{"id":1747,"date":"2024-05-28T01:35:56","date_gmt":"2024-05-28T01:35:56","guid":{"rendered":"https:\/\/www.nicekj.com\/?p=1747"},"modified":"2024-05-28T01:38:37","modified_gmt":"2024-05-28T01:38:37","slug":"huashijianhuabianshidunbianyikui-openai-sora-xiangguanjishudeyanjin","status":"publish","type":"post","link":"https:\/\/www.nicekj.com\/huashijianhuabianshidunbianyikui-openai-sora-xiangguanjishudeyanjin.html","title":{"rendered":"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb"},"content":{"rendered":"<blockquote>\n<p><strong><strong>\u672c\u6587\u7ecf\u539f\u4f5c\u8005 Ryota Kiuchi, Ph.D. \u6388\u6743\uff0c\u7531Baihai IDP\u7f16\u8bd1\u3002\u5982\u9700\u8f6c\u8f7d\u8bd1\u6587\uff0c\u8bf7\u8054\u7cfb\u83b7\u53d6\u6388\u6743\u3002<\/strong><\/strong><\/p>\n<p><strong>\u539f\u6587\u94fe\u63a5\uff1a<\/strong><\/p>\n<p><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Fhow-openais-sora-is-changing-the-game-an-insight-into-its-core-technologies-bd1ad17170df\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/how-openais-sora-is-changing-the-game-an-insight-into-its-core-technologies-bd1ad17170df\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/how-openais\u2026<\/a><\/p>\n<\/blockquote>\n<blockquote>\n<p><strong>\u7f16\u8005\u6309\uff1a<\/strong> \u8fd1\u671f\uff0cOpenAI \u53d1\u5e03\u901a\u7528\u89c6\u89c9\u5927\u6a21\u578b Sora \uff0c\u8fd9\u4e5f\u662f\u7ee7\u6587\u672c\u6a21\u578bChatGPT\u548c\u56fe\u7247\u6a21\u578bDall-E\u4e4b\u540e\uff0c\u53c8\u4e00\u6781\u5177\u98a0\u8986\u6027\u7684\u5927\u6a21\u578b\u4ea7\u54c1\uff0c\u4eba\u4eec\u91cd\u65b0\u601d\u8003\u4e86\u751f\u6210\u5f0f AI \u5728\u89c6\u89c9\u5185\u5bb9\u521b\u4f5c\u9886\u57df\u7684\u5e94\u7528\u524d\u666f\uff0c\u5185\u5bb9\u521b\u4f5c\u5de5\u4f5c\u6d41\u6709\u671b\u88ab\u98a0\u8986\u3002<\/p>\n<p>\u6211\u4eec\u4eca\u5929\u8981\u4e3a\u5927\u5bb6\u5206\u4eab\u7684\u8fd9\u7bc7\u535a\u6587\uff0c\u4f5c\u8005\u8ba4\u4e3a Sora \u4ee3\u8868\u4e86Transformer\u3001NaViT\u3001\u6269\u6563\u6a21\u578b\u7b49\u4e00\u7cfb\u5217\u89c6\u89c9AI\u6280\u672f\u7684\u878d\u5408\u521b\u65b0\uff0c\u662f\u8fc8\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7684\u91cd\u8981\u4e00\u6b65\u3002<\/p>\n<p>\u4f5c\u8005\u9996\u5148\u7b80\u8981\u4ecb\u7ecd\u4e86Sora\u7684\u529f\u80fd\uff0c\u7136\u540e\u8be6\u7ec6\u68b3\u7406\u4e86\u652f\u6301Sora\u7684\u5404\u9879\u6838\u5fc3\u6280\u672f\u5185\u5bb9\uff0c\u5305\u62ecTransformer\u3001ViT\u3001ViVit\u3001MAE\u3001NaViT\u3001\u6269\u6563\u6a21\u578b\u3001Latent Diffusion Models\u4ee5\u53ca\u6700\u5173\u952e\u7684Diffusion Transformer\u3002\u6700\u540e\uff0c\u4f5c\u8005\u9884\u6d4bSora\u672a\u6765\u5c06\u8fdb\u4e00\u6b65\u62d3\u5c55\u5e94\u7528\u8303\u56f4\uff0c\u8fdb\u519b\u4e09\u7ef4\u5efa\u6a21\u9886\u57df\uff0c\u5e76\u6700\u7ec8\u6210\u4e3a\u7c7b\u4f3c\u7269\u7406\u5f15\u64ce\u7684\u901a\u7528\u5206\u6790\u5de5\u5177\uff0c\u4e3a\u89c6\u89c9\u5185\u5bb9\u521b\u4f5c\u751a\u81f3\u5176\u4ed6\u5404\u4e2a\u9886\u57df\u5e26\u6765\u9769\u547d\u6027\u8fdb\u6b65\u3002Sora\u7684\u8bde\u751f\u9884\u793a\u7740\u591a\u6a21\u6001AI\u5c06\u9010\u6b65\u8d70\u5411\u6210\u719f\u4e0e\u666e\u53ca\uff0c\u4eba\u7c7b\u60f3\u8c61\u529b\u7684\u8fb9\u754c\u5c06\u5f97\u5230\u8fdb\u4e00\u6b65\u62d3\u5c55\u3002<\/p>\n<\/blockquote>\n<p><strong>\u4f5c\u8005 | Ryota Kiuchi, Ph.D.<\/strong><\/p>\n<p><strong>\u7f16\u8bd1&nbsp;|&nbsp;\u5cb3\u626c<\/strong><\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/219b5dffdc8058d9017623c2e66a9dcc.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Photo by Kaushik Panchal on Unsplash<\/p>\n<p>2024 \u5e74 2 \u6708 15 \u65e5\uff0c\u66fe\u5728 2022 \u5e74\u5e95\u53d1\u5e03 ChatGPT \u60ca\u8273\u4e16\u754c\u7684 OpenAI\uff0c\u518d\u6b21\u51ed\u501f Sora \u7684\u4eae\u76f8\u9707\u60ca\u4e16\u754c\u3002\u4e0d\u53ef\u5426\u8ba4\uff0c\u8fd9\u9879\u80fd\u591f\u6839\u636e\u6587\u5b57\u63d0\u793a\u8bcd\uff08text prompt\uff09\u5236\u4f5c\u957f\u8fbe\u4e00\u5206\u949f\u89c6\u9891\u7684\u6280\u672f\u5fc5\u5c06\u662f\u8fc8\u5411 AGI \u7684\u53c8\u4e00\u5ea7\u91cc\u7a0b\u7891\u3002<\/p>\n<p>\u5728\u8fd9\u7bc7\u535a\u6587\u4e2d\uff0c\u6211\u5c06\u6839\u636e OpenAI \u53d1\u5e03\u7684\u6280\u672f\u62a5\u544a\uff0c\u4ecb\u7ecd\u8fd9\u9879\u60ca\u4eba\u6280\u672f\u80cc\u540e\u7684\u57fa\u672c\u7814\u7a76\u65b9\u6cd5\u548c\u7814\u7a76\u5185\u5bb9\u3002<\/p>\n<p>\u987a\u4fbf\u63d0\u4e00\u4e0b\uff0c\u201cSora\u201d\u5728\u65e5\u8bed\u4e2d\u662f\u201c\u5929\u7a7a\u201d\u7684\u610f\u601d\u3002\u867d\u7136\u5b98\u65b9\u5c1a\u672a\u516c\u5e03\u8fd9\u4e00\u547d\u540d\u662f\u4f55\u7528\u610f\uff0c\u4f46\u9274\u4e8e OpenAI \u53d1\u5e03\u7684\u63a8\u6587\u4e2d\u6709\u4e00\u6bb5\u4ee5\u4e1c\u4eac\u4e3a\u4e3b\u9898\u7684\u89c6\u9891\uff0c\u56e0\u6b64\u53ef\u4ee5\u63a8\u6d4b\u8fd9\u4e2a\u731c\u6d4b\u662f\u6bd4\u8f83\u5408\u7406\u7684\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/ebba7be9f94618c8da3f75a22ef5f234.