7\u5c3a\u5ea6\u6240\u6709 token\uff1b<\/li>\n\n\n\n<li>\u6ce8\u610f\u529b\u673a\u5236\u8bbe\u8ba1\u4fdd\u7559\u5c3a\u5ea6\u4e4b\u95f4\u7684\u81ea\u56de\u5f52\u5173\u7cfb\uff0c\u540c\u65f6\u5141\u8bb8\u540c\u5c3a\u5ea6\u5185\u7684\u5e76\u884c\u9884\u6d4b\uff1b<\/li>\n\n\n\n<li>\u8bad\u7ec3\u635f\u5931\u91c7\u7528\u6807\u51c6\u7684\u4ea4\u53c9\u71b5\uff08cross-entropy\uff09\u635f\u5931\u6765\u6700\u5927\u5316\u6761\u4ef6\u6982\u7387 <math><semantics><mrow><mi>p<\/mi><mo stretchy=\"false\">(<\/mo><msub><mi>r<\/mi><mi>k<\/mi><\/msub><mo>\u2223<\/mo><msub><mi>r<\/mi><mrow><mo>&lt;<\/mo><mi>k<\/mi><\/mrow><\/msub><mo stretchy=\"false\">)<\/mo><\/mrow><annotation encoding=\"application\/x-tex\">p(r_k \\mid r_{&lt;k})<\/annotation><\/semantics><\/math>p(rk\u200b\u2223r&lt;k\u200b)\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u7ed3\u6784\u533a\u522b\u4e8e\u4f20\u7edf AR \u6a21\u578b\u9010 token \u987a\u5e8f\u9884\u6d4b\u7684\u7c92\u5ea6\uff0c\u4f7f\u9884\u6d4b\u8fc7\u7a0b\u66f4\u9ad8\u6548\u3001\u5c42\u6b21\u66f4\u6e05\u6670\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"490\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-10-1024x490.png\"  class=\"wp-image-861\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-10-1024x490.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-10-300x144.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-10-768x367.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-10.png 1068w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe3\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>3. \u4e0e\u73b0\u6709\u65b9\u6cd5\u7684\u6bd4\u8f83<\/strong><\/h1>\n\n\n\n<p>\u672c\u8bba\u6587\u4ece\u591a\u4e2a\u7ef4\u5ea6\u5bf9 VAR \u4e0e\u5176\u5b83\u751f\u6210\u6a21\u578b\u8fdb\u884c\u4e86\u8bc4\u4f30\uff1a<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3.1 <strong>\u4e0e\u4f20\u7edf AR \u6a21\u578b\u5bf9\u6bd4<\/strong><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>VAR \u76f8\u6bd4\u4f20\u7edf AR baseline \u5728 ImageNet 256\u00d7256 \u6570\u636e\u96c6\u4e0a\u6709\u663e\u8457\u63d0\u5347\uff1a\n<ul class=\"wp-block-list\">\n<li>Fr\u00e9chet Inception Distance (FID) \u4ece 18.65 \u63d0\u5347\u5230 1.73\uff1b<\/li>\n\n\n\n<li>Inception Score (IS) \u4ece 80.4 \u63d0\u5347\u5230 350.2\uff1b<\/li>\n\n\n\n<li>\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u7ea6 20\u00d7\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u63d0\u5347\u8868\u660e\u91c7\u7528\u5c3a\u5ea6\u9884\u6d4b\u4e0d\u4ec5\u4fdd\u8bc1\u4e86\u751f\u6210\u8d28\u91cf\uff0c\u8fd8\u6781\u5927\u63d0\u9ad8\u4e86\u6548\u7387\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3.2 <strong>\u4e0e\u6269\u6563 Transformer\uff08\u5982 DiT\uff09\u5bf9\u6bd4<\/strong><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u56fe\u50cf\u8d28\u91cf\u3001\u63a8\u7406\u901f\u5ea6\u3001\u6570\u636e\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\u7b49\u591a\u4e2a\u6307\u6807\u4e0a\uff0cVAR \u90fd\u4f18\u4e8e\u6269\u6563\u5f0f\u6a21\u578b\uff1b<\/li>\n\n\n\n<li>\u6269\u6563\u6a21\u578b\u867d\u7136\u751f\u6210\u8d28\u91cf\u9ad8\uff0c\u4f46\u63a8\u7406\u901f\u5ea6\u6162\u3001\u9700\u8f83\u591a\u91c7\u6837\u6b65\u9aa4\uff0cVAR \u5728\u8fd9\u4e9b\u65b9\u9762\u5c55\u73b0\u66f4\u4f18\u5e73\u8861\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>4. Scaling Laws \u4e0e\u6cdb\u5316\u80fd\u529b<\/strong><\/h1>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 <strong>Scaling Laws\uff08\u6269\u5c55\u5b9a\u5f8b\uff09<\/strong><\/h3>\n\n\n\n<p>\u8bba\u6587\u5b9e\u9a8c\u8bc1\u660e VAR \u5728\u6a21\u578b\u89c4\u6a21\u4e0e compute \u8d44\u6e90\u53d8\u5316\u65f6\u9075\u5faa<strong>\u5e42\u5f8b\u6269\u5c55\u5b9a\u5f8b\uff08power-law scaling laws\uff09<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u968f\u7740\u6a21\u578b\u53c2\u6570\u91cf\u548c\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u7684\u6269\u5927\uff0c\u751f\u6210\u8d28\u91cf\uff08\u5982 FID\/IS\uff09\u5448\u73b0\u6709\u89c4\u5f8b\u7684\u7ebf\u6027\u63d0\u5347\uff1b<\/li>\n\n\n\n<li>Pearson \u76f8\u5173\u7cfb\u6570\u8fd1 \u22120.998\uff0c\u8bf4\u660e\u8fd9\u79cd\u5b9a\u5f8b\u975e\u5e38\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e00\u70b9\u4e0e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728 natural language \u4efb\u52a1\u4e0a\u7684 scaling laws \u6781\u4e3a\u76f8\u4f3c\uff0c\u8868\u660e\u89c6\u89c9\u81ea\u56de\u5f52\u6a21\u578b\u5177\u6709\u7c7b\u4f3c\u901a\u7528\u6027\u548c\u53ef\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 <strong>\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\uff08Zero-Shot Generalization\uff09<\/strong><\/h3>\n\n\n\n<p>VAR \u5728\u4e0b\u5217\u4e0b\u6e38\u4efb\u52a1\u4e0a\u5747\u5c55\u793a\u4e86<strong>\u65e0\u9700\u989d\u5916\u8bad\u7ec3<\/strong>\u7684\u6cdb\u5316\u80fd\u529b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u56fe\u50cf <em>in-painting<\/em>\uff08\u5c40\u90e8\u4fee\u590d\uff09\uff1b<\/li>\n\n\n\n<li><em>out-painting<\/em>\uff08\u8fb9\u754c\u6269\u5c55\u751f\u6210\uff09\uff1b<\/li>\n\n\n\n<li>\u56fe\u50cf <em>editing<\/em>\uff08\u7f16\u8f91\u8bed\u4e49\u6539\u53d8\uff09\u7b49\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u8bf4\u660e VAR \u4e0d\u4ec5\u9002\u5408\u7eaf\u56fe\u50cf\u751f\u6210\uff0c\u8fd8\u53ef\u4f5c\u4e3a\u901a\u7528\u89c6\u89c9\u751f\u6210\u6a21\u578b\u7684\u4e00\u90e8\u5206\uff0c\u7c7b\u4f3c LLM \u5728\u8bed\u8a00\u4efb\u52a1\u7684\u901a\u7528\u6027\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>5. \u6570\u5b66\u4e0e\u8bad\u7ec3\u7ec6\u8282<\/strong><\/h1>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u5c3a\u5ea6\u7f16\u7801\u5668\u8bad\u7ec3<\/strong>\uff1a\u5b66\u4e60\u4ece\u56fe\u50cf\u5230\u5c3a\u5ea6 token maps \u7684\u7f16\u7801\uff1b<\/li>\n\n\n\n<li><strong>VAR transformer \u8bad\u7ec3<\/strong>\uff1a\u57fa\u4e8e cross-entropy \u6700\u5927\u5316\u76ee\u6807\u8bad\u7ec3\u5c3a\u5ea6\u9884\u6d4b\u53d8\u538b\u5668\u3002<\/li>\n<\/ol>\n<\/li>\n\n\n\n<li>\u6ce8\u610f\u529b\u63a9\u7801\u88ab\u8bbe\u8ba1\u4e3a\u786e\u4fdd\u5c3a\u5ea6\u95f4\u81ea\u56de\u5f52\u5173\u7cfb\u6b63\u786e\uff0c\u5373\u9884\u6d4b\u4e00\u4e2a\u5c3a\u5ea6\u65f6\u4ec5\u80fd\u770b\u5230\u66f4\u7c97\u5c3a\u5ea6\u53ca\u81ea\u8eab\u5df2\u751f\u6210\u7684\u4e0a\u4e0b\u6587\u3002<\/li>\n\n\n\n<li>\u6a21\u578b\u53ef\u7528\u6807\u51c6 