{"id":14351,"date":"2025-04-08T22:06:16","date_gmt":"2025-04-08T14:06:16","guid":{"rendered":"https:\/\/blog.konata.vip\/?p=14351"},"modified":"2025-04-08T22:16:24","modified_gmt":"2025-04-08T14:16:24","slug":"gpt-4o-%e7%94%9f%e5%9b%be%e6%8a%80%e6%9c%af%e7%9a%84%e4%b8%aa%e4%ba%ba%e7%8c%9c%e6%83%b3","status":"publish","type":"post","link":"https:\/\/blog.konata.vip\/?p=14351","title":{"rendered":"GPT-4o \u751f\u56fe\u6280\u672f\u7684\u4e2a\u4eba\u731c\u60f3"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"768\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7-1024x768.jpg\" alt=\"\" class=\"wp-image-14353\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7-1024x768.jpg 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7-300x225.jpg 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7-768x576.jpg 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7-1536x1152.jpg 1536w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/898ebd5bcb08d8a093fb4c001e7d3b7.jpg 1704w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p> \u56fe\u6e90\u5c0f\u7ea2\u4e66@\u6076\u9b54\u5e7c\u5d3d<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">GPT-4o \u7684\u795e\u5947\u9b54\u6cd5<\/h2>\n\n\n\n<p>\u6700\u8fd1GPT-4o\u751f\u56fe\u6a21\u578b\u6a2a\u7a7a\u51fa\u4e16\uff0c\u6548\u679c\u548c\u73a9\u6cd5\u4e0a\u90fd\u6709\u7a81\u7834\u6027\u7684\u8fdb\u5c55\uff0c\u4f5c\u4e3a\u4eceSD\u5f00\u59cb\u4e00\u8def\u8ddf\u8fdb\u6587\u751f\u56fe\u6a21\u578b\u6280\u672f\u7684\u4eba\u6765\u8bf4\u770b\u5230\u8fd9\u6837\u7684\u6548\u679c\u771f\u7684\u975e\u5e38\u9707\u60ca\uff0c\u5f88\u597d\u5947\u662f\u600e\u4e48\u8fbe\u5230\u7684\uff0c\u6240\u4ee5\u8fb9\u7814\u7a76\u8fb9\u6574\u7406\u4e86\u4e00\u4e0b\u76ee\u524d\u76f8\u5173\u7684\u6280\u672f\uff0c\u629b\u7816\u5f15\u7389\u4e00\u4e0b\uff0c\u5e0c\u671b\u6709\u66f4\u591a\u5927\u795e\u5206\u4eab\u8ba8\u8bba\uff08\u6700\u597d\u662fOpenAI\u53d1\u6280\u672f\u62a5\u544a\u591a\u900f\u9732\u70b9\uff09\u3002<\/p>\n\n\n\n<p>\u9996\u5148\u53ef\u80fd\u5f88\u591a\u670b\u53cb\u53ea\u5237\u5230\u8fc7\u7279\u522b\u706b\u4e5f\u4e89\u8bae\u7279\u522b\u5927\u7684\u90a3\u4e2a\u5409\u535c\u529b\u98ce\u683c\u5316\uff0c\u4f46\u8fd9\u5176\u5b9e\u4e0d\u662f\u4ec0\u4e48\u65b0\u9c9c\u7684\u73a9\u6cd5\u4e86\uff0c\u53ea\u4e0d\u8fc7\u5728\u6548\u679c\u4e0a\u53c8\u66f4\u8fdb\u4e00\u6b65\u3002\u66f4\u52a0\u503c\u5f97\u5173\u6ce8\u7684\u5176\u5b9e\u662f\u8fd9\u6b21GPT-4o\u51fa\u4e86\u8d85\u591a\u7279\u522b\u60ca\u4eba\u7684\u6280\u672f\uff0c\u57fa\u672c\u4e0a\u53ef\u4ee5\u8bf4\u662fGPT\u672cT\u957f\u4e86\u624b\u80fd\u753b\u753b\u4e86\uff0c\u4e0d\u518d\u5b58\u5728GPT\u751f\u6210prompt\u547d\u4ee4\u547c\u53eb DALLE3\u5bfc\u81f4\u7684\u4fe1\u606f\u635f\u8017\u4e86\u3002\uff08\u6700\u5f00\u59cbGPT\u8fd8\u4e0d\u4f1a\u81ea\u5df1\u753b\u56fe\u7684\u65f6\u5019\uff0c\u662f\u901a\u8fc7\u8c03\u7528\u6587\u751f\u56fe\u5927\u6a21\u578bDALLE3\u6765\u753b\u56fe\u7684\uff0c\u800c\u4e0d\u662f\u5b83\u672c\u8eab\u80fd\u4ea7\u751f\u56fe\u7247\uff09<\/p>\n\n\n\n<p>\u4f8b\u5982\u5b83\u53ef\u4ee5\u5199\u975e\u5e38\u975e\u5e38\u957f\u7684\u6587\u5b57\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"683\" height=\"1024\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/ChatGPT_Image_Mar_24__2025__07_55_11_AM-683x1024.webp\" alt=\"\" class=\"wp-image-14354\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/ChatGPT_Image_Mar_24__2025__07_55_11_AM-683x1024.webp 683w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/ChatGPT_Image_Mar_24__2025__07_55_11_AM-200x300.webp 200w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/ChatGPT_Image_Mar_24__2025__07_55_11_AM-768x1152.webp 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/ChatGPT_Image_Mar_24__2025__07_55_11_AM.webp 828w\" sizes=\"auto, (max-width: 683px) 100vw, 683px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u4f8b\u5982\u7ed9\u5b9a\u98ce\u683c\u53c2\u8003\u56fe\u8ba9\u5b83\u753b\u4e00\u4e2a\u7c7b\u4f3c\u7684\u6d77\u62a5\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"754\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407160937-1024x754.jpg\" alt=\"\" class=\"wp-image-14355\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407160937-1024x754.jpg 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407160937-300x221.jpg 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407160937-768x565.jpg 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407160937.jpg 1312w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u4f8b\u5982\u76f4\u63a5\u901a\u8fc7\u5bf9\u8bdd\u4fee\u6539\u56fe\u7247:<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"764\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407161154-1024x764.jpg\" alt=\"\" class=\"wp-image-14357\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407161154-1024x764.jpg 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407161154-300x224.jpg 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407161154-768x573.jpg 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/\u5fae\u4fe1\u622a\u56fe_20250407161154.jpg 1299w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u4f8b\u5982\u753b\u4e00\u4e9b\u6f2b\u753b\u6216\u8005\u63d2\u753b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"685\" height=\"1024\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Screenshot_2025-03-24_at_10.08.43_PM-685x1024.webp\" alt=\"\" class=\"wp-image-14358\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Screenshot_2025-03-24_at_10.08.43_PM-685x1024.webp 