{"id":270,"date":"2025-12-25T14:51:04","date_gmt":"2025-12-25T06:51:04","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2025\/12\/25\/minimax%e6%b5%b7%e8%9e%ba%e8%a7%86%e9%a2%91%e5%9b%a2%e9%98%9f%e9%a6%96%e6%ac%a1%e5%bc%80%e6%ba%90%ef%bc%9atokenizer%e4%b9%9f%e5%85%b7%e5%a4%87%e6%98%8e%e7%a1%ae%e7%9a%84scaling-law\/"},"modified":"2025-12-25T14:51:04","modified_gmt":"2025-12-25T06:51:04","slug":"minimax%e6%b5%b7%e8%9e%ba%e8%a7%86%e9%a2%91%e5%9b%a2%e9%98%9f%e9%a6%96%e6%ac%a1%e5%bc%80%e6%ba%90%ef%bc%9atokenizer%e4%b9%9f%e5%85%b7%e5%a4%87%e6%98%8e%e7%a1%ae%e7%9a%84scaling-law","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2025\/12\/25\/minimax%e6%b5%b7%e8%9e%ba%e8%a7%86%e9%a2%91%e5%9b%a2%e9%98%9f%e9%a6%96%e6%ac%a1%e5%bc%80%e6%ba%90%ef%bc%9atokenizer%e4%b9%9f%e5%85%b7%e5%a4%87%e6%98%8e%e7%a1%ae%e7%9a%84scaling-law\/","title":{"rendered":"MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u9996\u6b21\u5f00\u6e90\uff1aTokenizer\u4e5f\u5177\u5907\u660e\u786e\u7684Scaling Law"},"content":{"rendered":"<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/02b69e872d6ee0738292e185fd049453.jpeg'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/02b69e872d6ee0738292e185fd049453.jpeg\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u9996\u6b21\u5f00\u6e90\uff1aTokenizer\u4e5f\u5177\u5907\u660e\u786e\u7684Scaling Law\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/bb1da2c5a1ca5c97148a1e7515fdc4db.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/bb1da2c5a1ca5c97148a1e7515fdc4db.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u9996\u6b21\u5f00\u6e90\uff1aTokenizer\u4e5f\u5177\u5907\u660e\u786e\u7684Scaling Law\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/2025ccce58cd1acfb2f54fc74b45f98c.jpeg'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/2025ccce58cd1acfb2f54fc74b45f98c.jpeg\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u9996\u6b21\u5f00\u6e90\uff1aTokenizer\u4e5f\u5177\u5907\u660e\u786e\u7684Scaling Law\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/a32d6a13f517385209d7449abfe06e59.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2025\/12\/a32d6a13f517385209d7449abfe06e59.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u9996\u6b21\u5f00\u6e90\uff1aTokenizer\u4e5f\u5177\u5907\u660e\u786e\u7684Scaling Law\"\/><\/div><\/figure>\n<p>MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u4e0d\u85cf\u4e86\uff01<\/p>\n<p>\u9996\u6b21\u5f00\u6e90\u5c31\u63ed\u6653\u4e86\u4e00\u4e2a\u56f0\u6270\u884c\u4e1a\u5df2\u4e45\u7684\u95ee\u9898\u7684\u7b54\u6848\u2014\u2014<\/p>\n<p>\u4e3a\u4ec0\u4e48\u5f80\u7b2c\u4e00\u9636\u6bb5\u7684\u89c6\u89c9\u5206\u8bcd\u5668\u91cc\u7838\u518d\u591a\u7b97\u529b\uff0c\u4e5f\u65e0\u6cd5\u63d0\u5347\u7b2c\u4e8c\u9636\u6bb5\u7684\u751f\u6210\u6548\u679c\uff1f<\/p>\n<p>\u7ffb\u8bd1\u6210\u5927\u767d\u8bdd\u5c31\u662f\uff0c\u867d\u7136\u56fe\u50cf\/\u89c6\u9891\u751f\u6210\u6a21\u578b\u7684\u53c2\u6570\u8d8a\u505a\u8d8a\u5927\u3001\u7b97\u529b\u8d8a\u5806\u8d8a\u731b\uff0c\u4f46\u7528\u6237\u5b9e\u9645\u4f53\u9a8c\u4e0b\u6765\u603b\u6709\u4e00\u79cd\u5fae\u5999\u7684\u611f\u53d7\u2014\u2014\u8fd9\u4e9b\u5e9e\u5927\u7684\u6295\u5165\u4e0e\u4ea7\u51fa\u4f3c\u4e4e\u4e0d\u6210\u6b63\u6bd4\uff0c\u6a21\u578b\u79bb\u5b8c\u5168\u771f\u6b63\u53ef\u7528\u603b\u662f\u5dee\u4e00\u6bb5\u8ddd\u79bb\u3002<\/p>\n<p>So why\uff1f\u95ee\u9898\uff0c\u5927\u6982\u7387\u5c31\u51fa\u5728\u89c6\u89c9\u5206\u8bcd\u5668\uff08Tokenizer\uff09\u8fd9\u4e2a\u4e1c\u897f\u8eab\u4e0a\u4e86\u3002<\/p>\n<p>\u5f53\u7b97\u529b\u4e0d\u518d\u662f\u7b54\u6848\u65f6\uff0c\u771f\u6b63\u9700\u8981\u88ab\u91cd\u65b0\u5ba1\u89c6\u7684\uff0c\u5176\u5b9e\u662f\u751f\u6210\u6a21\u578b\u7684\u201c\u8d77\u70b9\u201d\u3002<\/p>\n<p>\u5728\u5f53\u524d\u4e3b\u6d41\u7684\u4e24\u9636\u6bb5\u751f\u6210\u6846\u67b6\u4e2d\uff08\u5206\u8bcd\u5668+\u751f\u6210\u6a21\u578b\uff09\uff0c\u4e1a\u754c\u5df2\u7ecf\u5728\u89c6\u89c9\u5206\u8bcd\u5668\u7684\u9884\u8bad\u7ec