GPU\/\u5206\u5e03\u5f0f\u67b6\u6784\u8bad\u7ec3\uff1b\u8bba\u6587\u4e5f\u5f00\u6e90\u4e86\u4ee3\u7801\u4e0e\u9884\u8bad\u7ec3\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>6. \u6838\u5fc3\u8d21\u732e\u4e0e\u610f\u4e49<\/strong><\/h1>\n\n\n\n<p>\u8be5\u8bba\u6587\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u63d0\u51fa\u89c6\u89c9\u81ea\u56de\u5f52\u751f\u6210\u8303\u5f0f VAR<\/strong>\uff1a\u901a\u8fc7\u201c\u4e0b\u4e00\u5c3a\u5ea6\u9884\u6d4b\u201d\u66ff\u4ee3\u4f20\u7edf\u5149\u6805\u626b\u63cf\uff0c\u63d0\u9ad8\u4e86 AR \u56fe\u50cf\u751f\u6210\u7684\u6548\u7387\u4e0e\u8d28\u91cf\u3002<\/li>\n\n\n\n<li><strong>\u9996\u6b21\u4f7f AR \u6a21\u578b\u5728\u751f\u6210\u8d28\u91cf\u548c\u6548\u7387\u4e0a\u8d85\u8d8a\u6269\u6563 Transformer<\/strong>\uff1a\u5728\u6807\u51c6 benchmark \u4e0a\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n\n\n\n<li><strong>\u9a8c\u8bc1\u89c6\u89c9 AR \u4e5f\u5177\u6709 LLM \u6837\u7684 scaling laws<\/strong>\uff1a\u4e3a\u7edf\u4e00\u89c6\u89c9\u4e0e\u8bed\u8a00\u751f\u6210\u6a21\u578b\u67b6\u6784\u63d0\u4f9b\u91cd\u8981\u7406\u8bba\u4f9d\u636e\u3002<\/li>\n\n\n\n<li><strong>\u5c55\u793a\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b<\/strong>\uff1a\u589e\u5f3a\u4e86\u89c6\u89c9\u751f\u6210\u6a21\u578b\u7684\u9002\u7528\u6027\u3002<\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"789\" height=\"1024\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-14-789x1024.png\"  class=\"wp-image-865\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-14-789x1024.png 789w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-14-231x300.png 231w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-14-768x997.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-14.png 951w\" sizes=\"auto, (max-width: 789px) 100vw, 789px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe4\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"881\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-13-1024x881.png\"  class=\"wp-image-864\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-13-1024x881.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-13-300x258.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-13-768x661.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-13.png 1051w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe5\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"442\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-12-1024x442.png\"  class=\"wp-image-863\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-12-1024x442.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-12-300x130.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-12-768x332.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/01\/image-12.png 1030w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe6\" alt=\"Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<h1 class=\"wp-block-heading\"><strong>7. \u5c40\u9650\u6027\u4e0e\u672a\u6765\u65b9\u5411<\/strong><\/h1>\n\n\n\n<p>\u867d\u7136 VAR \u5728\u591a\u4e2a\u6307\u6807\u4e2d\u8868\u73b0\u4f18\u8d8a\uff0c\u4f46\u4ecd\u6709\u4e00\u4e9b\u9700\u8981\u8fdb\u4e00\u6b65\u7814\u7a76\u7684\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u66f4\u9ad8\u5206\u8fa8\u7387\u751f\u6210\u7684\u8ba1\u7b97\u5f00\u9500<\/strong>\uff1a\u968f\u7740\u5c3a\u5ea6\u5c42\u6570\u589e\u52a0\uff0c\u6a21\u578b\u4ecd\u9700\u8981\u9ad8\u6548\u7684\u5e76\u884c\u673a\u5236\u548c\u7f13\u5b58\u7b56\u7565\uff08\u5982\u540e\u7eed\u7814\u7a76\u63d0\u51fa\u7684 AMS-KV \/ FastVAR\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4e0e\u6587\u672c\/\u8bed\u8a00\u8054\u5408\u751f\u6210\u7684\u878d\u5408<\/strong>\uff1a\u76ee\u524d\u4e3b\u8981\u7528\u4e8e\u56fe\u50cf\u751f\u6210\uff0c\u672a\u6765\u53ef\u4e0e\u591a\u6a21\u6001\u6a21\u578b\u878d\u5408\u3002<\/li>\n\n\n\n<li><strong>\u5728\u6781\u9ad8\u5206\u8fa8\u7387\u4e0e\u590d\u6742\u7ed3\u6784\u56fe\u50cf\u4e0a\u7684\u6269\u5c55\u80fd\u529b<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<p><strong>\u603b\u7ed3<\/strong>\uff1a\u8fd9\u7bc7\u8bba\u6587\u901a\u8fc7\u4e00\u4e2a\u5168\u65b0\u4e14\u7b80\u5355\u6709\u6548\u7684\u201c\u4ece\u7c97\u5230\u7ec6\u5c3a\u5ea6\u9884\u6d4b\u201d\u67b6\u6784\uff0c\u8ba9\u89c6\u89c9\u81ea\u56de\u5f52\u6a21\u578b\u9996\u6b21\u5728\u751f\u6210\u8d28\u91cf\u3001\u901f\u5ea6\u548c\u6269\u5c55\u6027\u4e0a\u8d85\u8d8a\u6269\u6563\u6a21\u578b\uff0c\u5c55\u793a\u51fa\u7c7b\u4f3c LLM \u7684 scaling laws \u548c\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\uff0c\u5177\u6709\u91cd\u8981\u7684\u7406\u8bba\u4e0e\u5b9e\u8df5\u610f\u4e49<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u7814\u7a76\u52a8\u673a\u4e0e\u80cc\u666f \u4f20\u7edf\u7684\u89c6\u89c9 \u81ea\u56de\u5f52\uff08Autoregressive\uff0cAR\uff09\u751f\u6210\u65b9\u6cd5\u901a\u5e38\u6309\u5149\u6805\u626b\u63cf\uff08raster-scan\uff09\u7684\u987a\u5e8f\u9010\u50cf\u7d20\u6216\u9010 patch \u9884\u6d4b\u4e0b\u4e00\u4e2a token\uff0c\u800c\u8fd9\u79cd\u7b56\u7565\uff1a \u4e0e\u6b64\u540c\u65f6\uff0c\u81ea\u56de\u5f52\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\uff08\u4f8b\u5982 GPT \u7cfb\u5217\uff09\u5c55\u73b0\u51fa\u53ef\u6269\u5c55\u6027\uff08scaling laws\uff09\u548c\u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b\u7b49\u4f18\u52bf\uff0c\u4f46\u8fd9\u4e9b\u7279\u6027\u5c1a\u672a\u5728\u89c6\u89c9\u751f\u6210\u9886\u57df\u5f97\u5230\u5145\u5206\u5229\u7528\u3002 \u672c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u5168\u65b0\u89c6\u89c9\u751f\u6210\u8303\u5f0f Visual AutoRegressive Modeling\uff08VAR\uff09\uff0c\u901a\u8fc7\u4ece\u7c97\u5230\u7ec6 &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=858\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":859,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-858","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/858","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=858"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/858\/revisions"}],"predecessor-version":[{"id":866,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/858\/revisions\/866"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/859"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=858"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=858"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=858"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}