685w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Screenshot_2025-03-24_at_10.08.43_PM-201x300.webp 201w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Screenshot_2025-03-24_at_10.08.43_PM.webp 692w\" sizes=\"auto, (max-width: 685px) 100vw, 685px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u66f4\u591a\u4f8b\u5b50\uff1a<br>\n<a href=\"https:\/\/openai.com\/index\/introducing-4o-image-generation\/\" title=\"https:\/\/openai.com\/index\/introducing-4o-image-generation\/\">https:\/\/openai.com\/index\/introducing-4o-image-generation\/<\/a><\/p>\n\n\n\n<p>\u8fd9\u8fd8\u53ea\u662f\u5b98\u7f51\u7684\u793a\u4f8b\uff0c\u8fd9\u51e0\u5929\u7f51\u53cb\u5df2\u7ecf\u73a9\u51fa\u66f4\u591a\u5c42\u51fa\u4e0d\u7a77\u7684\u82b1\u6837\u4e86\uff0c\u6bd4\u5982\u8f6c\u6210\u52a8\u68ee\u98ce\u683c\u3001\u628a\u5bb6\u732b\u62df\u4eba\u5316\u3001\u5236\u4f5c\u8868\u60c5\u5305\u7b49\u7b49\uff0c\u53ea\u80fd\u8bf4\u4eba\u7c7b\u5bf9GPT 4o\u751f\u56fe\u80fd\u529b\u7684\u5f00\u53d1\u8fd8\u4e0d\u8db31%\u3002\u8fd9\u4e9b\u4f8b\u5b50\uff0c\u90fd\u53ef\u4ee5\u4ece\u5404\u79cd\u610f\u4e49\u4e0a\u4f53\u73b0\u51fa\uff0cGPT\u6a21\u578b\u81ea\u5df1\u957f\u4e86\u4e2a\u4f1a\u753b\u753b\u7684\u624b\uff0c\u6bd4\u8c03\u7528\u53e6\u4e00\u4e2a\u7ed8\u56fe\u6a21\u578b\u5728\u4f7f\u7528\u4e0a\u6709\u591a\u5927\u7684\u53d8\u9769\u3002<\/p>\n\n\n\n<p>\u6bd4\u5982\u8bf4\u5236\u9020\u62b9\u8336\u90a3\u4e2a\u56db\u683c\uff0c\u5982\u679c\u662f\u4e4b\u524d\u7684\u6d41\u7a0b\uff0c\u53ef\u80fd\u9700\u8981\u5148\u8ba9GPT\u751f\u6210\u62b9\u8336\u505a\u6cd5\u7684\u56db\u4e2a\u6b65\u9aa4\uff0c\u7136\u540e\u518d\u8ba9\u5b83\u628a\u6b65\u9aa4\u8f6c\u6362\u6210\u6587\u751f\u56fe\u6a21\u578b\u7684prompt\uff0c\u518d\u8c03\u7528DALLE3\u628a\u56fe\u7247\u753b\u51fa\u6765\u3002\u6bcf\u4e00\u4e2a\u6b65\u9aa4\u90fd\u53ef\u80fd\u6709\u635f\u8017\uff0c\u5c24\u5176\u662f\u6700\u540e\u4e00\u6b65\uff0cDALLE3\u53ef\u80fd\u6ca1\u529e\u6cd5\u6709\u8fd9\u4e48\u597d\u7684\u6587\u672c\u5bf9\u5e94\u6027\u80fd\u7cbe\u51c6\u753b\u51fa\u8fd9\u6837\u7684\u5185\u5bb9\u3002\u4f46\u662fGPT\u5982\u679c\u672c\u6765\u5c31\u4f1a\u753b\u753b\uff0c\u4e8b\u60c5\u5c31\u4f1a\u53d8\u5f97\u5f88\u81ea\u7136\u4e86\u3002\u6211\u4eec\u90fd\u77e5\u9053GPT\u5185\u90e8\u80af\u5b9a\u662f\u5df2\u7ecf\u6709\u5173\u4e8e\u5982\u4f55\u5236\u4f5c\u62b9\u8336\u7684\u77e5\u8bc6\u7684\uff0c\u4e0d\u6b62\u5982\u6b64\uff0cGPT-4o\u8fd8\u662f\u4e00\u4e2a\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002\u591a\u6a21\u6001\u7684\u610f\u601d\u662f\u5b83\u80fd\u591f\u7406\u89e3\u5404\u79cd\u5f62\u5f0f\u7684\u4fe1\u606f\uff0c\u4f8b\u5982\u6587\u5b57\u3001\u56fe\u7247\u3001\u8bed\u8a00\u7b49\u3002\u6240\u4ee5\u5b83\u540c\u6837\u4e5f\u89c1\u8fc7\u5927\u91cf\u7684\u201c\u5982\u4f55\u5236\u4f5cxx\u201d\u8fd9\u6837\u7684\u6d77\u62a5\u3002\u4e8e\u662f\u5b83\u5c31\u53ef\u4ee5\u76f4\u63a5\u7ed3\u5408\u8fd9\u4e9b\u77e5\u8bc6\u751f\u6210\u8fd9\u4e2a\u7ed3\u679c\u3002<\/p>\n\n\n\n<p>\u53c8\u4f8b\u5982\u732b\u732b\u90a3\u5f20\u56fe\u7ed9\u5b9a\u53c2\u8003\u56fe\u6765\u753b\u65b0\u7684\u56fe\u548c\u7528\u5bf9\u8bdd\u7f16\u8f91\u56fe\u7247\u4e5f\u662f\u975e\u5e38\u81ea\u7136\u3002\u539f\u672c\u5982\u679c\u8981\u505a\u8fd9\u4e9b\uff0c\u7528\u8fc7SD\u7cfb\u5217\u5de5\u5177\u7684\u4eba\u4f1a\u77e5\u9053\uff0c\u4f60\u9700\u8981\u6709\u4e00\u4e2a\u5b9e\u73b0\u8fd9\u4e9b\u529f\u80fd\u7684workflow\uff0c\u628a\u5404\u79cd\u6a21\u5757\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u8ba9\u8fd9\u4e9b\u53c2\u8003\u56fe\u4ee5\u5404\u79cd\u5f62\u5f0f\u6ce8\u5165\u4fe1\u606f\u5230\u6a21\u578b\uff0c\u7136\u540ediffusion\u6a21\u578b\u518d\u63a5\u53d7\u8fd9\u4e9b\u4fe1\u606f\u4f5c\u4e3a\u751f\u6210\u56fe\u7247\u7684\u6761\u4ef6\u53bb\u5b8c\u6210\u6307\u4ee4\u3002\u4f46\u5176\u5b9ediffusion\u539f\u751f\u662f\u4e0d\u652f\u6301\u8fd9\u4e9b\u7f16\u8f91\u529f\u80fd\u7684\uff0c\u6240\u4ee5\u6bcf\u4e2a\u529f\u80fd\u90fd\u9700\u8981\u5355\u72ec\u4f5c\u4e3a\u4e00\u4e2a\u4efb\u52a1\u53bb\u8bad\u7ec3\u4f18\u5316\u3002\u4f46\u662f\u5176\u5b9e\u8fd9\u4e9b\u4efb\u52a1\u672c\u8d28\u4e0a\u90fd\u662f\u5dee\u4e0d\u591a\u7684\uff0c\u53ea\u4e0d\u8fc7diffusion\u6ca1\u6709auto regressive model\u8fd9\u6837\u4e00\u5207\u7686\u53eftokenized\u7684flexibility\uff08\u8fd9\u90e8\u5206\u6280\u672f\u540e\u9762\u518d\u89e3\u91ca\uff09\u3002\u6240\u4ee5GPT\u4e00\u65e6\u80fd\u591f\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u56fe\u7247\uff0c\u8fd9\u4e9b\u56fe\u7247\u7f16\u8f91\u64cd\u4f5c\u80fd\u505a\u5f97\u6bd4diffusion\u6a21\u578b\u597d\u662f\u975e\u5e38\u81ea\u7136\u800c\u7136\u6c34\u5230\u6e20\u6210\u7684\u4e8b\u60c5\u3002<\/p>\n\n\n\n<p>OpenAI\u5728\u505aGPT 3.5\uff08\u4e5f\u5c31\u662f\u6700\u521d\u7684ChatGPT\uff09\u7684\u65f6\u5019\u5c31\u5f88\u559c\u6b22\u8fd9\u79cd\u5927\u4e00\u7edf\u7684\u8def\u7ebf\uff0c\u5728\u5927\u5bb6\u90fd\u5728\u505a\u4e13\u7528\u6a21\u578b\uff08\u7ffb\u8bd1\u6a21\u578b\u3001\u6253\u5206\u6a21\u578b\u3001\u804a\u5929\u673a\u5668\u4eba\u7b49\uff09\u7684\u65f6\u5019\u575a\u6301\u505a\u901a\u7528\u6a21\u578b\uff08\u4e00\u4e2a\u6a21\u578b\u89e3\u51b3\u6240\u6709\u4efb\u52a1\uff09\u3002\u5b83\u4eec\u6700\u5f00\u59cb\u7684DALLE 1\u4e5f\u662f\u57fa\u4e8e\u81ea\u56de\u5f52\u7684\uff0c\u4eceDALLE 2\u5f00\u59cb\u81ea\u56de\u5f52\u6548\u679c\u88ab\u6269\u6563\u6a21\u578b\u8d85\u8d8a\uff0cDALLE 