3\u4e0a\u6295\u5165\u4e86\u5927\u91cf\u7b97\u529b\u4e0e\u6570\u636e\uff0c\u4f46\u4e00\u4e2a\u5c34\u5c2c\u7684\u4e8b\u5b9e\u662f\uff1a<\/p>\n<p>\u8fd9\u4e9b\u6210\u672c\uff0c\u51e0\u4e4e\u6ca1\u6709\u7ebf\u6027\u5730\u8f6c\u5316\u4e3a\u751f\u6210\u8d28\u91cf\u7684\u63d0\u5347\u3002<\/p>\n<p>\u800cMiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\uff0c\u4e0d\u6b62\u6311\u6218\u4e86\u8fd9\u4e00\u73b0\u5b9e\u2014\u2014\u7528\u5b9e\u9a8c\u8bc1\u660e\u201cTokenizer\u7684scaling\u80fd\u591f\u63d0\u5347\u6a21\u578b\u6027\u80fd\u201d\u3002<\/p>\n<p>\u66f4\u5173\u952e\u7684\u662f\uff0c\u8fd8\u5e26\u6765\u4e86\u4e00\u6b3e\u5f00\u7bb1\u5373\u7528\u3001\u4e13\u4e3a\u201c\u4e0b\u4e00\u4ee3\u751f\u6210\u6a21\u578b\u201d\u6253\u9020\u7684\u53ef\u6269\u5c55\u89c6\u89c9\u5206\u8bcd\u5668\u9884\u8bad\u7ec3\u6846\u67b6\u2014\u2014Visual Tokenizer Pre-training\uff08\u4ee5\u4e0b\u7b80\u79f0VTP\uff09\u3002<\/p>\n<p>\u53ea\u9700\u6362\u4e0a\u8fd9\u4e2a\u89c6\u89c9\u5206\u8bcd\u5668\uff0c\u5373\u53ef\u5728\u4e0d\u6539\u53d8\u4e0b\u6e38\u4e3b\u6a21\u578b\uff08\u5982DiT\uff09\u4efb\u4f55\u8bad\u7ec3\u914d\u7f6e\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u7aef\u5230\u7aef\u751f\u6210\u6027\u80fd\u7684\u500d\u6570\u7ea7\u63d0\u5347\u3002<\/p>\n<p>\u4e0b\u9762\u8be6\u7ec6\u6765\u770b\u2014\u2014<\/p>\n<p>\u8981\u7406\u89e3VTP\u5b58\u5728\u7684\u610f\u4e49\uff0c\u9996\u5148\u6211\u4eec\u9700\u8981\u56de\u5230\u66f4\u57fa\u7840\u7684\u95ee\u9898\u4e0a\u2014\u2014<\/p>\n<p>Tokenizer\u662f\u4ec0\u4e48\uff1f\u5b83\u53c8\u4e3a\u4f55\u5982\u6b64\u5173\u952e\uff1f<\/p>\n<p>\u4ee5AI\u751f\u56fe\u4e3a\u4f8b\uff0c\u5f53\u524d\u4e3b\u6d41\u7684\u751f\u56fe\u6a21\u578b\u51e0\u4e4e\u6e05\u4e00\u8272\u91c7\u7528\u4e86\u201c\u4e24\u9636\u6bb5\u201d\u751f\u6210\u6846\u67b6\uff1a<\/p>\n<p>\u7528\u4e00\u5f20\u56fe\u6765\u5f62\u5bb9\u8fd9\u4e2a\u8fc7\u7a0bbe like\uff08\u5148\u538b\u7f29\u540e\u8fd8\u539f\uff09\uff1a<\/p>\n<p>\u800cTokenizer\uff0c\u5c31\u662f\u5728\u7b2c\u4e00\u9636\u6bb5\u4e2d\u8d1f\u8d23\u538b\u7f29\u56fe\u7247\u4fe1\u606f\u7684\u7ec4\u4ef6\u3002<\/p>\n<p>\u5982\u679c\u6ca1\u6709\u8fd9\u4e2a\u7ec4\u4ef6\uff0c\u6269\u6563\u6a21\u578b\u5c31\u5f97\u76f4\u63a5\u5728\u6570\u767e\u4e07\u50cf\u7d20\u7ec4\u6210\u7684\u539f\u59cb\u50cf\u7d20\u7a7a\u95f4\u4e2d\u8fdb\u884c\u201c\u53bb\u566a\u201d\u548c\u751f\u6210\u2014\u2014\u5176\u8ba1\u7b97\u590d\u6742\u5ea6\u548c\u5185\u5b58\u5f00\u9500\u5c06\u662f\u5929\u6587\u6570\u5b57\u3002<\/p>\n<p>\u56e0\u6b64\uff0cTokenizer\u79f0\u5f97\u4e0a\u662f\u8ba9\u6a21\u578b\u80fd\u591f\u5728\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u8fd0\u884c\u8d77\u6765\u7684\u201c\u5173\u952e\u529f\u81e3\u201d\u3002<\/p>\n<p>\u9664\u6b64\u4e4b\u5916\uff0cTokenize\u8fc7\u7a0b\u6240\u5f97\u7684\u4f4e\u7ef4\u6f5c\u5728\u8868\u5f81\uff0c\u5176\u8d28\u91cf\u76f4\u63a5\u51b3\u5b9a\u4e86\u6700\u7ec8\u751f\u6210\u56fe\u50cf\u7684\u8bed\u4e49\u4fdd\u771f\u5ea6\u3001\u7ec6\u8282\u4e30\u5bcc\u5ea6\u4ee5\u53ca\u6574\u4f53\u771f\u5b9e\u6027\uff0c\u6240\u4ee5\u4e1a\u754c\u4e5f\u666e\u904d\u8ba4\u4e3a\u2014\u2014<\/p>\n<p>\u4ece\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u6765\u8bf4\uff0cTokenize\u51e0\u4e4e\u51b3\u5b9a\u4e86\u540e\u7eed\u751f\u6210\u6a21\u578b\u7684\u4e0a\u9650\u3002<\/p>\n<p>\u57fa\u4e8e\u6b64\uff0c\u4eba\u4eec\u5728\u7814\u7a76\u5982\u4f55\u63d0\u5347\u4e0b\u6e38\u751f\u6210\u6a21\u578b\u7684\u8d28\u91cf\u65f6\uff0c\u4e5f\u4e0d\u7ea6\u800c\u540c\u5730\u5c06\u91cd\u70b9\u653e\u5728\u4e86\u201c\u5982\u4f55\u7cbe\u51c6\u590d\u523b\u539f\u59cb\u50cf\u7d20\u201d\u4e0a\uff0c\u5373\u8981\u6c42\u6a21\u578b\u5c06\u538b\u7f29\u540e\u7684\u6f5c\u5728\u7279\u5f81\uff08latent\uff09\u5c3d\u53ef\u80fd\u7cbe\u51c6\u5730\u8fd8\u539f\u4e3a\u539f\u59cb\u56fe\u50cf\u3002<\/p>\n<p>\u800c\u95ee\u9898\uff0c\u4e5f\u6b63\u51fa\u5728\u8fd9\u91cc\u3002<\/p>\n<p>\u4f20\u7edfTokenizer\u5728\u8ffd\u6c42\u201c\u50cf\u7d20\u7ea7\u91cd\u5efa\u201d\u65f6\uff0c\u5f80\u5f80\u4f1a\u8fc7\u5ea6\u6