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">OpenAI&nbsp;\u901a\u8fc7&nbsp;X&nbsp;\u5411\u5168\u4e16\u754c\u5c55\u793aSora<\/p>\n<h1 data-id=\"heading-0\"><strong>\u76ee\u5f55<\/strong><\/h1>\n<p>01 Sora&nbsp;\u7684\u7b80\u5355\u4ecb\u7ecd<\/p>\n<p>02 \u5b83\u80cc\u540e\u7684\u76f8\u5173\u6280\u672f\u548c\u76f8\u5173\u7814\u7a76\u6709\u54ea\u4e9b\uff1f<\/p>\n<p>03&nbsp;\u8fd9\u4e9b\u7814\u7a76\u57fa\u7840\u52a0\u4e0aOpenAI\u7684\u52aa\u529b\u5171\u540c\u9020\u5c31\u4e86&nbsp;Sora<\/p>\n<p>04 \u5c55\u671b Sora&nbsp;\u7684\u672a\u6765<\/p>\n<h1 data-id=\"heading-1\"><strong>01 Sora&nbsp;\u7684\u7b80\u5355\u4ecb\u7ecd<\/strong><\/h1>\n<p>Sora \u662f\u7531 OpenAI \u5f00\u53d1\u7684\u4e00\u6b3e text-to-video \uff08\u6587\u751f\u89c6\u9891\uff09\u8f6c\u6362\u6a21\u578b\uff0c\u5176\u80fd\u529b\u548c\u5e94\u7528\u8303\u56f4\u6307\u5f15\u4e86\u73b0\u4ee3AI\u6280\u672f\u7684\u65b0\u53d1\u5c55\u65b9\u5411\u3002\u8be5\u6a21\u578b\u4e0d\u4ec5\u9650\u4e8e\u80fd\u591f\u751f\u6210\u51e0\u79d2\u949f\u7684\u89c6\u9891\uff0c\u751a\u81f3\u53ef\u4ee5\u521b\u5efa\u957f\u8fbe\u4e00\u5206\u949f\u7684\u89c6\u9891\uff0c\u5728\u4fdd\u6301\u9ad8\u8d28\u91cf\u7684\u540c\u65f6\u5fe0\u5b9e\u5730\u6ee1\u8db3\u7528\u6237\u7684\u6307\u4ee4\u3002\u5b83\u4eff\u4f5b\u80fd\u591f\u5c06\u5927\u5bb6\u7684\u68a6\u60f3\u53d8\u6210\u73b0\u5b9e\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/1cb53b63f6be9e52c6e92cfcb8e364a7.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">OpenAI Sora&nbsp;\u751f\u6210\u7684\u5185\u5bb9\u6f14\u793a<\/p>\n<h2 data-id=\"heading-2\"><strong>\u6839\u636e\u73b0\u5b9e\u4e16\u754c\u751f\u6210\u590d\u6742\u573a\u666f<\/strong><\/h2>\n<p>Sora \u53ef\u4ee5\u7406\u89e3 Prompt \u4e2d\u63cf\u8ff0\u7684\u5143\u7d20\u5728\u7269\u7406\u4e16\u754c\u4e2d\u5b58\u5728\u5f62\u5f0f\u548c\u8fd0\u4f5c\u65b9\u5f0f\uff08exist and operate\uff09\u3002\u8fd9\u4f7f\u5f97\u8be5\u6a21\u578b\u80fd\u591f\u51c6\u786e\u5730\u8868\u73b0\u7528\u6237\u671f\u671b\u5728\u89c6\u9891\u4e2d\u51fa\u73b0\u7684\u52a8\u4f5c\u548c\u884c\u4e3a\u3002\u4f8b\u5982\uff0c\u5b83\u53ef\u4ee5\u903c\u771f\u5730\u518d\u73b0\u4eba\u5954\u8dd1\u7684\u666f\u8c61\u6216\u81ea\u7136\u73b0\u8c61\u7684\u53d8\u5316\u3002\u6b64\u5916\uff0c\u5b83\u8fd8\u80fd\u7cbe\u786e\u518d\u73b0\u591a\u4e2a\u89d2\u8272\u7684\u7ec6\u8282\u3001\u52a8\u4f5c\u7c7b\u578b\u4ee5\u53ca\u4e3b\u4f53\u548c\u80cc\u666f\u7684\u5177\u4f53\u7ec6\u8282\u3002<\/p>\n<p>\u4ee5\u5f80\uff0c\u4f7f\u7528\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u8fdb\u884c\u89c6\u9891\u521b\u4f5c\u9762\u4e34\u7740\u4e00\u4e2a\u8270\u5de8\u6311\u6218\uff0c\u5373\u5982\u4f55\u5728\u4e0d\u540c\u573a\u666f\u4e4b\u95f4\u4fdd\u6301\u4e00\u81f4\u6027\u548c\u53ef\u518d\u73b0\u6027\u3002\u8fd9\u662f\u56e0\u4e3a\uff0c<strong>\u5728\u5355\u72ec\u751f\u6210\u6bcf\u4e2a\u573a\u666f\u6216\u6bcf\u4e00\u5e27\u65f6\uff0c\u8981\u5b8c\u5168\u7406\u89e3\u4e4b\u524d\u7684\u4e0a\u4e0b\u6587\u548c\u7ec6\u8282\uff0c\u5e76\u5c06\u5176\u9002\u5f53\u5730\u7ee7\u627f\u5230\u4e0b\u4e00\u4e2a\u573a\u666f\u4e2d\u662f\u4e00\u9879\u6781\u5176\u8270\u5de8\u7684\u6311\u6218\u3002<\/strong> \u7136\u800c\uff0c\u8be5\u6a21\u578b\u901a\u8fc7\u5c06&nbsp; <strong>\u201c\u5bf9\u5e26\u6709\u89c6\u89c9\u4e0a\u4e0b\u6587\u7684\u8bed\u8a00\u7684\u6df1\u523b\u7406\u89e3\u201d<\/strong> &nbsp;\u548c&nbsp; <strong>\u201c\u5bf9&nbsp;prompt&nbsp;\u7684\u51c6\u786e\u89e3\u8bfb\u201d<\/strong> &nbsp;\u76f8\u7ed3\u5408\uff0c\u4fdd\u8bc1\u4e86\u53d9\u4e8b\u7684\u4e00\u81f4\u6027\u3002\u5b83\u8fd8\u80fd\u4ece\u7ed9\u5b9a\u7684 prompt \u4e2d\u6355\u6349\u4eba\u7269\u7684\u60c5\u7eea\u548c\u4e2a\u6027\u7279\u5f81\uff0c\u5e76\u5c06\u5176\u63cf\u7ed8\u6210\u89c6\u9891\u4e2d\u5bcc\u6709\u8868\u73b0\u529b\u7684\u89d2\u8272\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/bb07f65a71a2a771c0a972ba81dae32b.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">The post by Bill Peebles&nbsp;(OpenAI)&nbsp;via X<\/p>\n<h1 data-id=\"heading-3\"><strong>02 \u5b83\u80cc\u540e\u7684\u76f8\u5173\u6280\u672f\u548c\u76f8\u5173\u7814\u7a76\u6709\u54ea\u4e9b\uff1f<\/strong><\/h1>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/9ec82fcc727c90eda68537157c86501f.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Photo by Markus Spiske on Unsplash<\/p>\n<p>Sora \u7684\u7814\u7a76\u5efa\u7acb\u5728\u5148\u524d\u56fe\u50cf\u6570\u636e\u751f\u6210\u6a21\u578b\u7814\u7a76\u7684\u57fa\u7840\u4e0a\u3002\u4e4b\u524d\u7684\u7814\u7a76\u91c7\u7528\u4e86\u591a\u79cd\u65b9\u6cd5\uff0c\u5982\u9012\u5f52\u7f51\u7edc\uff08recurrent networks\uff09\u3001\u751f\u6210\u5bf9\u6297\u7f51\u7edc\uff08GANs\uff09\u3001\u81ea\u56de\u5f52Transformers\u548c\u6269\u6563\u6a21\u578b\uff0c\u4f46\u901a\u5e38\u4e13\u6ce8\u4e8e\u67d0\u4e9b\u5355\u4e00\u7c7b\u522b\u7684\u89c6\u89c9\u6570\u636e\u3001\u8f83\u77ed\u7684\u89c6\u9891\u6216\u56fa\u5b9a\u5206\u8fa8\u7387\u7684\u89c6\u9891\u3002Sora \u8d85\u8d8a\u4e86\u8fd9\u4e9b\u9650\u5236\uff0c\u5e76\u4e14\u5728\u751f\u6210\u89c6\u9891\u7684\u6301\u7eed\u65f6\u95f4\u3001\u957f\u5bbd\u6bd4\u548c\u5c3a\u5bf8\u4e0a\u5f97\u5230\u4e86\u663e\u8457\u6539\u8fdb\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u5c06\u4ecb\u7ecd\u652f\u6301\u8fd9\u4e9b\u6539\u8fdb\u7684\u6838\u5fc3\u6280\u672f\u3002<\/p>\n<h2 data-id=\"heading-4\"><strong>2.1 Transformer<\/strong><\/h2>\n<blockquote>\n<p>Vaswani et al.&nbsp;(2017),&nbsp;\u201cAttention is all you need.\u201d<\/p>\n<\/blockquote>\n<p>Transformer\u662f\u4e00\u79cd\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff0c\u5b83\u5f7b\u5e95\u6539\u53d8\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u3002\u5b83\u7531 Vaswani \u7b49\u4eba\u4e8e 2017 \u5e74\u9996\u6b21\u63d0\u51fa\u3002<strong>\u8be5\u6a21\u578b\u6781\u5927\u5730\u514b\u670d\u4e86\u4f20\u7edf\u9012\u5f52\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u548c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u5b58\u5728\u7684\u77ed\u677f\uff0c\u4f5c\u4e3a\u4e00\u79cd\u521b\u65b0\u65b9\u6cd5\u652f\u6301\u7740\u5f53\u4eca\u7684\u5404\u79cd\u7a81\u7834\u6027\u6280\u672f\u3002<\/strong><\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/136b073b5891b474786ae3067e9b87c8.