3\u5c31\u662f\u5b8c\u5168\u57fa\u4e8e\u6269\u6563\u6a21\u578b\u4e86\u3002\u4f46\u73b0\u5728\u770b\u6765\uff0c\u5b83\u4eec\u5185\u90e8\u5e94\u8be5\u4e00\u76f4\u6ca1\u6709\u653e\u5f03\u8fc7\u81ea\u56de\u5f52\u7684\u8def\u7ebf\uff0c\u56e0\u4e3a\u591a\u6a21\u6001\u81ea\u56de\u5f52\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u628a\u6587\u5b57\u3001\u56fe\u7247\u3001\u58f0\u97f3\u7b49\u4fe1\u606f\u7edf\u4e00\u5230\u4e00\u8d77\uff0c\u4ece\u957f\u8fdc\u770b\u6765\u6709\u66f4\u5927\u7684\u597d\u5904\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">GPT-4o \u53ef\u80fd\u7684\u6280\u672f\u8def\u7ebf\u63a8\u6d4b<\/h2>\n\n\n\n<p>\u90a3\u4e48GPT-4o\u662f\u5982\u4f55\u505a\u5230\u7684\u5462\uff1f\u63a5\u4e0b\u6765\u6211\u8bf4\u7684\u662f\u6211\u7ed3\u5408\u793e\u533a\u8ba8\u8bba\u7684\u4e00\u4e9b\u731c\u60f3\uff0c\u4e0d\u4ee3\u8868\u5b9e\u9645\u4f7f\u7528\u7684\u6280\u672f\u3002<\/p>\n\n\n\n<p>\u9996\u5148\u662f\u4ee5CloseAI\u8457\u79f0\u7684OpenAI\u4f9d\u7136\u51e0\u4e4e\u6ca1\u6709\u900f\u9732\u4efb\u4f55\u6280\u672f\u7ec6\u8282\uff0c\u53ea\u8bf4\u4e86\u65b0\u6a21\u578b\u662f\u57fa\u4e8e\u81ea\u56de\u5f52\uff08auto-regression\uff09\u7684\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>The system uses an autoregressive approach \u2014 generating images sequentially from left to right and top to bottom, similar to how text is written \u2014 rather than the diffusion model technique used by most image generators (like DALL-E) that create the entire image at once. Goh speculates that this technical difference could be what gives Images in ChatGPT better text rendering and binding capabilities.<br>\n\u8be5\u7cfb\u7edf\u91c7\u7528\u81ea\u56de\u5f52\u65b9\u6cd5\u2014\u2014\u6309\u987a\u5e8f\u4ece\u5de6\u5230\u53f3\u3001\u4ece\u4e0a\u5230\u4e0b\u751f\u6210\u56fe\u50cf\uff0c\u7c7b\u4f3c\u4e8e\u6587\u672c\u7684\u4e66\u5199\u65b9\u5f0f\u2014\u2014\u800c\u4e0d\u662f\u5927\u591a\u6570\u56fe\u50cf\u751f\u6210\u5668\uff08\u5982 DALL-E\uff09\u4f7f\u7528\u7684\u6269\u6563\u6a21\u578b\u6280\u672f\uff0c\u540e\u8005\u4e00\u6b21\u521b\u5efa\u6574\u4e2a\u56fe\u50cf\u3002Goh \u63a8\u6d4b\uff0c\u8fd9\u79cd\u6280\u672f\u5dee\u5f02\u53ef\u80fd\u662f ChatGPT \u4e2d\u7684\u56fe\u50cf\u5177\u6709\u66f4\u597d\u7684\u6587\u672c\u6e32\u67d3\u548c\u7ed1\u5b9a\u529f\u80fd\u7684\u539f\u56e0\u3002<\/p>\n<\/blockquote>\n\n\n\n<p>(\u51fa\u81ea\uff1a <a href=\"https:\/\/www.theverge.com\/openai\/635118\/chatgpt-sora-ai-image-generation-chatgpt\">https:\/\/www.theverge.com\/openai\/635118\/chatgpt-sora-ai-image-generation-chatgpt<\/a>)<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"683\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/hero_image_1-whiteboard1-1024x683.webp\" alt=\"\" class=\"wp-image-14359\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/hero_image_1-whiteboard1-1024x683.webp 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/hero_image_1-whiteboard1-300x200.webp 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/hero_image_1-whiteboard1-768x512.webp 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/hero_image_1-whiteboard1.webp 1536w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u53e6\u5916\u5728\u5b98\u65b9\u793a\u4f8b\u7684\u8fd9\u5f20\u677f\u4e66\u4e0a\u5199\u4e86\u201ccompose autoregressive prior with a powerful decoder\u201d\uff0c\u53f3\u4e0b\u89d2\u4e5f\u5199\u4e86tokens-&gt;[transformer]-&gt;[diffusion] pixels\u3002\u6240\u4ee5\u73b0\u5728\u793e\u533a\u4e0a\u6700\u4e3b\u6d41\u7684\u63a8\u6d4b\u662f\u4f7f\u7528\u4e86\u57fa\u4e8etransformer\u7684\u81ea\u56de\u5f52\u67b6\u6784\uff0c\u914d\u5408\u5f3a\u5927\u7684diffusion decoder\uff0c\u8fbe\u5230\u7684\u73b0\u5728\u7684\u6548\u679c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u81ea\u56de\u5f52\u4e0etokenization<\/h3>\n\n\n\n<p>\u9996\u5148\u89e3\u91ca\u4e00\u4e0b\u81ea\u56de\u5f52\uff0c\u81ea\u56de\u5f52\u88ab\u5927\u91cf\u7528\u4e8e\u6587\u5b57\u9886\u57df\uff0c\u56e0\u4e3a\u5b83\u8bf4\u767d\u4e86\u5c31\u662f\u4e00\u4e2a\u63a5\u9f99\u7684\u8fc7\u7a0b\u3002\u81ea\u56de\u5f52\u6a21\u578b\u628a\u751f\u6210\u5185\u5bb9\u770b\u6210\u4e00\u7cfb\u5217\u65e0\u6cd5\u7a77\u4e3e\u4f46\u662f\u6709\u56fa\u5b9a\u8303\u56f4\u7684\u5e8f\u5217\uff0c\u8fd9\u4e2a\u5e8f\u5217\u53ef\u4ee5\u662f\u56fa\u5b9a\u957f\u5ea6\u7684\u4e5f\u53ef\u4ee5\u662f\u4e0d\u56fa\u5b9a\u7684\uff0c\u6bcf\u6b21\u751f\u6210\u7684\u65f6\u5019\u4f1a\u8f93\u5165\u4e4b\u524d\u7684\u6240\u6709\u7ed3\u679c\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\u201c\u4f60\u662f\u8c01\uff1f\u201d\u6a21\u578b\u5728\u56de\u7b54\u7684\u65f6\u5019\u5c31\u4f1a\u5148\u9884\u6d4b\u7b2c\u4e00\u4e2a\u5b57\u201c\u6211\u201d\uff0c\u7136\u540e\u8f93\u5165\u53d8\u6210\u4e86\u201c\u4f60\u662f\u8c01\uff1f\u6211\u201d\uff0c\u6a21\u578b\u4f1a\u8ba4\u4e3a\u4e0b\u4e00\u4e2a\u5b57\u5927\u6982\u7387\u662f\u201c\u662f\u201d\uff0c\u4e4b\u540e\u518d\u7ed9\u6a21\u578b\u8f93\u5165\u201c\u4f60\u662f\u8c01\uff1f\u6211\u662f\u201d\u4ee5\u6b64\u7c7b\u63a8\uff0c\u4e00\u76f4\u5230\u6a21\u578b\u8ba4\u4e3a\u53ef\u4ee5\u7ed3\u675f\u4e86\uff0c\u5c31\u4f1a\u8f93\u51fa\u4e00\u4e2a[END]\u7684\u6807\u8bb0\u3002\u9047\u5230\u8fd9\u4e2a\u6807\u8bb0\u5c31\u53ef\u4ee5\u505c\u6b62\u7ee7\u7eed\u751f\u6210\u4e86\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e2a\u81ea\u56de\u5f52\u6a21\u578b\u88ab\u5927\u91cf\u7528\u5728\u6587\u5b57\u4e0a\uff0c\u56e0\u4e3a\u6587\u5b57\u975e\u5e38\u5929\u7136\u5730\u7b26\u5408\u81ea\u56de\u5f52\u751f\u6210\u7684\u903b\u8f91\u3002\u81ea\u56de\u5f52\u6a21\u578b\u7684\u7ecf\u5178\u67b6\u6784\u5c31\u662ftransformer\uff0c\u56e0\u4e3atransformer\u5bf9\u4e8e\u8f83\u957f\u7684\u8f93\u5165\u6709\u7740\u6bd4\u4e4b\u524d\u7684\u7b97\u6cd5\u66f4\u4f18\u7684\u8ba1\u7b97\u65b9\u5f0f\uff08\u6ce8\u610f\u529b\u673a\u5236\uff0c\u8fd9\u4e2a\u4ee5\u540e\u6709\u673a\u4f1a\u518d\u89e3\u91ca\uff09\uff0c\u6240\u4ee5\u5728transformer\u95ee\u4e16\u4e4b\u540e\u5728NLP\u9886\u57df\u560e\u560e\u4e71\u6740\u4e00\u76f4\u5230\u73b0\u5728\u5927\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u5f97\u5982\u6b64\u8fc5\u731b\u3002<\/p>\n\n\n\n<p>\u4f46\u662f\u4e3b\u6d41\u7684\u6587\u751f\u56fe\u6a21\u578b\u5e76\u4e0d\u662f\u57fa\u4e8e\u81ea\u56de\u5f52\u67b6\u6784\u7684\uff0c\u800c\u662f\u6269\u6563\u6a21\u578b\u3002\u6269\u6563\u6a21\u578b\u662f\u4e00\u79cd\u975e\u81ea\u56de\u5f52\u6a21\u578b\uff08Non