c89\u8ff7\u4e8e\u6355\u6349\u56fe\u50cf\u7684\u5c40\u90e8\u7eb9\u7406\u3001\u566a\u70b9\u7b49\u4f4e\u5c42\u4fe1\u606f\uff0c\u4ece\u800c\u5ffd\u89c6\u4e86\u751f\u6210\u4efb\u52a1\u771f\u6b63\u4f9d\u8d56\u7684\u9ad8\u5c42\u8bed\u4e49\u4e0e\u6574\u4f53\u7ed3\u6784\u3002<\/p>\n<p>\u8fd9\u5c31\u50cf\u5b66\u6e23\u4e3a\u4e86\u5e94\u4ed8\u8003\u8bd5\uff0c\u53ea\u60f3\u7740\u6b7b\u8bb0\u786c\u80cc\u5404\u79cd\u8bd5\u9898\uff0c\u5374\u5b8c\u5168\u4e0d\u7406\u89e3\u9898\u76ee\u80cc\u540e\u7684\u77e5\u8bc6\u70b9\u4e0e\u903b\u8f91\u7ed3\u6784\u3002\u7ed3\u679c\u771f\u6b63\u5230\u4e86\u9700\u8981\u4e3e\u4e00\u53cd\u4e09\u3001\u89e3\u51b3\u65b0\u95ee\u9898\u7684\u65f6\u5019\uff0c\u4e00\u4e0b\u5c31\u50bb\u773c\u4e86\u3002<\/p>\n<p>\u5728\u8bba\u6587\u4e2d\uff0c\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u5c06\u8fd9\u4e00\u73b0\u8c61\u5b9a\u4e49\u4e3a\u201c\u9884\u8bad\u7ec3\u7f29\u653e\u95ee\u9898\u201d\uff08Pre-training Scaling Problem\uff09\uff1a<\/p>\n<p>\u66f4\u597d\u7684\u50cf\u7d20\u7ea7\u91cd\u5efa\u7cbe\u5ea6\u5e76\u4e0d\u80fd\u5e26\u6765\u66f4\u9ad8\u8d28\u91cf\u7684\u751f\u6210\u6548\u679c\u3002<\/p>\n<p>\u5982\u56fe\u6240\u793a\uff0c\u4f20\u7edf\u4ec5\u57fa\u4e8e\u201c\u91cd\u5efa\u201d\u76ee\u6807\u8bad\u7ec3\u7684Tokenizer\uff08\u7070\u7ebf\uff09\uff0c\u968f\u7740\u7b97\u529b\u6295\u5165\u589e\u52a0\uff0c\u6a21\u578b\u7684\u751f\u6210\u6027\u80fd\u4e0d\u4ec5\u6ca1\u6709\u63d0\u5347\uff0c\u53cd\u800c\u9010\u6e10\u8d8b\u4e8e\u9971\u548c\u751a\u81f3\u7565\u5fae\u4e0b\u964d\u3002\uff08ps\uff1agFID\u8861\u91cf\u751f\u6210\u8d28\u91cf\uff0c\u8d8a\u5c0f\u8d8a\u597d\uff09<\/p>\n<p>\u7531\u6b64\uff0c\u8bba\u6587\u5f97\u51fa\u7b2c\u4e00\u4e2a\u5173\u952e\u53d1\u73b0\uff1a\u91cd\u5efa\u505a\u5f97\u8d8a\u597d\uff0c\u751f\u6210\u53cd\u800c\u53ef\u80fd\u8d8a\u5dee\u3002\u6362\u8a00\u4e4b\uff0c\u4f20\u7edfVAE\u6240\u4f9d\u8d56\u7684\u7eaf\u91cd\u5efa\u8bad\u7ec3\u8303\u5f0f\uff0c\u5728\u9762\u5411\u751f\u6210\u4efb\u52a1\u65f6\u5b58\u5728\u6839\u672c\u6027\u5c40\u9650\u3002<\/p>\n<p>\u90a3\u4e48\uff0cVTP\u662f\u5982\u4f55\u7834\u5c40\u7684\u5462\uff1f<\/p>\n<p>\u4e8b\u5b9e\u4e0a\uff0cVTP\u5e76\u975e\u6a2a\u7a7a\u51fa\u4e16\uff0c\u800c\u662f\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u5728\u957f\u671f\u5b9e\u8df5\u4e2d\u53cd\u590d\u8bd5\u9519\u4e0e\u8ba4\u77e5\u8fed\u4ee3\u7684\u4ea7\u7269\u3002<\/p>\n<p>\u5728\u65e9\u671f\u5de5\u4f5cVA-VAE\u4e2d\uff0c\u56e2\u961f\u6ce8\u610f\u5230\uff1a<\/p>\n<p>\u6709\u4e9b\u201c\u538b\u7f29\u540e\u7684\u7279\u5f81\u201d\u80fd\u8ba9\u540e\u7eed\u7684\u6269\u6563\u6a21\u578b\u5b66\u5f97\u53c8\u5feb\u53c8\u597d\uff0c\u800c\u53e6\u4e00\u4e9b\u5219\u4e0d\u80fd\u3002<\/p>\n<p>\u57fa\u4e8e\u6b64\uff0c\u4ed6\u4eec\u63d0\u51fa\u4e86\u201c\u6613\u5b66\u6027\u201d\u8fd9\u4e00\u6982\u5ff5\u2014\u2014\u5373\u4e0d\u540c\u7684\u6f5c\u5728\u8868\u5f81\uff08latents\uff09\u5bf9\u4e8e\u4e0b\u6e38\u751f\u6210\u6a21\u578b\u7684\u5b66\u4e60\u96be\u5ea6\u622a\u7136\u4e0d\u540c\uff0c\u5e76\u671f\u671b\u901a\u8fc7\u4f18\u5316latent\u7684\u67d0\u4e9b\u201c\u53cb\u597d\u7279\u6027\u201d\uff08\u5982\u5747\u5300\u6027\u3001\u9891\u57df\u5206\u5e03\u7b49\uff09\u6765\u95f4\u63a5\u6539\u5584\u751f\u6210\u6548\u679c\u3002<\/p>\n<p>\u7136\u800c\u5b9e\u8df5\u53d1\u73b0\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u867d\u6709\u4e00\u5b9a\u6548\u679c\uff0c\u5374\u672a\u89e6\u53ca\u672c\u8d28\u3002\u6309\u56e2\u961f\u81ea\u5df1\u7684\u8bdd\u6765\u8bf4\u5c31\u662f\uff1a<\/p>\n<p>\u5c1d\u8bd5\u5c06\u6613\u5b66\u6027\u7b49\u4ef7\u4e3a\u53ef\u7cbe\u786e\u8ba1\u7b97\u7684\u5355\u4e00\u6307\u6807\uff08\u4f8b\u5982\u9891\u57df\u5206\u5e03\u3001\u5747\u5300\u6027\u3001scale\u4e0d\u53d8\u6027\u3001\u4f4e\u79e9\u6548\u5e94\u7b49\u7b49\uff09\u6709\u4e9b\u8fc7\u4e8e\u7406\u60f3\u4e86\uff0c\u8fd9\u4e9b\u6307\u6807\u53ef\u80fd\u548c\u6613\u5b66\u6027\u6709\u4e00\u5b9a\u5173\u8054\uff0c\u4f46\u8fdc\u4e0d\u662f\u5168\u90e8\u3002<\/p>\n<p>\u5c31\u662f\u8bf4\uff0c\u5f53\u4f18\u5316\u76ee\u6807