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Transformer&nbsp;\u6a21\u578b\u67b6\u6784\uff5cVaswani et al.&nbsp;(2017)<\/p>\n<p>RNN \u5b58\u5728\u7684\u95ee\u9898\uff1a<\/p>\n<ul>\n<li><strong>\u957f\u671f\u4f9d\u8d56\uff08long-term dependencies\uff09\u95ee\u9898<\/strong>\uff1a\u5c3d\u7ba1 RNN \u5728\u7406\u8bba\u4e0a\u53ef\u4ee5\u901a\u8fc7\u65f6\u95f4\u4f20\u9012\u4fe1\u606f\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u5f80\u5f80\u96be\u4ee5\u6355\u6349\u957f\u65f6\u95f4\u8de8\u5ea6\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/li>\n<li><strong>\u5e76\u884c\u5904\u7406\u5b58\u5728\u9650\u5236<\/strong>\uff1a\u7531\u4e8e RNN \u7684\u6bcf\u4e00\u6b65\u8ba1\u7b97\u90fd\u4f9d\u8d56\u4e8e\u524d\u4e00\u6b65\u7684\u8f93\u51fa\uff0c\u56e0\u6b64\u5fc5\u987b\u8fdb\u884c\u987a\u5e8f\u5904\u7406\uff08\u4f8b\u5982\uff0c\u6309\u987a\u5e8f\u9010\u4e2a\u5904\u7406\u6587\u672c\u4e2d\u7684\u5355\u8bcd\u6216\u53e5\u5b50\uff09\uff0c\u4ece\u800c\u65e0\u6cd5\u5229\u7528\u73b0\u4ee3\u8ba1\u7b97\u673a\u4f53\u7cfb\u7ed3\u6784\u63d0\u4f9b\u7684\u5e76\u884c\u5904\u7406\u4f18\u52bf\u3002\u8fd9\u5bfc\u81f4\u5728\u5927\u578b\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u6548\u7387\u4f4e\u4e0b\u3002<\/li>\n<\/ul>\n<p>CNN \u5b58\u5728\u7684\u95ee\u9898\uff1a<\/p>\n<ul>\n<li><strong>\u56fa\u5b9a\u7684\u611f\u53d7\u91ce\u5927\u5c0f\uff08receptive field size\uff09<\/strong> \uff1a\u867d\u7136 CNN \u64c5\u957f\u63d0\u53d6\u5c40\u90e8\u7279\u5f81\uff0c\u4f46\u5176\u56fa\u5b9a\u7684\u611f\u53d7\u91ce\u5927\u5c0f\u9650\u5236\u4e86\u5176\u5728\u6574\u4e2a\u4e0a\u4e0b\u6587\u4e2d\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb\uff08long-distance dependencies\uff09\u7684\u80fd\u529b\u3002<\/li>\n<li><strong>\u96be\u4ee5\u6a21\u62df\u81ea\u7136\u8bed\u8a00\u7684\u5c42\u6b21\u7ed3\u6784<\/strong>\uff1a\u4f7f\u7528CNN\u76f4\u63a5\u4e3a\u8bed\u8a00\u7684\u5c42\u6b21\u7ed3\u6784\u5efa\u6a21\u6781\u5177\u6311\u6218\u6027\uff0c\u53ef\u80fd\u4e0d\u8db3\u4ee5\u5b9e\u73b0\u6df1\u5c42\u6b21\u7684\u4e0a\u4e0b\u6587\u7406\u89e3\u3002<\/li>\n<\/ul>\n<p>Transformer \u7684\u65b0\u7279\u6027\uff1a<\/p>\n<ul>\n<li><strong>\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff1a\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u76f4\u63a5\u5efa\u6a21\u5e8f\u5217\u4e2d\u4efb\u610f\u4f4d\u7f6e\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u4ece\u800c\u76f4\u63a5\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u548c\u5e7f\u6cdb\u7684\u4e0a\u4e0b\u6587\u3002<\/li>\n<li><strong>\u80fd\u591f\u652f\u6301\u5e76\u884c\u5904\u7406<\/strong>\uff1a\u7531\u4e8e\u8f93\u5165\u6570\u636e\u662f\u4f5c\u4e3a\u4e00\u4e2a\u6574\u4f53\u4e00\u6b21\u6027\u5904\u7406\u7684\uff0c\u56e0\u6b64\u5b9e\u73b0\u4e86\u8ba1\u7b97\u7684\u9ad8\u5ea6\u5e76\u884c\u5316\uff0c\u5927\u5927\u52a0\u5feb\u4e86\u5728\u5927\u578b\u6570\u636e\u96c6\u4e0a\u7684\u8bad\u7ec3\u901f\u5ea6\u3002<\/li>\n<li><strong>\u53ef\u53d8\u7684\u611f\u53d7\u91ce\uff08receptive field\uff09<\/strong> \uff1a\u6ce8\u610f\u529b\u673a\u5236\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u6839\u636e\u9700\u8981\u52a8\u6001\u8c03\u6574\u201c\u611f\u53d7\u91ce\u201d\u7684\u5927\u5c0f\u3002\u8fd9\u610f\u5473\u7740\u6a21\u578b\u5728\u5904\u7406\u67d0\u4e9b\u4efb\u52a1\u6216\u6570\u636e\u65f6\uff0c\u53ef\u4ee5\u81ea\u7136\u5730\u5c06\u6ce8\u610f\u529b\u96c6\u4e2d\u5728\u5c40\u90e8\u4fe1\u606f\u4e0a\uff0c\u800c\u5728\u5176\u4ed6\u60c5\u51b5\u4e0b\uff0c\u5219\u53ef\u4ee5\u8003\u8651\u66f4\u5e7f\u6cdb\u7684\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n<p><em>\u6709\u5173 Transformer \u66f4\u8be6\u7ec6\u7684\u6280\u672f\u89e3\u91ca\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Ftransformers-141e32e69591\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/transformers-141e32e69591\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/transformer\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-5\"><strong>2.2 Vision Transformer&nbsp;(ViT)<\/strong><\/h2>\n<blockquote>\n<p>Dosovitskiy,&nbsp;et al.&nbsp;(2020),&nbsp;\u201cAn image is worth 16&#215;16 words:&nbsp;Transformers for image recognition at scale.\u201d<\/p>\n<\/blockquote>\n<p>\u5728\u8fd9\u9879\u7814\u7a76\u4e2d\uff0c\u98a0\u8986\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u7684 Transformer \u539f\u7406\u88ab\u5e94\u7528\u4e8e\u56fe\u50cf\u8bc6\u522b\u4e2d\uff0c\u4e3a\u89c6\u89c9\u6a21\u578b\u5f00\u8f9f\u4e86\u65b0\u7684\u65b9\u5411\u3002<\/p>\n<p><strong>Token&nbsp;\u548c&nbsp;Patch<\/strong><\/p>\n<p>\u5728\u539f\u59cb\u7684 Transformer \u8bba\u6587\u4e2d\uff0ctoken \u4e3b\u8981\u4ee3\u8868\u5355\u8bcd\u6216\u53e5\u5b50\u7684\u4e00\u90e8\u5206\uff0c\u5206\u6790\u8fd9\u4e9b token \u4e4b\u95f4\u7684\u5173\u7cfb\u53ef\u4ee5\u6df1\u5165\u7406\u89e3\u53e5\u5b50\u7684\u542b\u4e49\u3002\u5728\u8fd9\u9879\u7814\u7a76\u4e2d\uff0c\u4e3a\u4e86\u5c06 token \u7684\u6982\u5ff5\u5e94\u7528\u5230\u89c6\u89c9\u6570\u636e\u4e2d\uff0c\u56fe\u50cf\u88ab\u5212\u5206\u6210\u4e86 16&#215;16 \u7684\u5c0f\u5757\uff08patch\uff09\uff0c\u5e76\u4e14\u6bcf\u4e2a patch \u90fd\u88ab\u89c6\u4e3a Transformer \u4e2d\u7684\u4e00\u4e2a\u201ctoken\u201d\u3002\u8fd9\u79cd\u65b9\u6cd5\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u6bcf\u4e2a patch \u5728\u6574\u4e2a\u56fe\u50cf\u4e2d\u7684\u5173\u7cfb\uff0c\u4ece\u800c\u80fd\u591f\u57fa\u4e8e\u6b64\u8bc6\u522b\u548c\u7406\u89e3\u6574\u4e2a\u56fe\u50cf\u3002<strong>\u5b83\u8d85\u8d8a\u4e86\u4f20\u7edf CNN \u6a21\u578b\u5728\u56fe\u50cf\u8bc6\u522b\u4e2d\u4f7f\u7528\u7684\u56fa\u5b9a\u611f\u53d7\u91ce\u5927\u5c0f\u7684\u9650\u5236\uff0c\u80fd\u591f\u7075\u6d3b\u6355\u6349\u56fe\u50cf\u4e2d\u7684\u4efb\u4f55\u4f4d\u7f6e\u5173\u7cfb\u3002<\/strong><\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/c9ef7664d9272d1638b7e54a594b3f4e.