Auto-Regressive Generation\uff0cNAR\uff09\uff0c\u5b83\u7684\u56fe\u7247\u662f\u4e00\u6b21\u6027\u751f\u6210\u5168\u56fe\u518d\u9010\u6e10\u7ec6\u5316\u3002<\/p>\n\n\n\n<p>\u5f53\u7136openAI\u65e9\u57282020\u5e74\u5c31\u5c1d\u8bd5\u8fc7\u4e86\u81ea\u56de\u5f52\u751f\u6210\u56fe\u7247\uff1a<br>\n<a href=\"https:\/\/openai.com\/index\/image-gpt\/\">https:\/\/openai.com\/index\/image-gpt\/<\/a> \u5c31\u662f\u9010\u884c\u751f\u6210\u56fe\u7247\u7684\uff0c\u53ea\u4e0d\u8fc7\u540e\u9762\u81ea\u56de\u5f52\u8fd9\u6761\u6280\u672f\u8def\u7ebf\u9010\u6e10\u88ab\u6269\u6563\u6a21\u578b\u8d85\u8d8a\u53d6\u4ee3\u3002\u81ea\u56de\u5f52\u65b9\u5f0f\u751f\u6210\u56fe\u7247\u9047\u5230\u7684\u7b2c\u4e00\u4e2a\u95ee\u9898\u5c31\u662f\uff0c\u56fe\u7247\u7684\u53ef\u80fd\u6027\u592a\u591a\u4e86\u3002<\/p>\n\n\n\n<p>\u81ea\u56de\u5f52\u91cc\u6709\u4e00\u4e2a\u91cd\u8981\u7684\u6982\u5ff5\u53ebtoken\uff0c\u5c31\u662fGPT\u91cc\u6309token\u6536\u8d39\u7684\u90a3\u4e2atoken\u3002\u6240\u6709\u7684tokens\u7ec4\u6210codebook\uff0c\u8fd9\u4e2acodebook\u4e5f\u4e0d\u80fd\u592a\u5927\uff0c\u8d8a\u5927\u8ba1\u7b97\u91cf\u5c31\u8d8a\u5927\uff08\u81ea\u56de\u5f52\u7684\u6bcf\u4e00\u6b21\u751f\u6210\u53ef\u4ee5\u8fd1\u4f3c\u770b\u6210\u5206\u7c7b\u95ee\u9898\uff0c\u800c\u8bb0\u5f55\u6240\u6709\u7c7b\u522b\u7684\u8868\u683c\u5c31\u662f\u8fd9\u4e2acodebook\uff09\u3002\u81ea\u56de\u5f52\u672c\u8d28\u4e0a\u662ftoken\u63a5\u9f99\uff0c\u6bd4\u5982\u5728\u4e2d\u6587\u91cc\uff0c\u6bcf\u4e2a\u4e2d\u6587\u6c49\u5b57\u53ef\u4ee5\u662f\u4e00\u4e2atoken\uff0c\u56e0\u4e3a\u6c49\u5b57\u7684\u6570\u91cf\u662f\u6709\u9650\u7684\u3002\u5728\u82f1\u6587\u91cc\uff0c\u82f1\u6587\u5355\u8bcd\u4e5f\u53ef\u4ee5\u662f\u4e00\u4e2atoken\uff0c\u4e0d\u8fc7\u5b9e\u9645\u4e0a\u662f\u7528\u8bcd\u6839\u4f5c\u4e3atoken\uff0c\u56e0\u4e3a\u82f1\u6587\u968f\u65f6\u4f1a\u51fa\u73b0\u65b0\u9020\u7684\u5355\u8bcd\uff0c\u8fd9\u4e9b\u5355\u8bcd\u4f9d\u7136\u53ef\u4ee5\u7528\u5df2\u6709\u7684\u8bcd\u6839\u7ec4\u6210\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u6587\u5b57\u662f\u5f88\u597dtokenized\u7684\uff0c\u8fd9\u4e5f\u662f\u81ea\u56de\u5f52\u5728\u8bed\u8a00\u6a21\u578b\u88ab\u5e7f\u6cdb\u4f7f\u7528\u7684\u539f\u56e0\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u662f\u4e00\u4e2atoken\u53ef\u89c6\u5316\u7684\u5728\u7ebf\u5de5\u5177\u7684\u5206\u6790\u7ed3\u679c\uff1a<a href=\"https:\/\/tiktoken.aigc2d.com\/\">https:\/\/tiktoken.aigc2d.com\/<\/a>   <\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"707\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404225157-1024x707.png\" alt=\"\" class=\"wp-image-14360\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404225157-1024x707.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404225157-300x207.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404225157-768x530.png 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404225157.png 1183w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u6240\u4ee5\u57fa\u4e8e\u81ea\u56de\u5f52\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff0c\u672c\u8d28\u4e0a\u4f9d\u7136\u4e5f\u662f\u8fd9\u4e9btoken\u63a5\u9f99\uff0c\u53ea\u4e0d\u8fc7\u591a\u6a21\u6001\u652f\u6301\u56fe\u50cf\uff0c\u58f0\u97f3\u7b49\u4f5c\u4e3a\u8f93\u5165\u3001\u8f93\u51fa\uff0c\u90a3\u5c31\u5fc5\u987b\u8981\u628a\u8fd9\u4e9b\u5a92\u4ecb\u4e5f\u8f6c\u6362\u6210\u53ef\u7a77\u4e3e\u7684tokens\u3002<\/p>\n\n\n\n<p>\u76f4\u63a5\u62ff\u50cf\u7d20\u70b9\u4f5c\u4e3atoken\u53ef\u884c\u5417\uff1f\u4e00\u4e2a\u50cf\u7d20\u6709RGB\u4e09\u4e2a\u901a\u9053\uff0c\u6bcf\u4e2a\u901a\u9053\u67090~255\u4e2a\u53d6\u503c\uff0c\u4e00\u5171\u5c31\u662f255^3 \u79cd\u53ef\u80fd\uff0c\u662f\u53ef\u4ee5\u7a77\u4e3e\u7684\u3002\u4f46\u662f\u8fd9\u6837\u7684\u8bdd\u751f\u6210\u901f\u5ea6\u592a\u6162\u4e86\u3002\u4e00\u5f201024* 1024\u7684\u56fe\u76f8\u5f53\u4e8e\u4e00\u6bb5\u4e00\u767e\u4e07\u5b57\u7684\u6587\u7ae0\uff08\u4e3a\u65b9\u4fbf\u7406\u89e3\u5047\u8bbe\u4e00\u4e2a\u5b57\u662f\u4e00\u4e2atoken\uff09\u3002\u4e8e\u662f\u5f88\u81ea\u7136\u5730\uff0c\u5927\u5bb6\u60f3\u5230\u4e86\u4f7f\u7528auto-encoder\u5bf9\u56fe\u50cf\u8fdb\u884c\u538b\u7f29\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404223852-1024x576.png\" alt=\"\" class=\"wp-image-14361\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404223852-1024x576.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404223852-300x169.