\u88ab\u9650\u5236\u5728\u5c11\u6570\u53ef\u6d4b\u91cf\u6307\u6807\u4e0a\u65f6\uff0c\u6a21\u578b\u5f80\u5f80\u53ea\u662f\u5b66\u4f1a\u4e86\u5728\u7279\u5b9a\u6307\u6807\u4e0a\u201c\u5237\u5206\u201d\uff0c\u800c\u6ca1\u6709\u4ece\u6839\u672c\u4e0a\u5b66\u4f1a\u7406\u89e3\u56fe\u50cf\u7684\u7ed3\u6784\u4e0e\u8bed\u4e49\uff08\u672c\u8d28\u4e0a\u4ecd\u672a\u8131\u79bb\u201c\u91cd\u5efa\u201d\u8303\u7574\uff09\u3002<\/p>\n<p>\u800c\u7ecf\u8fc7\u6301\u7eed\u53cd\u601d\uff0c\u56e2\u961f\u9010\u6e10\u610f\u8bc6\u5230VA-VAE\u548c\u540e\u6765\u7684\u5f88\u591a\u5c1d\u8bd5\uff0c\u5176\u6df1\u5c42\u76ee\u6807\u5e76\u975e\u4ec5\u4ec5\u662f\u4f18\u5316\u51e0\u4e2a\u6307\u6807\uff0c\u800c\u662f\u5728\u8bd5\u56fe\u8ba9latents\u5177\u5907\u67d0\u4e9b\u66f4\u9ad8\u7ea7\u7684\u201c\u7ed3\u6784\u201d\u3002\u5176\u80cc\u540e\u903b\u8f91\u5728\u4e8e\uff1a<\/p>\n<p>\u5047\u5982\u67d0\u79cdlatents\u5bf9\u4e8e\u56fe\u50cf\u4e2d\u7684\u5b9e\u4f53\u3001\u4f4d\u7f6e\u5173\u7cfb\u5177\u6709\u9ad8\u5ea6\u7ed3\u6784\u5316\u7684\u8868\u8fbe\uff0c\u90a3\u4e48\u8fd9\u79cd\u7ed3\u6784\u5316\u8868\u8fbe\u5bf9\u4e8e\u4e0b\u6e38diffusion\u5efa\u6a21\u6765\u8bf4\u4e5f\u4f1a\u66f4\u7b80\u6d01\u3001\u66f4\u5bb9\u6613\uff0c\u8fd9\u4e9b\u65b9\u9762\u7684\u751f\u6210\u6548\u679c\u4e5f\u81ea\u7136\u4f1a\u66f4\u597d\u3002<\/p>\n<p>\u8fd9\u5c31\u597d\u6bd4\u53e6\u4e00\u79cd\u5c42\u9762\u4e0a\u7684\u201c\u5927\u9053\u81f3\u7b80\u201d\u2014\u2014\u5f53\u6a21\u578b\u771f\u6b63\u638c\u63e1\u4e86\u7269\u4f53\u3001\u5173\u7cfb\u4e0e\u8bed\u4e49\u8fd9\u4e9b\u89c6\u89c9\u4e16\u754c\u7684\u672c\u8d28\u89c4\u5f8b\u65f6\uff0c\u5b83\u80fd\u5e94\u5bf9\u7684\u4efb\u52a1\u81ea\u7136\u4e5f\u5c31\u8d8a\u591a\u4e86\uff08\u4ee5\u4e0d\u53d8\u5e94\u4e07\u53d8\uff09\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u4e0e\u5176\u4e3a\u201c\u6613\u5b66\u6027\u201d\u8bbe\u8ba1\u8106\u5f31\u7684\u4ee3\u7406\u6307\u6807\uff0c\u4e0d\u5982\u76f4\u63a5\u8ba9Tokenizer\u53bb\u5b66\u4e60\u4eba\u7c7b\u89c6\u89c9\u7cfb\u7edf\u672c\u8eab\u6240\u4f9d\u8d56\u7684\u90a3\u79cd\u9ad8\u5ea6\u7ed3\u6784\u5316\u3001\u8bed\u4e49\u5316\u7684\u901a\u7528\u8868\u5f81\u3002<\/p>\n<p>\u6bd5\u7adf\u50cfCLIP\u3001DINOv2\u8fd9\u6837\u7684\u73b0\u4ee3\u901a\u7528\u8868\u5f81\u5b66\u4e60\u65b9\u6cd5\uff0c\u6b63\u662f\u5728\u5927\u89c4\u6a21\u6570\u636e\u4e0a\u81ea\u52a8\u5316\u5730\u5b66\u4e60\u4e86\u8fd9\u79cd\u201c\u4eba\u7c7b\u5bf9\u9f50\u201d\u7684\u89c6\u89c9\u8868\u8fbe\uff0c\u624d\u5728\u591a\u79cd\u4e0b\u6e38\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u81f3\u6b64\uff0c\u56e2\u961f\u7684\u7814\u7a76\u91cd\u5fc3\u6b63\u5f0f\u4ece\u201c\u50cf\u7d20\u7ea7\u91cd\u5efa\u201d\u8f6c\u5411\u4e86\u5f3a\u8c03\u7406\u89e3\u529b\u7684\u201c\u901a\u7528\u8868\u5f81\u5b66\u4e60\u201d\u3002<\/p>\n<p>\u57fa\u4e8e\u8fd9\u4e00\u8ba4\u77e5\u8f6c\u53d8\uff0cVTP\u7684\u8bbe\u8ba1\u539f\u5219\u53d8\u5f97\u5f02\u5e38\u6e05\u6670\u3002\u56e2\u961f\u8868\u793a\uff1a<\/p>\n<p>\u6211\u4eec\u5e94\u8be5\u878d\u5408\u6240\u6709\u5df2\u77e5\u6709\u6548\u7684\u8868\u5f81\u5b66\u4e60\u65b9\u6cd5\u6765\u6784\u5efa\u4e00\u4e2aTokenizer\u3002<\/p>\n<p>\u5177\u4f53\u800c\u8a00\uff0cVTP\u8fd9\u6b21\u7387\u5148\u5b9e\u73b0\u4e86\u4ece\u53ea\u7ec3\u201c\u91cd\u5efa\u201d\u5230\u201c\u7406\u89e3\u3001\u91cd\u5efa\u3001\u751f\u6210\u201d\u7684\u8054\u5408\u4f18\u5316\u3002<\/p>\n<p>\u4e00\u662f\u901a\u8fc7\u56fe\u6587\u5bf9\u6bd4\u5b66\u4e60\uff0c\u5efa\u7acb\u9ad8\u5c42\u8bed\u4e49\u7406\u89e3\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u65f6\uff0c\u56e2\u961f\u4f1a\u7ed9\u6a21\u578b\u770b\u5927\u91cf\u56fe\u50cf-\u6587\u672c\u914d\u5bf9\u6570\u636e\uff08\u5982\u201c\u4e00\u53ea\u732b\u5728\u6c99\u53d1\u4e0a\u201d\u4e0e\u5bf9\u5e94\u56fe\u7247\uff09\uff0c\u7136\u540e\u8981\u6c42\u56fe\u7247\u7684\u7279\u5f81\u5411\u91cf\u548c\u6587\u5b57\u7684\u7279\u5f81\u5411\u91cf\u5c3d\u53ef\u80fd\u63a5\u8fd1\u3002<\/p>\n<p>\u8fd9\u548c\u4e4b\u524d\u7684CLIP\u601d\u8def\u4e00\u8109\u76f8\u627f\uff0c\u4f46