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">ViT&nbsp;\u6a21\u578b\u6982\u89c8\uff5cDosovitskiy,&nbsp;et al.&nbsp;(2020)<\/p>\n<p><em>\u6709\u5173 Vision Transformer (ViT)&nbsp;\u66f4\u8be6\u7ec6\u7684\u6280\u672f\u89e3\u91ca\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Fmachinelearningmastery.com%2Fthe-vision-transformer-model%2F\" target=\"_blank\" title=\"https:\/\/machinelearningmastery.com\/the-vision-transformer-model\/\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">machinelearningmastery.com\/the-vision-\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-6\"><strong>2.3 Video Vision Transformer&nbsp;(ViViT)<\/strong><\/h2>\n<blockquote>\n<p>Arnab,&nbsp;et al.&nbsp;(2021),&nbsp;\u201cVivit:&nbsp;A video vision transformer.\u201d<\/p>\n<\/blockquote>\n<p>ViViT \u8fdb\u4e00\u6b65\u6269\u5c55\u4e86 Vision Transformer \u7684\u6982\u5ff5\uff0c\u5c06\u5176\u5e94\u7528\u5230\u89c6\u9891\u7684\u591a\u7ef4\u6570\u636e\u4e0a\u3002\u89c6\u9891\u6570\u636e\u66f4\u52a0\u590d\u6742\uff0c\u56e0\u4e3a\u5b83\u65e2\u5305\u542b\u9759\u6001\u56fe\u50cf\u4fe1\u606f\uff08\u7a7a\u95f4\u5143\u7d20\uff09\uff0c\u53c8\u5305\u542b\u968f\u65f6\u95f4\u53d8\u5316\u7684\u52a8\u6001\u4fe1\u606f\uff08\u65f6\u95f4\u5143\u7d20\uff09\u3002<strong>ViViT \u5c06\u89c6\u9891\u5206\u89e3\u4e3a patch \uff0c\u5e76\u5c06\u5176\u89c6\u4e3a Transformer \u6a21\u578b\u4e2d\u7684 token\u3002<\/strong> \u5f15\u5165 patch &nbsp;\u540e\uff0cViViT \u80fd\u591f\u540c\u65f6\u6355\u6349\u89c6\u9891\u4e2d\u7684\u9759\u6001\u548c\u52a8\u6001\u5143\u7d20\uff0c\u5e76\u5bf9\u5b83\u4eec\u4e4b\u95f4\u7684\u590d\u6742\u5173\u7cfb\u8fdb\u884c\u5efa\u6a21\u3002&nbsp;<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/7c825ef52ec95302691bffa317f432b2.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Tubelet&nbsp;(\u65f6\u7a7a\u8f93\u5165\u91cf)&nbsp;\u5d4c\u5165\u56fe\u50cf&nbsp;\uff5cArnab,&nbsp;et al.&nbsp;(2021)<\/p>\n<p><em>\u6709\u5173 ViViT \u7684\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Fmedium.com%2Faiguys%2Fvivit-video-vision-transformer-648a5fff68a4\" target=\"_blank\" title=\"https:\/\/medium.com\/aiguys\/vivit-video-vision-transformer-648a5fff68a4\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">medium.com\/aiguys\/vivi\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-7\"><strong>2.4 Masked Autoencoders&nbsp;(MAE)&nbsp;\u5e26\u6709\u63a9\u7801\u7684\u81ea\u7f16\u7801\u5668<\/strong><\/h2>\n<blockquote>\n<p>He,&nbsp;et al.&nbsp;(2022),&nbsp;\u201cMasked autoencoders are scalable vision learners.\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u9879\u7814\u7a76\u901a\u8fc7\u4f7f\u7528\u4e00\u79cd\u88ab\u79f0\u4e3a\u5e26\u6709\u63a9\u7801\u7684\u81ea\u7f16\u7801\u5668\uff08Masked Autoencoder\uff09\u7684\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u65b9\u6cd5\uff0c\u663e\u8457\u6539\u5584\u4e86\u4f20\u7edf\u4e0a<strong>\u4e0e\u9ad8\u7ef4\u5ea6\u548c\u6d77\u91cf\u4fe1\u606f\u76f8\u5173\u7684\u5927\u578b\u6570\u636e\u96c6\u8bad\u7ec3\u4e2d<\/strong>\u5b58\u5728\u7684<strong>\u8ba1\u7b97\u6210\u672c\u9ad8\u6602\u548c\u4f4e\u6548\u7387\u95ee\u9898<\/strong>\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c<strong>\u901a\u8fc7\u5bf9\u8f93\u5165\u56fe\u50cf\u7684\u90e8\u5206\u5185\u5bb9\u8fdb\u884c\u63a9\u7801\u5904\u7406\uff0c\u7f51\u7edc\u88ab\u8bad\u7ec3\u6765\u9884\u6d4b\u9690\u85cf\u90e8\u5206\u7684\u4fe1\u606f\uff0c\u4ece\u800c\u66f4\u6709\u6548\u5730\u5b66\u4e60\u56fe\u50cf\u4e2d\u7684\u91cd\u8981\u7279\u5f81\u548c\u7ed3\u6784\uff0c\u5e76\u83b7\u5f97\u4e30\u5bcc\u7684\u89c6\u89c9\u6570\u636e\u8868\u5f81\u3002<\/strong> \u8fd9\u4e2a\u8fc7\u7a0b\u4f7f\u5f97\u6570\u636e\u7684\u538b\u7f29\uff08compression \uff09\u548c\u8868\u5f81\u5b66\u4e60\uff08representation learning\uff09\u66f4\u52a0\u9ad8\u6548\uff0c\u964d\u4f4e\u4e86\u8ba1\u7b97\u6210\u672c\uff0c\u5e76\u589e\u5f3a\u4e86\u4e0d\u540c\u7c7b\u578b\u7684\u89c6\u89c9\u6570\u636e\u4ee5\u53ca\u89c6\u89c9\u4efb\u52a1\u7684\u591a\u6837\u6027\u3002<\/p>\n<p>\u8fd9\u9879\u7814\u7a76\u7684\u65b9\u6cd5\u8fd8\u4e0e BERT\uff08Bidirectional Encoder Representations from Transformers\uff09\u7b49\u8bed\u8a00\u6a21\u578b\u7684\u6f14\u53d8\u5bc6\u5207\u76f8\u5173\u3002\u867d\u7136 BERT \u901a\u8fc7 Masked Language Modeling\uff08MLM\uff09\u5b9e\u73b0\u4e86\u5bf9\u6587\u672c\u6570\u636e\u7684\u6df1\u5ea6\u4e0a\u4e0b\u6587\u7406\u89e3\uff0c\u4f46 He \u7b49\u4eba\u5219\u5c06\u7c7b\u4f3c\u7684\u63a9\u7801\u6280\u672f\u5e94\u7528\u4e8e\u89c6\u89c9\u6570\u636e\uff0c\u5b9e\u73b0\u4e86\u5bf9\u56fe\u50cf\u7684\u66f4\u6df1\u5c42\u6b21\u7406\u89e3\u548c\u8868\u793a\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/42c211debafc9b226be0ab179bf9b2ae.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Masked Autoencoders\uff5cHe,&nbsp;et al.&nbsp;(2022)<\/p>\n<p><em>\u6709\u5173 MAE \u7684\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Fpaper-explained-masked-autoencoders-are-scalable-vision-learners-9dea5c5c91f0\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/paper-explained-masked-autoencoders-are-scalable-vision-learners-9dea5c5c91f0\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/paper-expla\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-8\"><strong>2.5 Native Resolution Vision Transformer&nbsp;(NaViT)<\/strong><\/h2>\n<blockquote>\n<p>Dehghani,&nbsp;et al.