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404223852-768x432.png 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404223852.png 1097w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u6709\u4e86auto-encoder\u4ee5\u540e\u5c31\u53ef\u4ee5\u628a\u538b\u7f29\u540e\u7684token\u62ff\u53bb\u8bad\u7ec3\u81ea\u56de\u5f52\u7f51\u7edc\uff0c\u7136\u540e\u8ba9\u81ea\u56de\u5f52\u7f51\u7edc\u505atoken\u63a5\u9f99\uff0c\u518d\u62ffdecoder\u628atoken\u8f6c\u6362\u56de\u56fe\u7247\u4e86\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"546\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404224024-1024x546.png\" alt=\"\" class=\"wp-image-14362\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404224024-1024x546.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404224024-300x160.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404224024-768x409.png 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404224024.png 1100w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\uff08\u56fe\u6e90\uff1a\u674e\u5b8f\u6bc5\u300a\u751f\u6210\u5f0fAI\u5bfc\u8bba 2024\u300b\uff09<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u56fe\u50cfTokenizer\u8fdb\u9636\u4e4b\u8def<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">VQ-VAE<\/h4>\n\n\n\n<p>\u63d0\u5230Auto-encoder\u4e00\u5b9a\u4f1a\u7b2c\u4e00\u4e2a\u60f3\u5230VAE\uff08\u4ee5\u524d\u5199\u7684\u4ecb\u7ecdVAE\u548cGAN\u7684\u6587\u7ae0\uff0c\u5efa\u8bae\u5148\u770b\u770b\uff1a<a href=\"https:\/\/blog.konata.vip\/?p=248\">\u300a\u4e3a\u4ec0\u4e48\u795e\u7ecf\u7f51\u7edc\u53ef\u4ee5\u521b\u9020\u672a\u5b58\u5728\u8fc7\u7684\u4eba\u8138\u300b<\/a> \uff09\u3002VAE\u5c31\u662f\u4e00\u4e2a\u57fa\u4e8e\u5377\u79ef\u7684\u5bf9\u79f0\u7684\u538b\u7f29-\u89e3\u7801\u7ed3\u6784\uff0c\u5b83\u53ef\u4ee5\u628a\u56fe\u7247\u538b\u7f29\u5230\u4e00\u4e2a\u975e\u5e38\u5c0f\u7684\u7a7a\u95f4\u4e0a\uff08latent space\uff09\u7136\u540e\u518d\u89e3\u538b\u7f29\u8fd8\u539f\u4e3a\u539f\u56fe\u3002\u77e5\u540d\u7684stable diffusion\uff08\u548c\u540e\u7eed\u7684\u7edd\u5927\u90e8\u5206\u4e3b\u6d41\u6587\u751f\u56fe\u6a21\u578b\uff09\u90fd\u6709\u4f7f\u7528VAE\u6765\u5148\u628a\u56fe\u7247\u8fdb\u884c\u538b\u7f29\uff0c\u518d\u5728latent\u7ef4\u5ea6\u4e0a\u505a\u53bb\u566a\u3002VAE\u7684\u6027\u80fd\u662f\u76f4\u63a5\u548c\u751f\u56fe\u6548\u679c\u6302\u94a9\u7684\uff0c\u6bd4\u5982\u4e4b\u524dSD\u5f88\u96be\u751f\u6210\u4f4e\u5206\u8fa8\u7387\u7684\u5c0f\u4eba\u8138\uff0c\u5982\u679c\u4f60\u76f4\u63a5\u628a\u56fe\u7247\u8fc7\u4e00\u904dVAE\u4f60\u4f1a\u53d1\u73b0\u53ea\u662f\u7528VAE\u590d\u539f\u56de\u53bb\u4ee5\u540e\u8138\u5c31\u7cca\u4e86\uff0c\u56e0\u4e3a\u4fe1\u606f\u88ab\u9ad8\u5ea6\u538b\u7f29\u4e86\uff0c\u800c\u4e0d\uff08\u53ea\uff09\u662f\u56e0\u4e3a\u751f\u56fe\u8fc7\u7a0b\u7684\u4e0d\u51c6\u786e\u9020\u6210\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"444\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232415-1024x444.png\" alt=\"\" class=\"wp-image-14363\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232415-1024x444.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232415-300x130.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232415-768x333.png 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232415.png 1198w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u81ea\u56de\u5f52\u548cdiffusion model\u4f7f\u7528\u7684VAE\u4e0d\u4e00\u6837\u7684\u5730\u65b9\u5728\u4e8e\uff0c\u81ea\u56de\u5f52\u7684codebook\u5927\u5c0f\u662f\u6709\u9650\u7684\uff0c\u800cVAE\u7684latent space\u662f\u8fde\u7eed\u7684\uff0c\u6240\u4ee5\u4f1a\u591a\u51fa\u4e00\u4e2a\u79bb\u6563\u5316\uff08\u91cf\u5316\uff09\u7684\u6b65\u9aa4\uff0c\u4e5f\u5c31\u662fVQ-VAE\u3002\u8fd9\u91cc\u7684VQ\u5c31\u662fvector quantization\uff0c\u5411\u91cf\u91cf\u5316\u3002\u5411\u91cf\u91cf\u5316\u53ef\u4ee5\u7c7b\u6bd4\u7406\u89e3\u4e3a\u4e00\u4e2ahash\u6216\u8005\u805a\u7c7b\u8fc7\u7a0b\uff0c\u628a\u8fde\u7eed\u7684\u503c\u4ee5\u7279\u5b9a\u65b9\u5f0f\u6620\u5c04\u5230codebook\u4e2d\u3002\u91cf\u5316\u7684\u672c\u8d28\u662f\u4e3a\u4e86\u8fdb\u4e00\u6b65\u538b\u7f29\u4fe1\u606f\u548c\u964d\u7ef4\uff0c\u540c\u65f6\u4e5f\u662f\u81ea\u56de\u5f52\u751f\u6210\u7684\u524d\u63d0\uff08\u9884\u6d4b\u5143\u7d20\u9700\u8981\u662f\u53ef\u7a77\u4e3e\u7684\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"516\" height=\"205\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232741.png\" alt=\"\" class=\"wp-image-14364\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232741.png 516w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250404232741-300x119.png 300w\" sizes=\"auto, (max-width: 516px) 100vw, 516px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u56fe\u6e90\uff1a<a href=\"https:\/\/www.breezedeus.com\/article\/image-tokenizers\">\u56fe\u50cf Token \u5316\uff1a\u89c6\u89c9\u6570\u636e\u8f6c\u6362\u7684\u5173\u952e\u6280\u672f<\/a><\/p>\n\n\n\n<p>\u4f46\u662f\u5982\u679c\u628a\u4e00\u5f20\u56fe\u7247\u7684latent vector\u5bf9\u5e94\u5230codebook\u91cc\u7684\u4e00\u4e2atoken\uff0c\u5f88\u660e\u663edecoder\u7684\u4fe1\u606f\u91cf\u662f\u4e0d\u591f\u7684\uff0c\u56e0\u4e3a\u628a\u4fe1\u606f\u538b\u7f29\u5f97\u592a\u6781\u9650\u4e86\u3002\u6240\u4ee5\u901a\u5e38\u4f1a\u628a\u56fe\u7247\u5207\u4e3a\u4e00\u4e2a\u4e2a\u5c0f\u683c\uff08patch\uff09\uff0c\u8fd9\u6837\u5c31\u591f\u7528\u4e86\uff0c\u8fd9\u4e2a\u64cd\u4f5c\u7684\u540d\u5b57\u53ebpatchify\u3002\u5bf9\u5e94\u7684\uff0c\u5982\u679c\u57fa\u4e8e\u81ea\u56de\u5f52\uff0c\u56fe\u7247\u4e5f\u4f1a\u4ee5\u4e00\u4e2a\u4e00\u4e2atoken\u6309\u987a\u5e8f\u751f\u6210\uff0c\u6700\u540e\u7ecf\u8fc7\u89e3\u7801\u7ec4\u6210\u4e00\u5f20\u5927\u56fe\uff0c\u8fd9\u4e2a\u64cd\u4f5c\u53eb\u505aUnpatchify\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">VQ-GAN<\/h4>\n\n\n\n<p>\u63a5\u4e0b\u6765\u7684\u8def