\u76ee\u6807\u66f4\u805a\u7126\u3002CLIP\u8ffd\u6c42\u7684\u662f\u5e7f\u4e49\u7684\u56fe\u6587\u5339\u914d\u80fd\u529b\uff0c\u800cVTP\u5219\u8981\u6c42Tokenizer\u5728\u538b\u7f29\u56fe\u50cf\u4e3a\u7d27\u51d1\u7684\u6f5c\u5728\u7f16\u7801\u65f6\uff0c\u5fc5\u987b\u4fdd\u7559\u4e0e\u6587\u672c\u5bf9\u9f50\u7684\u8bed\u4e49\u7ed3\u6784\u3002<\/p>\n<p>\u8fd9\u610f\u5473\u7740\uff0c\u751f\u6210\u6a21\u578b\u540e\u7eed\u63a5\u6536\u5230\u7684\u6bcf\u4e00\u4e2a\u6f5c\u5728\u8868\u793a\uff0c\u672c\u8eab\u5c31\u5df2\u7ecf\u662f\u201c\u6709\u8bed\u4e49\u7684\u201d\u2014\u2014\u5b83\u4e0d\u53ea\u662f\u4e00\u5806\u50cf\u7d20\u7684\u538b\u7f29\uff0c\u800c\u662f\u643a\u5e26\u4e86\u201c\u732b\u201d\u3001\u201c\u6c99\u53d1\u201d\u7b49\u6982\u5ff5\u53ca\u5176\u5173\u7cfb\u7684\u7ed3\u6784\u5316\u8868\u8fbe\u3002\u548c\u4ee5\u5f80\u76f8\u6bd4\uff0c\u201c\u8bed\u4e49\u6ce8\u5165\u201d\u88ab\u63d0\u524d\u5230\u4e86\u538b\u7f29\u9636\u6bb5\u3002<\/p>\n<p>\u4e8c\u662f\u901a\u8fc7\u81ea\u76d1\u7763\u5b66\u4e60\uff0c\u638c\u63e1\u7a7a\u95f4\u7ed3\u6784\u4e0e\u89c6\u89c9\u5e38\u8bc6\u3002<\/p>\n<p>\u5728VTP\u6846\u67b6\u4e2d\uff0c\u81ea\u76d1\u7763\u5b66\u4e60\u6a21\u5757\u901a\u8fc7\u7ed3\u5408\u63a9\u7801\u56fe\u50cf\u5efa\u6a21\uff08MIM\uff09 \u548c\u81ea\u84b8\u998f\uff08DINO\u98ce\u683c\uff09\uff0c\u201c\u8feb\u4f7f\u201d\u6a21\u578b\u53bb\u7406\u89e3\u56fe\u7247\u7684\u7a7a\u95f4\u5173\u7cfb\u548c\u7269\u4f53\u7ed3\u6784\u3002<\/p>\n<p>\u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u201c\u903c\u7740\u201dTokenizer\u6765\u56de\u7b54\u4e0d\u540c\u95ee\u9898\uff1a<\/p>\n<p>\u8fd9\u4e00\u5957\u6d41\u7a0b\u8d70\u4e0b\u6765\uff0c\u6a21\u578b\u5bf9\u56fe\u50cf\u7684\u7406\u89e3\u4fbf\u4e0d\u518d\u505c\u7559\u5728\u50cf\u7d20\u5c42\u9762\uff0c\u800c\u662f\u6784\u5efa\u8d77\u4e86\u7ed3\u6784\u5316\u7684\u89c6\u89c9\u8ba4\u77e5\u3002<\/p>\n<p>\u4e09\u662f\u901a\u8fc7\u91cd\u5efa\uff0c\u4fdd\u7559\u5fc5\u8981\u7684\u5e95\u5c42\u89c6\u89c9\u7ec6\u8282\u3002<\/p>\n<p>VTP\u4f9d\u7136\u4fdd\u7559\u4e86\u91cd\u5efa\u76ee\u6807\uff0c\u4f46\u5b83\u7684\u5b9a\u4f4d\u53d1\u751f\u4e86\u53d8\u5316\u2014\u2014\u5176\u4f5c\u7528\u4e0d\u518d\u662f\u201c\u8d8a\u51c6\u8d8a\u597d\u201d\uff0c\u800c\u662f\u786e\u4fdd\u6f5c\u5728\u7a7a\u95f4\u4e2d\u4ecd\u7136\u4fdd\u6709\u751f\u6210\u6240\u9700\u7684\u7eb9\u7406\u3001\u989c\u8272\u4e0e\u8fb9\u7f18\u7b49\u7ec6\u8282\u4fe1\u606f\u3002<\/p>\n<p>\u8fd9\u4e09\u91cd\u76ee\u6807\u5e76\u975e\u5b64\u7acb\u8fdb\u884c\uff0c\u800c\u662f\u901a\u8fc7\u4e00\u4e2a\u7edf\u4e00\u7684\u591a\u4efb\u52a1\u635f\u5931\u51fd\u6570\u8fdb\u884c\u8054\u5408\u4f18\u5316\uff1a<\/p>\n<p>\u6700\u7ec8\uff0cVTP\u5f97\u5230\u7684\u4e0d\u662f\u4e00\u4e2a\u5728\u67d0\u4e2a\u6307\u6807\u4e0a\u201c\u7279\u522b\u4f18\u79c0\u201d\u7684\u538b\u7f29\u5668\uff0c\u800c\u662f\u4e00\u4e2a\u81ea\u5e26\u7406\u89e3\u80fd\u529b\u3001\u5bf9\u751f\u6210\u6a21\u578b\u975e\u5e38\u53cb\u597d\u7684\u89c6\u89c9Tokenizer\u3002<\/p>\n<p>\u8fd9\u91cc\u4e5f\u8865\u5145\u4e00\u4e2a\u7ec6\u8282\uff1a\u4e3a\u4ec0\u4e48\u4ed6\u4eec\u4e0d\u50cf\u5f88\u591a\u5de5\u4f5c\u4e00\u6837\u76f4\u63a5\u4f7f\u7528\u73b0\u6709\u7684CLIP\u6216DINOv2\u4f5c\u4e3aTokenizer\uff0c\u800c\u975e\u8981\u6295\u5165\u5de8\u5927\u7b97\u529b\u4ece\u5934\u9884\u8bad\u7ec3\u4e00\u4e2a\uff1f<\/p>\n<p>\u5176\u6838\u5fc3\u601d\u8003\u5728\u4e8e\u4e24\u70b9\uff08\u56e2\u961f\u539f\u8bdd\u5982\u4e0b\uff09\uff1a<\/p>\n<p>\u8868\u5f81\u771f\u7684\u5f88\u91cd\u8981\uff0c\u6211\u4eec\u60f3\u8981\u505a\u5230\u5f88\u6781\u81f4\u3002\u5728\u6211\u4eec\u7684\u89c6\u89d2\u4e0b\u8868\u5f81\u5305\u542b\u4e86\u81ea\u76d1\u7763\u3001\u5bf9\u6bd4\u5b66\u4e60\u3001\u751a\u81f3\u662f\u91cd\u5efa\uff08\u8fd9\u4e9b\u53ea\u662f\u5df2\u77e5\u6bd4\u8f83\u6210\u719f\u7684\u65b9\u6cd5\uff0c\u5b9e\u9645\u4e0a\u7406\u60f3\u7684\u8868\u5f81\u8fdc\u4e0d\u6b62\u8fd9\u4e9b\uff09\uff0c\u5e02\u9762\u4e0a\u6ca1\u6709\u4e00\u4e2a\u80fd\u591f\u5f88\u597d\u5730\u878d\u6c47\u8fd9\u4e9b\u65b9\u6cd5\u7684\u