&nbsp;(2023),&nbsp;\u201cPatch n\u2019Pack:&nbsp;NaViT,&nbsp;a Vision Transformer for any Aspect Ratio and Resolution.\u201d<\/p>\n<\/blockquote>\n<p>\u672c\u7814\u7a76\u63d0\u51fa\u4e86Native Resolution &nbsp;ViTransformer\uff08NaViT\uff09\uff0c\u8be5\u6a21\u578b\u65e8\u5728\u8fdb\u4e00\u6b65\u6269\u5c55 Vision Transformer\uff08ViT\uff09\u7684\u9002\u7528\u6027\uff0c\u4f7f\u5176\u9002\u7528\u4e8e\u4efb\u4f55\u957f\u5bbd\u6bd4\u6216\u5206\u8fa8\u7387\u7684\u56fe\u50cf\u3002<\/p>\n<p><strong>\u4f20\u7edf&nbsp;ViT&nbsp;\u9762\u4e34\u7684\u6311\u6218<\/strong><\/p>\n<p>Vision Transformer \u5f15\u5165\u4e86\u4e00\u79cd\u5f00\u521b\u6027\u7684\u65b9\u6cd5\uff0c\u901a\u8fc7\u5c06\u56fe\u50cf\u5212\u5206\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684 patches \uff0c\u5e76\u5c06\u8fd9\u4e9b patches \u89c6\u4e3a tokens \uff0c\u5c06 transformer \u6a21\u578b\u5e94\u7528\u4e8e\u56fe\u50cf\u8bc6\u522b\u4efb\u52a1\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u5047\u8bbe\u6a21\u578b\u9488\u5bf9\u7279\u5b9a\u5206\u8fa8\u7387\u6216\u957f\u5bbd\u6bd4\u8fdb\u884c\u4e86\u9488\u5bf9\u6027\u7684\u4f18\u5316\uff0c\u56e0\u6b64\u5bf9\u4e8e\u4e0d\u540c\u5c3a\u5bf8\u6216\u5f62\u72b6\u7684\u56fe\u50cf\uff0c\u9700\u8981\u5bf9\u6a21\u578b\u8fdb\u884c\u8c03\u6574\u3002\u8fd9\u662f\u4e00\u4e2a\u6bd4\u8f83\u5927\u7684\u9650\u5236\uff0c\u56e0\u4e3a\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684\u5e94\u7528\u901a\u5e38\u9700\u8981\u5904\u7406\u5404\u79cd\u5c3a\u5bf8\u548c\u957f\u5bbd\u6bd4\u7684\u56fe\u50cf\u3002<\/p>\n<p><strong>NaViT&nbsp;\u7684\u521b\u65b0<\/strong><\/p>\n<p>NaViT \u53ef\u9ad8\u6548\u5904\u7406\u4efb\u4f55\u957f\u5bbd\u6bd4\u6216\u5206\u8fa8\u7387\u7684\u56fe\u50cf\uff0c\u5141\u8bb8\u5b83\u4eec\u76f4\u63a5\u8f93\u5165\u6a21\u578b\u800c\u65e0\u9700\u4e8b\u5148\u8c03\u6574\u3002<strong>Sora \u4e5f\u5c06\u8fd9\u79cd\u7075\u6d3b\u6027\u5e94\u7528\u4e8e\u89c6\u9891\u573a\u666f\uff0c\u901a\u8fc7\u65e0\u7f1d\u5904\u7406\u5404\u79cd\u5c3a\u5bf8\u548c\u5f62\u72b6\u7684\u89c6\u9891\u548c\u56fe\u50cf\uff0c\u5927\u5927\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u7075\u6d3b\u6027\u548c\u9002\u5e94\u6027\u3002<\/strong><\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/aab351822cb7193fd3e6a04bb16cd7b6.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Dehghani,&nbsp;et al.&nbsp;(2023)<\/p>\n<h2 data-id=\"heading-9\"><strong>2.6 Diffusion Models<\/strong><\/h2>\n<blockquote>\n<p>Sohl-Dickstein,&nbsp;et al.&nbsp;(2015),&nbsp;\u201cDeep unsupervised learning using nonequilibrium thermodynamics.\u201d<\/p>\n<\/blockquote>\n<p>\u9664\u4e86 Transformer\uff0c\u6269\u6563\u6a21\u578b\u4e5f\u662f\u652f\u6301 Sora \u7684\u9aa8\u5e72\u6280\u672f\u3002\u8fd9\u9879\u7814\u7a76\u4e3a\u6269\u6563\u6a21\u578b\u5960\u5b9a\u4e86\u7406\u8bba\u57fa\u7840\uff0c\u6269\u6563\u6a21\u578b\u662f\u4e00\u79cd\u5229\u7528\u975e\u5e73\u8861\u70ed\u529b\u5b66\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u3002\u6269\u6563\u6a21\u578b\u5f15\u5165\u4e86\u6269\u6563\u8fc7\u7a0b\u7684\u6982\u5ff5\uff0c\u8be5\u8fc7\u7a0b\u4ece\u968f\u673a\u566a\u58f0\uff08\u6ca1\u6709\u4efb\u4f55\u6a21\u5f0f\uff08pattern\uff09\u7684\u6570\u636e\uff09\u5f00\u59cb\uff0c\u9010\u6e10\u53bb\u9664\u566a\u58f0\uff0c\u4ece\u800c\u521b\u5efa\u4e0e\u5b9e\u9645\u56fe\u50cf\u6216\u89c6\u9891\u76f8\u4f3c\u7684\u6570\u636e\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u60f3\u8c61\u4e00\u4e0b\uff0c\u4e00\u5f00\u59cb\u53ea\u6709\u968f\u673a\u7684\u70b9\uff0c\u7136\u540e\u9010\u6e10\u53d8\u6210\u7f8e\u4e3d\u98ce\u666f\u6216\u4eba\u7269\u7684\u89c6\u9891\u3002\u8fd9\u79cd\u65b9\u6cd5\u540e\u6765\u88ab\u5e94\u7528\u4e8e\u56fe\u50cf\u548c\u58f0\u97f3\u7b49\u590d\u6742\u6570\u636e\u7684\u751f\u6210\uff0c\u4fc3\u8fdb\u4e86\u9ad8\u8d28\u91cf\u751f\u6210\u6a21\u578b\u7684\u53d1\u5c55\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/71885b285eb2cdd37bce1ea759271ced.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">\u53bb\u566a\u8fc7\u7a0b\u7684\u56fe\u50cf\uff5c\u56fe\u7247\u6765\u6e90\uff1aOpenAI<\/p>\n<blockquote>\n<p>Ho et al.&nbsp;(2020),&nbsp;\u201cDenoising diffusion probabilistic models.\u201d<\/p>\n<p>Nichol and Dhariwal&nbsp;(2021),&nbsp;\u201cImproved denoising diffusion probabilistic models.\u201d<\/p>\n<\/blockquote>\n<p>\u5728 Sohl-Dickstein \u7b49\u4eba\uff082015\uff09\u63d0\u51fa\u7684\u7406\u8bba\u6846\u67b6\u57fa\u7840\u4e0a\uff0c\u5f00\u53d1\u51fa\u4e86\u88ab\u79f0\u4e3a Denoising Diffusion Probabilistic Models\uff08DDPM\uff09\u7684\u5b9e\u7528\u6570\u636e\u751f\u6210\u6a21\u578b\u3002\u8fd9\u79cd\u6a21\u578b\u5728\u9ad8\u8d28\u91cf\u56fe\u50cf\u751f\u6210\u9886\u57df\u53d6\u5f97\u4e86\u7279\u522b\u663e\u8457\u7684\u6210\u679c\uff0c\u8bc1\u660e\u4e86\u6269\u6563\u6a21\u578b\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><strong>\u6269\u6563\u6a21\u578b\u5bf9&nbsp;Sora&nbsp;\u7684\u5f71\u54cd<\/strong><\/p>\n<p>\u901a\u5e38\u60c5\u51b5\u4e0b\uff0c\u8981\u8bad\u7ec3\u673a\u5668\u5b66\u4e60\u6a21\u578b\uff0c\u9700\u8981\u5927\u91cf\u6807\u6ce8\u6570\u636e\uff08\u6bd4\u5982\uff0c\u544a\u8bc9\u6a21\u578b\u201c\u8fd9\u662f\u4e00\u5f20\u732b\u7684\u56fe\u50cf\u201d\uff09\u3002\u7136\u800c\uff0c\u6269\u6563\u6a21\u578b\u4e5f\u53ef\u4ee5\u4ece\u672a\u88ab\u6807\u6ce8\u7684\u6570\u636e\u4e2d\u5b66\u4e60\uff0c\u4f7f\u5176\u80fd\u591f\u5229\u7528\u4e92\u8054\u7f51\u4e0a\u5927\u91cf\u7684\u89c6\u89c9\u5185\u5bb9\u6765\u751f\u6210\u5404\u79cd\u7c7b\u578b\u7684\u89c6\u9891\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c<strong>Sora \u53ef\u4ee5\u901a\u8fc7\u89c2\u5bdf\u4e0d\u540c\u7684\u89c6\u9891\u548c\u56fe\u50cf\uff0c\u5b66\u4e60\u5230\u201c\u4ec0\u4e48\u662f\u4e00\u4e2a\u6b63\u5e38\u89c6\u9891\u7684\u6837\u5b50\u201d\u3002<\/strong><\/p>\n<p><em>\u6709\u5173 Diffusion Models \u7684\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Fdiffusion-models-made-easy-8414298ce4da\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/diffusion-models-made-easy-8414298ce4da\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/diffusion-m\u2026<\/a><\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Funderstanding-the-denoising-diffusion-probabilistic-model-the-socratic-way-445c1bdc5756\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/understanding-the-denoising-diffusion-probabilistic-model-the-socratic-way-445c1bdc5756\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/understandi\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-10\"><strong>2.7 Latent Diffusion Models<\/strong><\/h2>\n<blockquote>\n<p>Rombach,&nbsp;et al.&nbsp;(2022),&nbsp;\u201cHigh-resolution image synthesis with latent diffusion models.\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u9879\u7814\u7a76\u4e3a\u5229\u7528\u6269\u6563\u6a21\u578b\uff08diffusion models\uff09\u5408\u6210\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u8fd9\u4e00\u9886\u57df\u505a\u51fa\u4e86\u91cd\u5927\u8d21\u732e\u3002\u5b83\u63d0\u51fa\u4e86\u4e00\u79cd\u65b9\u6cd5\uff0c\u4e0e\u76f4\u63a5\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u76f8\u6bd4\uff0c\u8be5\u65b9\u6cd5\u901a\u8fc7\u5229\u7528\u9690\u7a7a\u95f4\uff08latent space\uff09\u4e2d\u7684\u6269\u6563\u6a21\u578b\uff0c\u5728\u4fdd\u8bc1\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u5927\u5927\u964d\u4f4e\u4e86\u8ba1\u7b97\u6210\u672c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5b83\u901a\u8fc7\u5bf9\u5728\u9690\u7a7a\u95f4\uff08\u4e00\u4e2a\u5bb9\u7eb3\u56fe\u50cf\u538b\u7f29\u8868\u5f81\u7684\u4f4e\u7ef4\u7a7a\u95f4\uff09\u4e2d\u8868\u793a\u7684\u6570\u636e\u8fdb\u884c\u7f16\u7801\u5e76\u5f15\u5165\u6269\u6563\u8fc7\u7a0b\uff0c\u53ef\u4ee5\u7528\u66f4\u5c11\u7684\u8ba1\u7b97\u8d44\u6e90\u5b9e\u73b0\u76ee\u6807\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u64cd\u4f5c\u56fe\u50cf\u3002<\/p>\n<p><strong>Sora \u5c06\u8fd9\u4e00\u6280\u672f\u5e94\u7528\u4e8e\u89c6\u9891\u6570\u636e\uff0c\u5c06\u89c6\u9891\u7684\u65f6\u95f4+\u7a7a\u95f4\u6570\u636e\u538b\u7f29\u5230\u8f83\u4f4e\u7ef4\u5ea6\u7684\u9690\u7a7a\u95f4\uff0c\u7136\u540e\u5c06\u5176\u5206\u89e3\u4e3a\u65f6\u7a7a\u788e\u7247\uff08spatiotemporal patches\uff09\u3002\u8fd9\u79cd\u9ad8\u6548\u7684\u9690\u7a7a\u95f4\u6570\u636e\u5904\u7406\u548c\u751f\u6210\u80fd\u529b\uff0c\u5728\u4f7f Sora \u80fd\u591f\u66f4\u5feb\u5730\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u89c6\u89c9\u5185\u5bb9\u65b9\u9762\u53d1\u6325\u4e86\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002<\/strong><\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/7419d9283556720a244f6861bdc20062.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Image of visual encoding\uff5cImage Credit&nbsp;(OpenAI)<\/p>\n<p><em>\u6709\u5173 Latent Diffusion Models \u7684\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Fpaper-explained-high-resolution-image-synthesis-with-latent-diffusion-models-f372f7636d42\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/paper-explained-high-resolution-image-synthesis-with-latent-diffusion-models-f372f7636d42\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/paper-expla\u2026<\/a><\/em><\/p>\n<h2 data-id=\"heading-11\"><strong>2.8 Diffusion Transformer&nbsp;(DiT)<\/strong><\/h2>\n<blockquote>\n<p>Peebles and Xie.&nbsp;(2023),&nbsp;\u201cScalable diffusion models with transformers.\u201d<\/p>\n<\/blockquote>\n<p>\u8fd9\u9879\u7814\u7a76\u53ef\u80fd\u662f\u5b9e\u73b0 Sora \u6700\u5173\u952e\u7684\u90e8\u5206\u3002\u6b63\u5982 OpenAI \u53d1\u5e03\u7684\u6280\u672f\u62a5\u544a\u6240\u8ff0\uff0cSora \u91c7\u7528\u7684\u4e0d\u662f\u666e\u901a\u7684 transformer \uff0c\u800c\u662f diffusion transformer\uff08DiT\uff09\u3002<\/p>\n<blockquote>\n<p>Importantly,&nbsp;Sora is a diffusion transformer.&nbsp;(via OpenAI Sora technical report)<\/p>\n<\/blockquote>\n<p>\u8fd9\u9879\u7814\u7a76\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684\u6a21\u578b\uff0c<strong>\u7528&nbsp;Transformer&nbsp;\u7ed3\u6784\u66ff\u4ee3\u4e86\u6269\u6563\u6a21\u578b\u4e2d\u5e38\u7528\u7684&nbsp;U-net&nbsp;\u7ec4\u4ef6<\/strong>\u3002\u8fd9\u79cd\u7ed3\u6784\u901a\u8fc7 Transformer \u5bf9 latent patches \u7684\u64cd\u4f5c\u5b9e\u73b0 Latent Diffusion Model\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u66f4\u9ad8\u6548\u5730\u5904\u7406image patches\uff0c\u4ece\u800c\u5728\u6709\u6548\u5229\u7528\u8ba1\u7b97\u8d44\u6e90\u7684\u540c\u65f6\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u3002\u4e0e 2022 \u5e74&nbsp;Stability AI \u5ba3\u5e03\u7684 Stable Diffusion \u4e0d\u540c\uff0c\u5f15\u5165\u8fd9\u79cd Transformer \u88ab\u8ba4\u4e3a\u6709\u52a9\u4e8e\u66f4\u81ea\u7136\u5730\u751f\u6210\u89c6\u9891\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/2598cbe4906e1f75a40971bb0e898b44.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Diffusion Transformers\u751f\u6210\u7684\u56fe\u50cf\uff5cPeebles and Xie.