\u7ebf\u641eCV\u7684\u5e94\u8be5\u5f88\u719f\u6089\u3002\u53ef\u4ee5\u770b\u51fa\u6765\u56fe\u50cftokenize\u6709\u4e24\u4e2a\u6700\u91cd\u8981\u7684\u6307\u6807\uff0c\u4e00\u4e2a\u662f\u56fe\u7247\u538b\u7f29\u5ea6\uff0c\u4e00\u4e2a\u662f\u56fe\u7247\u8fd8\u539f\u5ea6\u3002\u800c\u8fd8\u539f\u56fe\u7247\u7684\u8d28\u91cf\u548cDecoder\u7684\u80fd\u529b\u662f\u5f3a\u76f8\u5173\u7684\u3002VAE\u5728\u56fe\u50cf\u751f\u6210\u4e0a\u6709\u5404\u79cd\u786c\u4f24\uff0c\u4f8b\u5982\u56fe\u7247\u4f1a\u8d8b\u5411\u4e8e\u6a21\u7cca\u3002\u4e8e\u662f\u5f88\u81ea\u7136\u5730\uff0c\u5927\u5bb6\u5c31\u60f3\u5230\u589e\u52a0\u5bf9\u6297loss\u6765\u63d0\u5347\u6548\u679c\uff0c\u8fd9\u5c31\u662fVQ-GAN\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"493\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250407013419.png\" alt=\"\" class=\"wp-image-14366\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250407013419.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250407013419-300x144.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250407013419-768x370.png 768w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><br> \u56fe\u6e90\uff1a <a href=\"https:\/\/compvis.github.io\/taming-transformers\/\">Taming Transformers for High-Resolution Image Synthesis<\/a><\/p>\n\n\n\n<p>\u53ef\u4ee5\u770b\u5230\u4ece\u7ed3\u6784\u4e0a\u4e3b\u8981\u662f\u6bd4VQ-VAE\u591a\u4e86\u4e00\u4e2a\u5224\u522b\u5668\uff0c\u540c\u65f6\u8bad\u7ec3\u4e0a\u4e5f\u4f7f\u7528\u4e86GAN\u5e38\u7528\u7684\u611f\u77e5\u635f\u5931\u4ee3\u66ffVAE\u7684\u50cf\u7d20\u7ea7\u635f\u5931\uff08\u5982MSE\uff09\uff0c\u53ef\u4ee5\u7b80\u5355\u7406\u89e3\u4e3a\u524d\u8005\u5bf9\u6bd4\u7684\u662f\u751f\u6210\u6570\u636e\u548c\u8bad\u7ec3\u6570\u636e\u66f4\u9ad8\u5c42\u7ea7\u7684\u7279\u5f81\u5bf9\u5e94\u6027\uff08\u901a\u8fc7\u9884\u8bad\u7ec3\u7684\u795e\u7ecf\u7f51\u7edc\u63d0\u53d6\u7279\u5f81\uff09\uff0c\u540e\u8005\u5219\u662f\u7b80\u5355\u8ba1\u7b97\u4e24\u5f20\u56fe\u7247\u7684\u8ddd\u79bb\uff08\u6bcf\u4e2a\u50cf\u7d20\u4e4b\u95f4\u5dee\u591a\u5c11\uff09\u3002\u56fe\u7247tokenizer\u5e76\u4e0d\u662f\u81ea\u56de\u5f52\u7684\u4e13\u5c5e\uff0c\u540e\u7eed\u7684\u4e00\u4e9b\u6269\u6563\u6a21\u578b\u5982SD 3\u4e5f\u5f15\u5165\u4e86VQ-GAN\u6765\u63d0\u5347\u751f\u6210\u8d28\u91cf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">FlowMo<\/h4>\n\n\n\n<p>\u6309\u7167\u8fd9\u4e2a\u53d1\u5c55\u8109\u7edc\uff0c\u63a5\u4e0b\u6765\u5c31\u8be5\u8f6e\u5230diffusion model\u51fa\u573a\u4e86\u3002\u674e\u98de\u98de\u56e2\u961f\u51fa\u4e86\u4e00\u7bc7\u65b0\u7684paper\u53ebFlow to the Mode\uff08\u611f\u89c9\u5728\u73a9fly me to the moon\u7684\u6897\uff09\uff0c\u7b80\u79f0FlowMo\uff0c\u5c31\u662f\u91c7\u7528\u4e86\u76ee\u524d\u6700\u6d41\u884c\u7684MMDiT\uff08\u591a\u6a21\u6001transformer\u7ed3\u6784\u7684\u6269\u6563\u6a21\u578b\uff09\u548cRectified flow\uff08\u6bd4\u4f20\u7edf\u6269\u6563\u6a21\u578b\u66f4\u9ad8\u6548\u7684\u4e00\u79cd\u201c\u8d70\u76f4\u7ebf\u201d\u7684\u6c42\u89e3\u65b9\u6cd5\uff09\u7684\u65b9\u6cd5\u8bad\u7ec3Decoder\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"323\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250406175249-1024x323.png\" alt=\"\" class=\"wp-image-14365\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250406175249-1024x323.png 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250406175249-300x95.png 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250406175249-768x243.png 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/Pasted-image-20250406175249.png 1504w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p> \u56fe\u6e90\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2503.11056\">Flow to the Mode<\/a><\/p>\n\n\n\n<p>FlowMo\u7684Encoder\u548cDecoder\u67b6\u6784\u90fd\u662fMMDiT\uff0c\u4e0d\u542b\u4efb\u4f55\u5377\u79ef\u5c42\uff0c\u5176\u4e2dDecoder\u76f8\u6bd4\u4e4b\u524d\u7684\u67b6\u6784\uff0c\u8fd8\u591a\u4e00\u4e2at\u4f5c\u4e3a\u8f93\u5165\uff0c\u5bf9\u5e94\u6269\u6563\u6a21\u578b\u7684\u591a\u6b65\u6e32\u67d3\u3002FlowMo\u5728\u8bad\u7ec3\u65f6\u5206\u4e3a\u4e86\u4e24\u4e2a\u9636\u6bb5\uff0c\u7b2c\u4e00\u4e2a\u9636\u6bb5Decoder\u53ea\u505a\u4e00\u6b65\u53bb\u566a\uff0c\u8bad\u7ec3\u597dEncoder\u3001VQ\u3001Decoder\u540e\uff0c\u7b2c\u4e8c\u4e2a\u9636\u6bb5\u8fd8\u4f1a\u56fa\u5b9aEncoder\uff0c\u5355\u72ec\u5bf9Decoder\u57288\u6b65\u53bb\u566a\u4e0a\u8fdb\u4e00\u6b65\u8bad\u7ec3\uff0c\u4ee5\u63d0\u5347Decoder\u7684\u751f\u6210\u6548\u679c\u3002\u8fd9\u4e2a\u65b9\u6cd5\u4e5f\u5728\u56fe\u50cf\u538b\u7f29\u4e0a\u53d6\u5f97\u4e86\u76ee\u524d\u6700\u597d\u7684\u6548\u679c\uff0c\u4e5f\u5c31\u662f\u53ef\u4ee5\u5728\u6700\u5c11\u7684\u4fe1\u606f\u91cf\uff08BPP\uff0cBits per Pixel\uff09\u4e0a\u6700\u597d\u5730\u8fd8\u539f\u56fe\u7247\u3002\u5f53\u7136\u6211\u60f3\u8017\u65f6\u80af\u5b9a\u4e5f\u662f\u6bd4\u4e4b\u524d\u7684\u65b9\u6cd5\u6162\u5f88\u591a\u7684\uff0cGPT-4o\u7684\u56fe\u50cf\u751f\u6210\u786e\u5b9e\u662f\u6162\u4e00\u4e9b\uff0c\u4e0d\u8fc7\u76f8\u6bd4\u5e26\u6765\u7684\u597d\u5904\uff0c\u8fd9\u70b9\u8017\u65f6\u4e5f\u662f\u53ef\u4ee5\u63a5\u53d7\u7684\u3002\u800c\u4e14\u76f8\u6bd4\u4e8e\u8bad\u7ec3\u4e0a\u7684\u6781\u5927\u5f00\u9500\uff0c\u901a\u8fc7\u9002\u5f53\u5ef6\u957f\u63a8\u7406\u8017\u65f6\u6765\u63d0\u5347\u6548\u679c\u4e5f\u662f\u4e00\u4e2a\u6027\u4ef7\u6bd4\u6781\u9ad8\u7684\u64cd\u4f5c\uff0c\u4f8b\u5982Deepseek R1\u7684\u6df1\u5ea6\u601d\u8003\u6216\u8005\u8bf4Chain of