6a21\u578b\uff0c\u6211\u4eec\u9700\u8981\u81ea\u5df1\u8bad\u7ec3\u4e00\u4e2a\u3002<\/p>\n<p>\u57fa\u4e8e\u8868\u5f81\u7684Tokenizer\u65b9\u6848\u5177\u5907scaling\u7684\u6f5c\u529b\uff0c\u9884\u8bad\u7ec3\u662f\u6700\u5408\u7406\u7684\u5b9e\u73b0\u65b9\u5f0f\u3002\u5982\u679c\u76f4\u63a5\u5229\u7528\u5df2\u6709\u6a21\u578b\u84b8\u998f\u6216\u8005\u8fc1\u79fb\uff0c\u4f1a\u56e0\u4e3asetting\u8fc7\u4e8e\u590d\u6742\u800c\u7834\u574fscaling\u7684\u6027\u8d28\uff0c\u4e5f\u4f1a\u53d7\u9650\u4e8e\u5df2\u6709\u7684\u6a21\u578b\u89c4\u683c\u800c\u65e0\u6cd5\u505a\u5145\u5206\u7684\u8bba\u8bc1\u3002<\/p>\n<p>\u6240\u4ee5\uff0c\u9009\u62e9\u201c\u4ece\u5934\u5f00\u59cb\u201d\u7684VTP\uff0c\u5176\u5b9e\u9645\u8868\u73b0\u7a76\u7adf\u5982\u4f55\u5462\uff1f<\/p>\n<p>\u6982\u62ec\u800c\u8a00\uff0c\u56e2\u961f\u901a\u8fc7VTP\u5f97\u51fa\u4e86\u4e24\u5927\u5173\u952e\u53d1\u73b0\uff1a<\/p>\n<p>\u5148\u8bf4\u7b2c\u4e00\u70b9\u3002<\/p>\n<p>VTP\u7528\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5982\u679c\u53ea\u505a\u91cd\u5efa\u7684\u8bdd\uff0c\u6295\u5165\u8d44\u6e90\u8d8a\u591a\u751f\u6210\u53cd\u800c\u53ef\u80fd\u8d8a\u5dee\u3002<\/p>\n<p>\u4e0b\u56fe\u4e2d\uff0c\u4ee3\u8868\u91cd\u5efa\u80fd\u529b\u7684rFID\u4ece2.0\u964d\u81f30.5\uff0c\u91cd\u5efa\u6548\u679c\u53d8\u597d\uff1b\u4f46\u4ee3\u8868\u751f\u6210\u80fd\u529b\u7684gFID\u4ece55.04\u5347\u81f358.56\uff0c\u751f\u6210\u6548\u679c\u53d8\u5dee\u3002<\/p>\n<p>\u800c\u4e00\u65e6\u6ce8\u5165\u201c\u7406\u89e3\u529b\u201d\uff08\u5373\u5f15\u5165CLIP\u56fe\u6587\u5bf9\u6bd4\/SSL\u81ea\u76d1\u7763\u4efb\u52a1\uff09\uff0c\u6a21\u578b\u7684\u751f\u6210\u8d28\u91cf\uff08gFID\uff09\u4f1a\u968f\u7740\u7406\u89e3\u80fd\u529b\uff08Linear Probe\uff09\u7684\u63d0\u5347\u800c\u540c\u6b65\u53d8\u597d\u2014\u2014<\/p>\n<p>\u4e8c\u8005\u5448\u73b0\u660e\u663e\u7684\u6b63\u76f8\u5173\uff0c\u4e14\u8fd9\u79cd\u63d0\u5347\u4f1a\u968f\u7740\u8bad\u7ec3\u8ba1\u7b97\u91cf\u589e\u52a0\u6301\u7eed\u63a8\u8fdb\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7f3a\u5c11\u201c\u7406\u89e3\u529b\u201d\u7684\u4f20\u7edfAE Only\u65b9\u6848\uff0c\u5176\u751f\u6210\u8d28\u91cf\u548c\u7406\u89e3\u80fd\u529b\u5f88\u5feb\u9677\u5165\u505c\u6ede\u3002<\/p>\n<p>\u66f4\u8fdb\u4e00\u6b65\uff0c\u56e2\u961f\u5728\u540c\u7b49\u7b97\u529b\u4e0b\u5bf9\u6bd4\u4e86\u4e0d\u540c\u7ec4\u5408\uff0c\u8bc1\u660e\u201cCLIP+SSL+AE\u201d\u7684\u8054\u5408\u8bad\u7ec3\u65b9\u5f0f\u4e0a\u9650\u6700\u9ad8\uff0c\u751f\u6210\u4e0e\u7406\u89e3\u6307\u6807\u5747\u6700\u4f18\u3002<\/p>\n<p>\u57fa\u4e8e\u6b64\uff0c\u56e2\u961f\u8bad\u7ec3\u7684VTP\u5728\u7406\u89e3\u3001\u91cd\u5efa\u3001\u751f\u6210\u65b9\u9762\u5747\u4ea4\u51fa\u4e86\u4e0d\u9519\u7684\u7b54\u5377\u2014\u2014<\/p>\n<p>\u5728ImageNet\u4e0a\u7684\u96f6\u6837\u672c\u5206\u7c7b\u51c6\u786e\u7387\u8fbe\u523078.2%\uff0c\u8d85\u8fc7\u4e86\u539f\u7248CLIP\u768475.5%\uff0c\u5df2\u7ecf\u5177\u5907\u5f3a\u901a\u7528\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff1b\u5728\u91cd\u5efa\u80fd\u529b\u4e0a\u8d85\u8fc7\u4e86Stable Diffusion\u7684VAE\uff0c rFID\u4f4e\u81f30.36\uff1b\u5728\u751f\u6210\u80fd\u529b\u4e0a\u8d85\u8fc7\u4e86\u6b64\u524d\u7684\u6539\u8fdb\u65b9\u6cd5VA-VAE\uff0cgFID\u4f4e\u81f32.81\u3002<\/p>\n<p>\u66f4\u5173\u952e\u7684\u662f\uff0c\u5728\u8fbe\u5230\u76f8\u540c\u751f\u6210\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\uff0cVTP\u7684\u8bad\u7ec3\u6536\u655b\u901f\u5ea6\u6bd4LDM\u5feb5.7\u500d\u3001\u6bd4VA-VAE\u5feb4.1\u500d\uff0c\u5927\u5e45\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n<p>\u8fd9\u4e00\u7cfb\u5217\u8868\u73b0\u5171\u540c\u5370\u8bc1\u4e86\u56e2\u961f\u7684\u5224\u65ad\uff1aTokenizer\u7684\u201c\u8bed\u4e49\u7406\u89e3\u529b\u201d\u800c\u975e\u5355\u7eaf\u7684\u50cf\u7d20\u91cd\u5efa\u7cbe\u5ea6\uff0c\u624d\u662f\u9a71\u52a8\u751f\u6210\u6027\u80fd\u4e0e\u6548\u7387\u63d0\u5347\u7684\u6838\u5fc3\u52a8\u529b\u3002<\/p>\n<p>\u518d\u8bf4\u7b2c\u4e8c\u70b9\uff0c\u4e5f\u662f\u66f4\u5177\u7a81\u7834\u6027\u7684\u4e00\u70b9\u3002<\/p>\n<p>\u56e2\u961f\u5728\u5b9e\u9a8c\u4e2d\u53d1\u73b0\uff0cVTP\u9996\u6b21\u5c55\u793a\u4e86Tokenizer\u7684Scaling