&nbsp;(2023)<\/p>\n<p>\u6b64\u5916\uff0c\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4ed6\u4eec\u7684\u9a8c\u8bc1\u7ed3\u679c\u8bc1\u660e\u4e86 <strong>DiT&nbsp;\u5177\u5907\u53ef\u6269\u5c55\u6027\uff0c\u4e3a&nbsp;Sora&nbsp;\u7684\u5b9e\u73b0\u505a\u51fa\u4e86\u91cd\u5927\u8d21\u732e<\/strong>\u3002\u5177\u5907\u53ef\u6269\u5c55\u6027\u610f\u5473\u7740\u6a21\u578b\u7684\u6027\u80fd\u80fd\u591f\u968f\u7740 Transformer \u7684\u6df1\u5ea6\/\u5bbd\u5ea6\uff08\u4f7f\u6a21\u578b\u66f4\u590d\u6742\uff09\u6216\u8f93\u5165 token \u6570\u91cf\u7684\u589e\u52a0\u800c\u63d0\u9ad8\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/463395239b9fdc46dd0386e23d375e97.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Diffusion Transformers&nbsp;\u7684\u53ef\u6269\u5c55\u6027\uff5cPeebles and Xie.&nbsp;(2023)<\/p>\n<ul>\n<li>Gflops\uff08\u8ba1\u7b97\u6027\u80fd\uff09\uff1a\u8ba1\u7b97\u673a\u8ba1\u7b97\u901f\u5ea6\u7684\u5ea6\u91cf\u5355\u4f4d\uff0c\u76f8\u5f53\u4e8e\u6bcf\u79d2\u5341\u4ebf\u6b21\u6d6e\u70b9\u8fd0\u7b97\u3002\u5728\u672c\u6587\u4e2d\uff0c\u7f51\u7edc\u590d\u6742\u5ea6\uff08network complexity\uff09\u901a\u8fc7 Gflops \u8fdb\u884c\u8861\u91cf\u3002<\/li>\n<li>FID\uff08Fr\u00e9chet Inception Distance\uff09\uff1a\u8fd9\u662f\u56fe\u50cf\u751f\u6210\u7684\u8bc4\u4f30\u6307\u6807\u4e4b\u4e00\uff0c\u6570\u503c\u8d8a\u5c0f\u8868\u793a\u51c6\u786e\u6027\u8d8a\u9ad8\u3002\u5b83\u901a\u8fc7\u6d4b\u91cf\u751f\u6210\u56fe\u50cf\u548c\u771f\u5b9e\u56fe\u50cf\u7684\u7279\u5f81\u5411\u91cf\u4e4b\u95f4\u7684\u8ddd\u79bb\u6765\u5b9a\u91cf\u8bc4\u4f30\u751f\u6210\u56fe\u50cf\u7684\u8d28\u91cf\u3002<\/li>\n<\/ul>\n<p>Kaplan&nbsp;\u7b49\u4eba\uff082020\uff09\u548c&nbsp;Brown&nbsp;\u7b49\u4eba\uff082020\uff09\u5df2\u7ecf\u8bc1\u5b9e\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u5df2\u7ecf\u89c2\u5bdf\u5230\u4e86\u8fd9\u4e00\u70b9\uff08\u8bd1\u8005\u6ce8\uff1a\u6b64\u5904\u5e94\u5f53\u6307\u7684\u662f\u201c\u5b58\u5728\u53ef\u6269\u5c55\u6027\u201d\uff09\uff0c\u8fd9\u4e5f\u662f\u652f\u6301 ChatGPT \u521b\u65b0\u6210\u529f\u80cc\u540e\u7684\u5173\u952e\u7279\u6027\u3002<\/p>\n<blockquote>\n<p>Kaplan et al.&nbsp;(2020),&nbsp;\u201cScaling Laws for Neural Language Models.\u201d<\/p>\n<p>Brown,&nbsp;et al.&nbsp;(2020),&nbsp;\u201cLanguage models are few-shot learners.\u201d<\/p>\n<\/blockquote>\n<p>\u4e0e\u4f20\u7edf\u7684\u6269\u6563\u6a21\u578b\uff08diffusion models\uff09\u76f8\u6bd4\uff0c\u7531\u4e8e&nbsp;Transformer&nbsp;\u7684\u4f18\u52bf\uff0c<strong>\u5b83\u80fd\u4ee5\u66f4\u4f4e\u7684\u8ba1\u7b97\u6210\u672c\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf<\/strong>\uff0c\u800c\u8fd9\u4e00\u663e\u8457\u7279\u70b9\u8868\u660e\uff0c\u4f7f\u7528\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\u751a\u81f3\u53ef\u4ee5\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u3002Sora \u5c06\u8fd9\u9879\u6280\u672f\u5e94\u7528\u4e8e\u89c6\u9891\u751f\u6210\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/19abf4ac19bb090852b6aa431bdc806e.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">\u89c6\u9891\u751f\u6210\u7684\u53ef\u6269\u5c55\u6027\uff5cImage Credit&nbsp;(OpenAI)<\/p>\n<p><em>\u6709\u5173 DiT \u7684\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u4ee5\u4e0b\u94fe\u63a5\uff1a<\/em><\/p>\n<p><em><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Fyoutu.be%2FeTBG17LANcI\" target=\"_blank\" title=\"https:\/\/youtu.be\/eTBG17LANcI\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">youtu.be\/eTBG17LANcI<\/a><\/em><\/p>\n<h1 data-id=\"heading-12\"><strong>03 \u8fd9\u4e9b\u7814\u7a76\u57fa\u7840\u52a0\u4e0aOpenAI\u7684\u52aa\u529b\u5171\u540c\u9020\u5c31\u4e86&nbsp;Sora<\/strong><\/h1>\n<h2 data-id=\"heading-13\"><strong>3.1&nbsp;\u53ef\u53d8\u7684\u89c6\u9891\u65f6\u957f\u3001\u5206\u8fa8\u7387\u3001\u957f\u5bbd\u6bd4<\/strong><\/h2>\n<p>\u4e3b\u8981\u5f97\u76ca\u4e8eNaViT\uff0cSora\u80fd\u591f\u751f\u6210widescreen 1920x1080p\u89c6\u9891\u3001vertical 1080&#215;1920\u89c6\u9891\u4ee5\u53ca\u4ecb\u4e8e\u4e24\u8005\u4e4b\u95f4\u7684\u6240\u6709\u89c6\u9891\u3002\u8fd9\u610f\u5473\u7740\u5b83\u53ef\u4ee5\u4e3a\u5404\u79cd\u8bbe\u5907\u7c7b\u578b\u521b\u5efa\u4efb\u4f55\u5206\u8fa8\u7387\u7684\u89c6\u89c9\u5185\u5bb9\u3002<\/p>\n<h2 data-id=\"heading-14\"><strong>3.2&nbsp;\u4f7f\u7528\u56fe\u50cf\u548c\u89c6\u9891\u4f5c\u4e3a&nbsp;Prompt<\/strong><\/h2>\n<p>\u76ee\u524d\uff0cSora \u4ee5 text-to-video \u7684\u683c\u5f0f\u5b9e\u73b0\u89c6\u9891\u751f\u6210\uff0c\u5373\u901a\u8fc7\u6587\u672c\u63d0\u793a\u8bcd\u7ed9\u51fa\u6307\u4ee4\u751f\u6210\u89c6\u9891\u3002\u4e0d\u8fc7\uff0c\u4ece\u524d\u9762\u7684\u7814\u7a76\u4e2d\u4e0d\u96be\u770b\u51fa\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528\u73b0\u6709\u7684\u56fe\u7247\u6216\u89c6\u9891\u4f5c\u4e3a\u8f93\u5165\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u6587\u5b57\u3002\u8fd9\u6837\uff0cSora \u5c31\u53ef\u4ee5\u5c06\u56fe\u50cf\u5236\u4f5c\u6210\u52a8\u753b\uff0c\u6216\u5c06\u73b0\u6709\u89c6\u9891\u7684\u8fc7\u53bb\u6216\u672a\u6765\u60f3\u8c61\u6210\u89c6\u89c9\u5185\u5bb9\u5e76\u8f93\u51fa\u3002<\/p>\n<h2 data-id=\"heading-15\"><strong>3.3 3D consistency<\/strong><\/h2>\n<p>\u867d\u7136\u4e0d\u6e05\u695a\u4e0a\u8ff0\u7814\u7a76\u5982\u4f55\u76f4\u63a5\u53c2\u4e0e\u5176\u4e2d\uff0c\u5e2e\u52a9\u5b9e\u73b0\u8fd9\u4e00\u7279\u6027\u3002\u4f46 Sora \u53ef\u4ee5\u751f\u6210\u5177\u6709dynamic camera motion\u6548\u679c\uff08\u8bd1\u8005\u6ce8\uff1adynamic camera motion \u8868\u660e\u89c6\u9891\u4e0d\u662f\u9759\u6b62\u4e0d\u52a8\u7684\uff0c\u800c\u662f\u968f\u7740\u65f6\u95f4\u53d8\u5316\u800c\u79fb\u52a8\u3001\u65cb\u8f6c\u6216\u6539\u53d8\u89c6\u89d2\u3002\uff09\u7684\u89c6\u9891\u3002\u968f\u7740\u201c\u6444\u50cf\u673a\u201d\u7684\u79fb\u52a8\u548c\u65cb\u8f6c\uff0c\u4eba\u7269\u548c\u573a\u666f\u5143\u7d20\u80fd\u591f\u5728\u4e09\u7ef4\u7a7a\u95f4\u4e2d\u4fdd\u6301\u4e00\u81f4\u5730\u79fb\u52a8\u3002<\/p>\n<h1 data-id=\"heading-16\"><strong>04 \u5c55\u671bSora\u7684\u672a\u6765<\/strong><\/h1>\n<p>\u8fd9\u7bc7\u535a\u6587\u8be6\u7ec6\u4ecb\u7ecd\u4e86OpenAI\u7528\u4e8e\u751f\u6210\u89c6\u9891\u7684\u901a\u7528\u89c6\u89c9\u6a21\u578b Sora \u80cc\u540e\u7684\u6280\u672f\u3002\u4e00\u65e6 Sora \u80fd\u591f\u5411\u516c\u4f17\u5f00\u653e\uff0c\u8ba9\u66f4\u591a\u4eba\u4f7f\u7528\uff0c\u5fc5\u5c06\u5728\u5168\u7403\u8303\u56f4\u5185\u4ea7\u751f\u66f4\u52a0\u91cd\u5927\u7684\u5f71\u54cd\u3002<\/p>\n<p>\u8fd9\u4e00\u7a81\u7834\u6240\u5e26\u6765\u7684\u5f71\u54cd\u9884\u8ba1\u5c06\u6db5\u76d6\u89c6\u9891\u521b\u4f5c\u7684\u5404\u4e2a\u65b9\u9762\uff0c\u4f46\u636e\u9884\u6d4b\uff0c<strong>Sora \u53ef\u80fd\u5728\u89c6\u9891\u9886\u57df\u624e\u6839\u540e\u7ee7\u7eed\u8fdb\u519b\u4e09\u7ef4\u5efa\u6a21\u9886\u57df\u3002<\/strong> \u5c4a\u65f6\uff0c\u4e0d\u4ec5\u5bf9\u89c6\u9891\u521b\u4f5c\u8005\u4ea7\u751f\u5f71\u54cd\uff0c\u5c31\u8fde\u865a\u62df\u7a7a\u95f4\uff08\u5982\u5143\u5b87\u5b99\uff09\u4e2d\u7684\u89c6\u89c9\u6548\u679c\u5236\u4f5c\u4e5f\u80fd\u5f88\u5feb\u7531\u4eba\u5de5\u667a\u80fd\u8f7b\u677e\u751f\u6210\u3002<\/p>\n<p>\u4e0b\u56fe\u5df2\u7ecf\u6697\u793a\u4e86\u8fd9\u79cd\u60c5\u51b5\u672a\u6765\u53ef\u80fd\u4f1a\u51fa\u73b0\uff1a<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/ba7c782635ae78f5cb28834e207a1a75.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Martin Nebelong&nbsp;\u901a\u8fc7&nbsp;X&nbsp;\u53d1\u5e03\u7684\u4e0e&nbsp;Micael Rublof&nbsp;\u4ea7\u54c1\u76f8\u5173\u7684\u5e16\u5b50<\/p>\n<p>\u76ee\u524d\uff0cSora\u88ab\u90e8\u5206\u4eba\u8ba4\u4e3a\u201c\u4ec5\u4ec5\u201d\u662f\u4e00\u4e2a\u89c6\u9891\u751f\u6210\u6a21\u578b\uff0c\u4f46Nvidia\u7684Jim Fan\u6697\u793a\u5b83\u53ef\u80fd\u662f\u4e00\u4e2a\u6570\u636e\u9a71\u52a8\u7684\u7269\u7406\u5f15\u64ce\u3002<strong>\u4eba\u5de5\u667a\u80fd\u6709\u53ef\u80fd\u4ece\u5927\u91cf\u771f\u5b9e\u4e16\u754c\u7684\u89c6\u9891\u548c\uff08\u867d\u7136\u6ca1\u6709\u660e\u786e\u63d0\u5230\uff09\u9700\u8981\u8003\u8651\u7269\u7406\u884c\u4e3a\u7684\u89c6\u9891\uff08\u5982\u865a\u5e7b\u5f15\u64ce\u4e2d\u7684\u89c6\u9891\uff09\u4e2d\u7406\u89e3\u7269\u7406\u89c4\u5f8b\u548c\u73b0\u8c61\u3002<\/strong> \u5982\u679c\u662f\u8fd9\u6837\uff0c\u90a3\u4e48\u5728\u4e0d\u4e45\u7684\u5c06\u6765\u51fa\u73b0 text-to-3D \u6a21\u578b\u7684\u53ef\u80fd\u6027\u4e5f\u662f\u975e\u5e38\u9ad8\u7684\u3002<\/p>\n<p>\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.nicekj.com\/wp-content\/uploads\/replace\/d73cbd3e692c775897bee9dce22f3993.png\" alt=\"\u5316\u662f\u6e10\u5316\uff0c\u53d8\u662f\u987f\u53d8\uff1a\u4e00\u7aa5 OpenAI Sora \u76f8\u5173\u6280\u672f\u7684\u6f14\u8fdb\" \/><\/figure>\n<\/p>\n<p align=\"center\">Jim Fan\u2019s intriguing post via X<\/p>\n<p><strong>Thanks for reading!<\/strong><\/p>\n<p><strong>END<\/strong><\/p>\n<p><strong><strong>\u672c\u6587\u7ecf\u539f\u4f5c\u8005\u6388\u6743\uff0c\u7531Baihai IDP\u7f16\u8bd1\u3002\u5982\u9700\u8f6c\u8f7d\u8bd1\u6587\uff0c\u8bf7\u8054\u7cfb\u83b7\u53d6\u6388\u6743\u3002<\/strong><\/strong><\/p>\n<p><strong>\u539f\u6587\u94fe\u63a5\uff1a<\/strong><\/p>\n<p><a href=\"https:\/\/link.juejin.cn?target=https%3A%2F%2Ftowardsdatascience.com%2Fhow-openais-sora-is-changing-the-game-an-insight-into-its-core-technologies-bd1ad17170df\" target=\"_blank\" title=\"https:\/\/towardsdatascience.com\/how-openais-sora-is-changing-the-game-an-insight-into-its-core-technologies-bd1ad17170df\" ref=\"nofollow noopener noreferrer\" rel=\"noopener\">towardsdatascience.com\/how-openais\u2026<\/a><\/p>","protected":false},"excerpt":{"rendered":"<p>\u6211\u4eec\u4eca\u5929\u8981\u4e3a\u5927\u5bb6\u5206\u4eab\u7684\u8fd9\u7bc7\u535a\u6587\uff0c\u4f5c\u8005\u8ba4\u4e3a Sora \u4ee3\u8868\u4e86Transformer\u3001NaViT\u3001\u6269\u6563\u6a21\u578b\u7b49\u4e00\u7cfb\u5217\u89c6\u89c9AI\u6280\u672f\u7684\u878d\u5408\u521b\u65b0\uff0c\u662f\u8fc8\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7684\u91cd\u8981\u4e00\u6b65\u3002<\/p>\n","protected":false},"author":1,"featured_media":11594,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_title":"","rank_math_description":"","rank_math_focus_keyword":"","views":"4","footnotes":""},"categories":[3],"tags":[136,126,127,128,129],"collection":[],"class_list":["post-1747","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-fenlei2","tag-136","tag-gpt","tag-ai","tag-128","tag-129"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1747","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/comments?post=1747"}],"version-history":[{"count":0,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1747\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media\/11594"}],"wp:attachment":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media?parent=1747"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/categories?post=1747"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/tags?post=1747"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/collection?post=1747"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}