thought\u3002<\/p>\n\n\n\n<p>\u6709\u4e86\u8fd9\u7c7b\u66f4\u5f3a\u529b\u7684\u56fe\u50cftokenization\u548c\u8bad\u7ec3\u7b56\u7565\u7684\u63d0\u5347\uff0c\u6700\u540e\u7684\u7ed3\u679c\u5c31\u662fGPT-4o\u4e00\u65e6\u8bad\u7ec3\u597d\uff0c\u5728\u56fe\u50cf\u751f\u6210\u3001\u63a7\u5236\u3001\u7f16\u8f91\u3001\u8bed\u4e49\u7406\u89e3\u7b49\u7ef4\u5ea6\u8d85\u51fa\u73b0\u6709\u65b9\u6cd5\u5f88\u591a\u5f88\u591a\uff0c\u5e76\u4e14\u662f\u4e00\u4e2a\u901a\u7528\u6a21\u578b\u800c\u4e0d\u662f\u4e13\u5bb6\u6a21\u578b\uff08\u53ea\u4e13\u7cbe\u4e8e\u7279\u5b9a\u4efb\u52a1\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u81ea\u56de\u5f52\uff1aNext-Token Prediction is All You Need<\/h3>\n\n\n\n<p>\u4ee5Emu3\u4e3a\u4f8b\uff0c\u6211\u4eec\u6765\u770b\u770b\u57fa\u4e8e\u81ea\u56de\u5f52\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u6709\u54ea\u4e9b\u7279\u6b8a\u4e4b\u5904\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">Token\u7edf\u4e00\u5316<\/h4>\n\n\n\n<p>Emu3\u662f\u4e00\u4e2a\u7eaf\u7cb9\u57fa\u4e8e\u81ea\u56de\u5f52\u9884\u6d4b\u4e0b\u4e00\u4e2atoken\uff0c\u628a\u56fe\u50cf\u3001\u6587\u672c\u548c\u89c6\u9891\u7f16\u7801\u4e3a\u540c\u4e00\u4e2a\u79bb\u6563\u7a7a\u95f4\u8bad\u7ec3\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002\u548c\u7eaf\u6587\u672cLLM\u76f8\u6bd4\uff0c\u5b83\u7684codebook\u9664\u4e86\u6587\u672ctoken\u8fd8\u5305\u542b\u4e86\u89c6\u89c9token\uff08\u57fa\u4e8eSBER-MoVQGAN\uff09\uff0c\u4e3a\u4e86\u7edf\u4e00\u683c\u5f0f\uff0c\u4f7f\u7528\u4e86\u5982\u4e0b\u7684\u7279\u6b8a\u6807\u8bb0\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;BOS]{\u6587\u672c}&#91;SOV]{\u5143\u6570\u636e}&#91;SOT]{\u89c6\u89c9tokens}&#91;EOV]&#91;EOS]  <\/code><\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>[BOS]\u548c[EOS]\u662f\u6587\u672c\u7684\u5f00\u59cb\u548c\u7ed3\u675f\u7b26\u53f7<\/li>\n\n\n\n<li>[SOV]\u548c[EOV]\u662f\u89c6\u89c9\u8f93\u5165\u7684\u5f00\u59cb\u548c\u7ed3\u675f\u7b26\u53f7<\/li>\n\n\n\n<li>[SOT]\u662f\u89c6\u89c9tokens\u7684\u5f00\u59cb\u7b26\u53f7<\/li>\n<\/ul>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u6837\u7684\u7ed3\u6784\uff0c\u7eaf\u6587\u672c\u6570\u636e\u548c\u56fe\u6587\u6570\u636e\u5bf9\uff08\u56fe\u7247+caption\uff09\u90fd\u53ef\u4ee5\u7edf\u4e00\u8f93\u5165\u7ed9\u6a21\u578b\u8bad\u7ec3\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u5927\u6a21\u578b\u8bad\u7ec3\u7b56\u7565<\/h4>\n\n\n\n<p>\u7136\u540e\u662f\u8bad\u7ec3\u7b56\u7565\uff0c\u8fd9\u90e8\u5206\u548c\u7eaf\u6587\u672cLLM\u6bd4\u8f83\u76f8\u4f3c\u3002\u8ba9\u6211\u4eec\u5148\u56de\u987e\u4e00\u4e0bChatGPT\u4e3a\u9996\u7684LLM\u8bad\u7ec3\u6d41\u7a0b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u9884\u8bad\u7ec3\u9636\u6bb5\uff1a\u4e5f\u5c31\u662f\u51c6\u5907\u6d77\u91cf\u7684\u6570\u636e\u8ba9\u6a21\u578b\u770b\uff0c\u63d0\u4f9b\u4e30\u5bcc\u7684\u524d\u7f6e\u77e5\u8bc6\u3002\u6b64\u65f6\u6a21\u578b\u53ea\u4f1a\u6a21\u4eff\u4eba\u7c7b\u7684\u6570\u636e\uff0c\u4f8b\u5982\u4f60\u95ee\u6a21\u578b\u201c\u4e16\u754c\u4e0a\u6700\u9ad8\u7684\u5c71\u5cf0\u662f\u4ec0\u4e48\uff1f\u201d\u6a21\u578b\u6709\u53ef\u80fd\u4f1a\u8f93\u51fa\uff1a\u201cA\uff0c\u963f\u5c14\u5351\u65af\u5c71\uff0cB\uff0c\u73e0\u7a46\u6717\u739b\u5cf0\u2026\u2026\uff09\u56e0\u4e3a\u5b83\u4e5f\u6709\u89c1\u8fc7\u8fd9\u6837\u7684\u6570\u636e\uff0c\u5728\u548c\u4eba\u7c7b\u9700\u6c42\u505a\u5bf9\u9f50\uff08alignment\uff09\u4e4b\u524d\uff0c\u5b83\u4e0d\u80fd\u7406\u89e3\u4f60\u8981\u5b83\u5e72\u4ec0\u4e48\uff0c\u53ea\u662f\u4e00\u5473\u6a21\u4eff\u3002<\/li>\n\n\n\n<li>\u76d1\u7763\u5fae\u8c03\u9636\u6bb5\uff1a\u4e5f\u5c31\u662f\u4f7f\u7528\u4e00\u4e9b\u9ad8\u8d28\u91cf\u7684\u4eba\u7c7b\u6807\u8bb0\u6570\u636e\uff0c\u8ba9\u6a21\u578b\u7684\u8f93\u51fa\u65b9\u5f0f\u66f4\u7b26\u5408\u6211\u4eec\u7684\u4f7f\u7528\u4e60\u60ef\uff0c\u5728ChatGPT\u91cc\uff0c\u5c31\u662f\u4f7f\u7528GPT 3\u5728\u4e00\u4e9b\u95ee\u7b54\u6570\u636e\u4e0a\u5fae\u8c03\u3002\u4f8b\u5982\u7ed9\u4ed6\u201d[\u7528\u6237]\uff1a\u201c\u4e16\u754c\u4e0a\u6700\u9ad8\u7684\u5c71\u5cf0\u662f\u4ec0\u4e48\uff1f[\u6a21\u578b]\uff1a\u73e0\u7a46\u6717\u739b\u5cf0\u3002\u201c\u8fd9\u4e9b\u6570\u636e\u4ee5\u540e\uff0c\u5b83\u53ef\u80fd\u5c31\u53ef\u4ee5\u7406\u89e3\u81ea\u5df1\u9700\u8981\u7ed9\u7528\u6237\u7b54\u6848\u3002<\/li>\n\n\n\n<li>\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\uff1a\u7ecf\u8fc7SFT\u4ee5\u540e\u6a21\u578b\u5df2\u7ecf\u53ef\u4ee5\u6709\u4e00\u4e2a\u80fd\u56de\u7b54\u4eba\u7c7b\u95ee\u9898\u7684\u8bed\u8a00\u6a21\u578b\u4e86\uff0c\u6b64\u65f6\u53ef\u4ee5\u8ba9\u4eba\u7c7b\u5bf9\u6a21\u578b\u7684\u7b54\u6848\u8fdb\u884c\u6253\u5206\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u751f\u6210\u66f4\u80fd\u8ba9\u4eba\u7c7b\u6ee1\u610f\u7684\u7b54\u6848\uff0c\u8fd9\u5c31\u662f\u5927\u540d\u9f0e\u9f0e\u7684RLHF\uff08Reinforcement Learning from Human Feedback\uff0c\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff09\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u5728Emu3\u4e5f\u6709\u7c7b\u4f3c\u7684\u51e0\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u9884\u8bad\u7ec3\u9636\u6bb5\uff1a\u51c6\u5907\u5927\u91cf\u7684\u6587\u672c\u3001\u56fe\u7247\u3001\u89c6\u9891\u8d44\u6599\u8fdb\u884c\u8bad\u7ec3\uff0c\u5176\u4e2d\u5148\u4f7f\u7528\u6587\u672c\u548c\u56fe\u7247\u8bad\u7ec3\uff0c\u518d\u52a0\u5165\u89c6\u9891\u6570\u636e\u3002<\/li>\n\n\n\n<li>\u8d28\u91cf\u5fae\u8c03\u9636\u6bb5\uff1a\u4f7f\u7528\u9ad8\u8d28\u91cf\u6570\u636e\u8fdb\u884c\u5fae\u8c03\uff0c\u5982\u66f4\u9ad8\u7b5b\u9009\u6807\u51c6\u548c\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u4f18\u8d28\u56fe\u7247\u3002<\/li>\n\n\n\n<li>DPO\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\uff1a\u4f7f\u7528SFT\u6a21\u578b\u751f\u6210\u591a\u4e2a\u7ed3\u679c\uff0c\u901a\u8fc7\u4eba\u7c7b\u6253\u5206\u4f5c\u4e3a\u53cd\u9988\u8fdb\u884c\u5f3a\u5316\u8bad\u7ec3\uff0c\u6253\u5206\u7684\u91cd\u70b9\u5728\u4e8e\u751f\u6210\u8d28\u91cf\u548c\u6587\u672c\u5bf9\u9f50\u7a0b\u5ea6\u3002 <\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u662fEmu3\u7684\u751f\u56fe\u6548\u679c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"1008\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504009ffe081e58294d1c01-1-1024x1008.webp\" alt=\"\" class=\"wp-image-14368\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504009ffe081e58294d1c01-1-1024x1008.webp 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504009ffe081e58294d1c01-1-300x295.webp 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504009ffe081e58294d1c01-1-768x756.webp 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504009ffe081e58294d1c01-1.webp 1290w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4ece\u8fd9\u4e2a\u8bba\u6587\u7684\u6548\u679c\u6765\u770b\uff0c\u81ea\u56de\u5f52\u751f\u6210\u56fe\u7247\u7684\u7ed3\u679c\u4e5f\u662f\u6f5c\u529b\u5f88\u5927\u7684\uff0c\u4e0d\u8fc7\u8fd9\u7bc7\u8bba\u6587\u6ca1\u6709\u5728\u56fe\u50cf\u7f16\u8f91\u3001\u98ce\u683c\u8f6c\u6362\u7b49\u56fe\u50cf\u9886\u57df\u5e38\u89c1\u7684\u4efb\u52a1\u4e0a\u505a\u8bad\u7ec3\u548c\u8bc4\u6d4b\uff0c\u731c\u60f3GPT-4o\u5e94\u8be5\u589e\u52a0\u4e86\u4e0d\u5c11\u8fd9\u7c7b\u4efb\u52a1\u7684SFT\u6570\u636e\u53bb\u505a\u5bf9\u9f50\u3002\u8fd9\u7c7b\u8bad\u7ec3\u6570\u636e\u53ef\u4ee5\u901a\u8fc7\u4ece\u4e00\u4e9b\u4e13\u5bb6\u6a21\u578b\u4e2d\u84b8\u998f\u3001\u4eba\u5de5PS\u7b49\u65b9\u5f0f\u5236\u9020\uff0c\u518d\u4f7f\u7528\u591a\u6a21\u6001\u89c6\u89c9\u7406\u89e3\u6a21\u578b\uff08\u76f8\u6bd4\u751f\u56fe\u6a21\u578b\u8f83\u4e3a\u6210\u719f\u7684\u4e00\u9879\u6280\u672f\uff09\u53bb\u6253\u5206\u7b5b\u9009\u6570\u636e\u3002\u5927\u81f4\u6d41\u7a0b\u53ef\u4ee5\u53c2\u8003\u8fd9\u7bc7<a href=\"https:\/\/arxiv.org\/pdf\/2411.07199\">OmniEdit<\/a>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"764\" src=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504007ecf408e5b90489901-1024x764.webp\" alt=\"\" class=\"wp-image-14369\" srcset=\"https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504007ecf408e5b90489901-1024x764.webp 1024w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504007ecf408e5b90489901-300x224.webp 300w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504007ecf408e5b90489901-768x573.webp 768w, https:\/\/blog.konata.vip\/wp-content\/uploads\/2025\/04\/000100022504007ecf408e5b90489901.webp 1207w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u8bed<\/h2>\n\n\n\n<p>\u81ea\u56de\u5f52\u505a\u56fe\u50cf\u751f\u6210\u5e76\u4e0d\u662f\u4e00\u4e2a\u5f88\u65b0\u9c9c\u7684\u4e8b\u60c5\uff0c\u4f46\u7531\u4e8e\u8bad\u7ec3\u56f0\u96be\u3001\u751f\u6210\u8f83\u6162\u88ab\u6269\u6563\u6a21\u578b\u53d6\u4ee3\u3002\u4f46\u662f\u6269\u6563\u6a21\u578b\u5f88\u96be\u50cf\u81ea\u56de\u5f52\u6a21\u578b\u8fd9\u6837\u505a\u4e00\u4e2a\u5f88\u5b8c\u6574\u7684\u901a\u7528\u6a21\u578b\uff0c\u6bcf\u4e2a\u4efb\u52a1\u90fd\u6709\u4e13\u95e8\u7684\u65b9\u6cd5\u53bb\u8fdb\u884c\u9002\u914d\uff0c\u5e76\u4e14\u56fe\u6587\u4e4b\u95f4\u7684\u4fe1\u606f\u4ea4\u4e92\u4e5f\u4f1a\u6709\u635f\u5931\u3002GPT 3.5\uff08\u7b2c\u4e00\u4ee3ChatGPT\uff09\u76f8\u6bd4\u4e8eGPT 3\u5728\u6a21\u578b\u67b6\u6784\u4e0a\u5e76\u6ca1\u6709\u975e\u5e38\u660e\u663e\u7684\u6539\u8fdb\uff0c\u4f46\u662f\u901a\u8fc7\u6781\u5927\u6269\u5145\u8bad\u7ec3\u96c6\u3001\u53c2\u6570\u91cf\uff0c\u589e\u52a0\u5927\u91cf\u7684\u4eba\u7c7b\u6807\u6ce8\u505aSFT\u548c\u5f3a\u5316\u5b66\u4e60\uff0c\u5728\u6027\u80fd\u4e0a\u5c31\u7acb\u523b\u628a\u6a21\u578b\u63d0\u5347\u5230\u975e\u5e38\u53ef\u7528\u7684\u9ad8\u5ea6\u3002\u6211\u60f3GPT-4o\u5e94\u8be5\u4e5f\u662f\u7c7b\u4f3c\u7684\u6280\u672f\u8def\u7ebf\uff0c\u76f8\u6bd4\u7b80\u5355\u6613\u61c2\u7684\u539f\u7406\uff0c\u6548\u679c\u7684\u63d0\u5347\u66f4\u91cd\u8981\u7684\u8fd8\u662f\u4f9d\u8d56\u80cc\u540e\u7684\u5de5\u7a0b\u548c\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u56fe\u6e90\u5c0f\u7ea2\u4e66@\u6076\u9b54\u5e7c\u5d3d GPT-4o \u7684\u795e\u5947\u9b54\u6cd5 \u6700\u8fd1GPT-4o\u751f\u56fe\u6a21\u578b\u6a2a\u7a7a\u51fa\u4e16\uff0c\u6548\u679c\u548c\u73a9\u6cd5\u4e0a\u90fd\u6709\u7a81\u7834\u6027\u7684\u8fdb\u5c55 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-14351","post","type-post","status-publish","format-standard","hentry","category-research"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/posts\/14351","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=14351"}],"version-history":[{"count":3,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/posts\/14351\/revisions"}],"predecessor-version":[{"id":14372,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=\/wp\/v2\/posts\/14351\/revisions\/14372"}],"wp:attachment":[{"href":"https:\/\/blog.konata.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=14351"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=14351"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.konata.vip\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=14351"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}