Law\uff0c\u5373\u751f\u6210\u6027\u80fd\u53ef\u4ee5\u968f\u9884\u8bad\u7ec3\u4e2d\u6295\u5165\u7684\u8ba1\u7b97\u91cf\u3001\u53c2\u6570\u91cf\u548c\u6570\u636e\u89c4\u6a21\u6709\u6548\u589e\u957f\u3002<\/p>\n<p>\u4ec5\u4ece\u7b97\u529b\u7ef4\u5ea6\u5bf9\u6bd4\uff0c\u5728\u4e0d\u6539\u52a8\u4e0b\u6e38DiT\u6807\u51c6\u8bad\u7ec3\u914d\u7f6e\u7684\u524d\u63d0\u4e0b\uff0c\u5355\u7eaf\u5c06Tokenizer\u7684\u9884\u8bad\u7ec3\u8ba1\u7b97\u91cf\u653e\u5927\uff0cVTP\u5c31\u80fd\u4e3a\u6700\u7ec8\u751f\u6210\u6548\u679c\u5e26\u676565.8%\u7684\u6027\u80fd\u63d0\u5347\uff0c\u4e14\u63d0\u5347\u66f2\u7ebf\u4ecd\u672a\u89e6\u9876\u3002<\/p>\n<p>\u53cd\u89c2\u4f20\u7edf\u81ea\u7f16\u7801\u5668\uff08AE\uff09\uff0c\u5176\u6027\u80fd\u5728\u4ec5\u6295\u5165\u7ea61\/10\u8ba1\u7b97\u91cf\u65f6\u4fbf\u5df2\u9971\u548c\uff0c\u4e14\u7ee7\u7eed\u589e\u52a0\u7b97\u529b\u4e0d\u4ec5\u6536\u76ca\u5fae\u4e4e\u5176\u5fae\uff0c\u751a\u81f3\u53ef\u80fd\u5bfc\u81f4\u751f\u6210\u8d28\u91cf\u5012\u9000\u3002<\/p>\n<p>\u4ee5\u4e0a\u53d1\u73b0\u8868\u660e\uff0c\u63a5\u4e0b\u6765\u9664\u4e86\u5728\u4e3b\u6a21\u578b\u4e0a\u6295\u5165\u66f4\u591a\u53c2\u6570\/\u7b97\u529b\/\u6570\u636e\u4e4b\u5916\uff0c\u8fd8\u53ef\u4ee5\u901a\u8fc7Tokenizer\u7684scaling\u6765\u63d0\u5347\u6574\u4e2a\u751f\u6210\u7cfb\u7edf\u7684\u6027\u80fd\u3002<\/p>\n<p>\u8fd9\u4e2a\u7ed3\u8bba\uff0c\u4e4d\u4e00\u770b\u53ef\u80fd\u4f1a\u8ba9\u4eba\u6709\u70b9\u8f6c\u4e0d\u8fc7\u5f2f\uff1a\u4ec0\u4e48\u65f6\u5019\uff0cTokenizer\u4e5f\u5f00\u59cb\u8c08scaling\u4e86\uff1f<\/p>\n<p>\u5728\u5927\u6a21\u578b\u8bed\u5883\u91cc\uff0c\u201cScaling Law\u201d\u51e0\u4e4e\u5929\u7136\u53ea\u5c5e\u4e8e\u4e3b\u6a21\u578b\u2014\u2014\u53c2\u6570\u66f4\u5927\u3001\u6570\u636e\u66f4\u591a\u3001\u7b97\u529b\u66f4\u731b\uff0c\u6027\u80fd\u5c31\u8be5\u7ee7\u7eed\u5f80\u4e0a\u8d70\u3002\u81f3\u4e8eTokenizer\uff0c\u5219\u957f\u671f\u88ab\u89c6\u4f5c\u4e00\u4e2a\u201c\u524d\u7f6e\u6a21\u5757\u201d\uff0c\u4e3b\u6253\u4e00\u4e2a\u591f\u7528\u5c31\u884c\uff0c\u505a\u5b8c\u91cd\u5efa\u4efb\u52a1\u5c31\u9000\u573a\u3002<\/p>\n<p>\u4f46VTP\u7684\u51fa\u73b0\u5374\u6539\u53d8\u4e86\u8fd9\u4e00\u73b0\u5b9e\uff0c\u56e2\u961f\u8868\u793a\uff1a<\/p>\n<p>VTP\u5728latents\u6613\u5b66\u6027\u548c\u901a\u7528\u8868\u5f81\u5b66\u4e60\u4e4b\u95f4\u5efa\u7acb\u8d77\u4e86\u660e\u786e\u5173\u8054\uff0c\u4ece\u800c\u7b2c\u4e00\u6b21\u5c06Tokenizer\u4f5c\u4e3ascaling\u7684\u4e3b\u89d2\uff0c\u5c55\u73b0\u51fa\u5168\u9762\u7684scaling\u66f2\u7ebf\u548c\u6269\u5c55\u65b9\u5411\u3002<\/p>\n<p>\u5c31\u662f\u8bf4\uff0c\u5982\u679cTokenizer\u53ea\u662f\u88ab\u7528\u6765\u7cbe\u51c6\u590d\u523b\u50cf\u7d20\uff0c\u90a3\u4e48\u65e0\u8bba\u600e\u4e48\u5806\u7b97\u529b\uff0c\u63d0\u5347\u90fd\u5c06\u5f88\u5feb\u89c1\u9876\uff1b\u800c\u4e00\u65e6Tokenizer\u5b66\u5230\u7684\u662f\u5177\u5907\u8bed\u4e49\u7ed3\u6784\u3001\u5bf9\u4e0b\u6e38\u66f4\u53cb\u597d\u7684\u901a\u7528\u8868\u5f81\uff0c\u4e8b\u60c5\u5c31\u5b8c\u5168\u4e0d\u4e00\u6837\u4e86\u3002<\/p>\n<p>\u6bd4\u5982\u5bf9\u6574\u4e2a\u884c\u4e1a\u6765\u8bf4\uff0c\u7531\u4e8eVTP\u5728Tokenizer\u5c42\u9762\u5c31\u5df2\u7ecf\u7edf\u4e00\u4e86\u8bed\u4e49\u5bf9\u9f50\u3001\u7ed3\u6784\u8ba4\u77e5\u548c\u7ec6\u8282\u8868\u8fbe\uff0c\u56e0\u6b64\u5176\u4ea7\u51fa\u7684\u89c6\u89c9\u8868\u5f81\u5929\u7136\u5177\u5907\u591a\u4efb\u52a1\u534f\u540c\u7684\u6f5c\u529b\u3002<\/p>\n<p>\u8fd9\u6709\u70b9\u50cf\u5148\u628a\u4e16\u754c\u7ffb\u8bd1\u6210\u4e86\u4e00\u79cd\u7edf\u4e00\u3001\u7ed3\u6784\u5316\u7684\u201c\u89c6\u89c9\u8bed\u8a00\u201d\u3002<\/p>\n<p>\u4e00\u65e6\u8fd9\u5957\u8bed\u8a00\u786e\u5b9a\u4e0b\u6765\uff0c\u4e0b\u6e38\u65e0\u8bba\u662f\u7406\u89e3\u4efb\u52a1\uff08\u5982\u5206\u7c7b\u3001\u68c0\u6d4b\uff09\uff0c\u8fd8\u662f\u751f\u6210\u4efb\u52a1\uff08\u5982\u56fe\u50cf\u5408\u6210\u3001\u7f16\u8f91\uff09\uff0c\u90fd\u4e0d\u518d\u9700\u8981\u5404\u81ea\u201c\u91cd\u65b0\u5b66\u600e\u4e48\u63cf\u8ff0\u8fd9\u4e2a\u4e16\u754c\u201d\uff0c\u800c\u53ea\u662f\u7ad9\u5728\u540c\u4e00\u5957\u5e95\u5c42\u8868\u8fbe\u4e4b\u4e0a\uff0c\u505a\u4e0d\u540c\u7684\u4e8b\u60c5\u3002<\/p>\n<p>\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u770b\uff0cVTP\u5929\u7136\u9002\u5408\u7528\u6765\u6784\u5efa\u201c\u7406\u89e3-\u751f\u6210\u7edf\u4e00\u6a21\u578b\u201d\u3002\u6b63\u5982\u56e2\u961f\u6240\u8a00\uff1a<\/p>\n<p>Tokenizer\u5c42\u9762\u7684\u7edf\u4e00\uff0c\u662f\u66f4\u672c\u8d28\u7684\u7edf\u4e00\u3002<\/p>\n<p>\u4e5f\u56e0\u6b64\uff0c\u6b64\u6b21VTP\u7684\u5f00\u6e90\u5c31\u4e0d\u5355\u5355\u53ea\u662f\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7ec4\u4ef6\u90a3\u4e48\u7b80\u5355\u4e86\u3002<\/p>\n<p>\u5176\u66f4\u5927\u7684\u4ef7\u503c\u6216\u8bb8\u5728\u4e8e\uff0c\u5b83\u4e3a\u6574\u4e2a\u884c\u4e1a\u63d0\u4f9b\u4e86\u4e00\u6761\u65b0\u7684\u3001\u4e14\u88ab\u5b9e\u9a8c\u8bc1\u660e\u884c\u4e4b\u6709\u6548\u7684\u8def\u5f84\u2014\u2014<\/p>\n<p>\u5728\u4e3b\u6a21\u578b\u4e4b\u5916\uff0cTokenizer\u540c\u6837\u662f\u4e00\u4e2a\u503c\u5f97\u957f\u671f\u6295\u5165\u3001\u4e14\u5177\u5907\u660e\u786escaling\u56de\u62a5\u7684\u6838\u5fc3\u73af\u8282\u3002<\/p>\n<p>\u76ee\u524d\uff0cVTP\u7684\u8bba\u6587\u548c\u6a21\u578b\u6743\u91cd\u5df2\u7ecf\u516c\u5f00\uff0c\u52a8\u624b\u80fd\u529b\u5f3a\u7684\u670b\u53cb\u4e5f\u53ef\u4ee5\u5c1d\u8bd5\u4f53\u9a8c\u4e0b\uff1a<\/p>\n<p>\u6362\u4e00\u4e2a\u89c6\u89c9Tokenizer\uff0c\u6a21\u578b\u6027\u80fd\u5c31\u80fd\u53d8\u5f97\u4e0d\u4e00\u6837\u7684feeling\uff08\u624b\u52a8\u72d7\u5934\uff09\u3002<\/p>\n<p>\u3010\u4f20\u9001\u95e8\u3011\u4ee3\u7801\uff1ahttps:\/\/github.com\/MiniMax-AI\/VTP\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/2512.13687v1\u6a21\u578b\uff1ahttps:\/\/huggingface.co\/collections\/MiniMaxAI\/vtp<\/p>\n<blockquote>\n<p><strong>\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n<p>\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n<p>\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<p>\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>MiniMax\u6d77\u87ba\u89c6\u9891\u56e2\u961f\u4e0d\u85cf\u4e86\uff01<\/p>\n<p>\u9996\u6b21\u5f00\u6e90\u5c31\u63ed\u6653\u4e86\u4e00\u4e2a\u56f0\u6270\u884c\u4e1a\u5df2\u4e45\u7684\u95ee\u9898\u7684\u7b54\u6848\u2014\u2014<\/p>\n<p>\u4e3a\u4ec0\u4e48\u5f80\u7b2c\u4e00\u9636\u6bb5\u7684\u89c6\u89c9\u5206\u8bcd\u5668\u91cc\u7838\u518d\u591a\u7b97\u529b\uff0c\u4e5f\u65e0\u6cd5\u63d0\u5347\u7b2c\u4e8c\u9636\u6bb5\u7684\u751f\u6210\u6548\u679c\uff1f<\/p>\n<p>\u7ffb\u8bd1\u6210\u5927\u767d\u8bdd\u5c31\u662f\uff0c\u867d\u7136\u56fe\u50cf\/\u89c6\u9891\u751f\u6210\u6a21\u578b\u7684\u53c2\u6570\u8d8a\u505a\u8d8a\u5927\u3001\u7b97\u529b\u8d8a\u5806\u8d8a\u731b\uff0c\u4f46\u7528\u6237\u5b9e\u9645\u4f53\u9a8c\u4e0b\u6765\u603b\u6709\u4e00\u79cd\u5fae\u5999\u7684\u611f\u53d7\u2014\u2014\u8fd9\u4e9b\u5e9e\u5927\u7684\u6295\u5165\u4e0e\u4ea7\u51fa\u4f3c\u4e4e\u4e0d\u6210\u6b63\u6bd4\uff0c\u6a21\u578b\u79bb\u5b8c\u5168\u771f\u6b63\u53ef\u7528\u603b\u662f\u5dee\u4e00\u6bb5\u8ddd\u79bb\u3002<\/p>\n<p>So why\uff1f\u95ee\u9898\uff0c\u5927\u6982\u7387\u5c31\u51fa\u5728\u89c6\u89c9\u5206\u8bcd\u5668\uff08Toke<\/p>\n","protected":false},"author":1,"featured_media":266,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[26,27],"class_list":["post-270","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-minimax","tag-27"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/270","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=270"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/270\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/266"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=270"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=270"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=270"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}