{"id":18087,"date":"2024-01-30T17:51:54","date_gmt":"2024-01-30T09:51:54","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=18087"},"modified":"2024-01-30T17:51:54","modified_gmt":"2024-01-30T09:51:54","slug":"%e6%8a%8a%e5%9b%be%e5%83%8f%e8%a7%86%e4%b8%ba%e5%a4%96%e8%af%ad%ef%bc%8c%e5%bf%ab%e6%89%8b%e3%80%81%e5%8c%97%e5%a4%a7%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%aa%b2%e7%be%8edalle-3","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/01\/30\/18087\/","title":{"rendered":"\u628a\u56fe\u50cf\u89c6\u4e3a\u5916\u8bed\uff0c\u5feb\u624b\u3001\u5317\u5927\u591a\u6a21\u6001\u5927\u6a21\u578b\u5ab2\u7f8eDALLE-3"},"content":{"rendered":"
\u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u7edf\u4e00\u56fe\u6587\u8868\u793a\uff0c\u5feb\u624b\u4e0e\u5317\u5927\u5408\u4f5c\u63d0\u51fa\u57fa\u5ea7\u6a21\u578b LaVIT \u5237\u699c\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u4efb\u52a1\u3002<\/div>\n
\n
\n

\u5f53\u524d\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5982 GPT\u3001LLaMA \u7b49\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u80fd\u591f\u7406\u89e3\u548c\u751f\u6210\u590d\u6742\u7684\u6587\u672c\u5185\u5bb9\u3002\u4f46\u4f60\u662f\u5426\u60f3\u8fc7\uff0c\u5982\u679c\u80fd\u591f\u5c06\u5927\u8bed\u8a00\u6a21\u578b\u8fd9\u5f3a\u5927\u7684\u7406\u89e3\u548c\u751f\u6210\u80fd\u529b\u8fc1\u79fb\u5230\u591a\u6a21\u6001\u6570\u636e\u4e0a\uff0c\u5c31\u53ef\u4ee5\u8f7b\u677e\u7406\u89e3\u6d77\u91cf\u7684\u56fe\u50cf\u4e0e\u89c6\u9891\uff0c\u5e76\u8f85\u52a9\u521b\u4f5c\u56fe\u6587\u5e76\u8302\u7684\u5185\u5bb9\u3002\u8fd1\u671f\uff0c\u6765\u81ea\u5feb\u624b\u548c\u5317\u5927\u5408\u4f5c\u7684\u6700\u65b0\u591a\u6a21\u6001\u5927\u6a21\u578b LaVIT, \u6b63\u5728\u8ba9\u8fd9\u4e2a\u60f3\u6cd5\u9010\u6b65\u53d8\u4e3a\u73b0\u5b9e\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

    \n
  • \u8bba\u6587\u6807\u9898\uff1aUnified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization<\/li>\n
  • \u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2309.04669<\/li>\n
  • \u4ee3\u7801\u6a21\u578b\u5730\u5740\uff1ahttps:\/\/github.com\/jy0205\/LaVIT<\/li>\n<\/ul>\n

    \u6a21\u578b\u603b\u89c8<\/h4>\n

    LaVIT \u4f5c\u4e3a\u4e00\u4e2a\u65b0\u578b\u7684\u901a\u7528\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\uff0c\u53ef\u4ee5\u50cf\u8bed\u8a00\u6a21\u578b\u90a3\u6837\uff0c\u65e2\u80fd\u591f\u7406\u89e3\u4e5f\u80fd\u751f\u6210\u89c6\u89c9\u5185\u5bb9\u3002LaVIT \u7ee7\u627f\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u6210\u529f\u7684\u8bad\u7ec3\u8303\u5f0f\uff0c\u5373\u4ee5\u81ea\u56de\u5f52\u7684\u65b9\u5f0f\u9884\u6d4b\u4e0b\u4e00\u4e2a\u56fe\u50cf\u6216\u6587\u672c token\u3002\u5728\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u5176\u53ef\u4ee5\u5145\u5f53\u4e00\u4e2a\u591a\u6a21\u6001\u901a\u7528\u63a5\u53e3\uff0c\u65e0\u9700\u8fdb\u4e00\u6b65\u7684\u5fae\u8c03\uff0c\u5c31\u53ef\u4ee5\u6267\u884c\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u3002\u4f8b\u5982\uff0cLaVIT \u5177\u6709\u4ee5\u4e0b\u7684\u80fd\u529b\uff1a<\/p>\n

    \u5b9e\u73b0\u9ad8\u8d28\u91cf\u6587\u672c\u5230\u56fe\u50cf\u7684\u751f\u6210\uff1aLaVIT \u80fd\u591f\u6839\u636e\u7ed9\u5b9a\u7684\u6587\u672c\u63d0\u793a\u751f\u6210\u9ad8\u8d28\u91cf\u3001\u591a\u79cd\u7eb5\u6a2a\u6bd4\u548c\u9ad8\u7f8e\u611f\u7684\u56fe\u50cf\u3002\u5176\u56fe\u50cf\u751f\u6210\u80fd\u529b\u4e0e\u6700\u5148\u8fdb\u7684\u56fe\u50cf\u751f\u6210\u6a21\u578b\uff08\u5982 Parti\u3001SDXL \u548c DALLE-3\uff09\u76f8\u5ab2\u7f8e\u3002<\/p>\n

    <\/p>\n

    \u6839\u636e\u591a\u6a21\u6001\u63d0\u793a\u8fdb\u884c\u56fe\u50cf\u751f\u6210\uff1a\u7531\u4e8e\u5728 LaVIT \u4e2d\uff0c\u56fe\u50cf\u548c\u6587\u672c\u90fd\u88ab\u7edf\u4e00\u8868\u793a\u4e3a\u79bb\u6563\u5316\u7684 token\uff0c\u56e0\u6b64\u5176\u53ef\u4ee5\u63a5\u53d7\u591a\u79cd\u6a21\u6001\u7ec4\u5408\uff08\u4f8b\u5982\u6587\u672c\u3001\u56fe\u50cf + \u6587\u672c\u3001\u56fe\u50cf + \u56fe\u50cf\uff09\u4f5c\u4e3a\u63d0\u793a\uff0c\u751f\u6210\u76f8\u5e94\u7684\u56fe\u50cf\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u4efb\u4f55\u5fae\u8c03\u3002<\/p>\n

    <\/p>\n

    \u7406\u89e3\u56fe\u50cf\u5185\u5bb9\u5e76\u56de\u7b54\u95ee\u9898\uff1a\u5728\u7ed9\u5b9a\u8f93\u5165\u56fe\u50cf\u7684\u60c5\u51b5\u4e0b\uff0cLaVIT \u80fd\u591f\u9605\u8bfb\u56fe\u50cf\u5185\u5bb9\u5e76\u7406\u89e3\u5176\u8bed\u4e49\u3002\u4f8b\u5982\uff0c\u6a21\u578b\u53ef\u4ee5\u4e3a\u8f93\u5165\u7684\u56fe\u50cf\u63d0\u4f9b caption \u5e76\u56de\u7b54\u76f8\u5e94\u7684\u95ee\u9898\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u65b9\u6cd5\u6982\u89c8<\/h4>\n

    LaVIT \u7684\u6a21\u578b\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5176\u6574\u4e2a\u4f18\u5316\u8fc7\u7a0b\u5305\u62ec\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n

    <\/p>\n

    \u56fe\uff1aLaVIT \u6a21\u578b\u7684\u6574\u4f53\u67b6\u6784<\/p>\n

    \u9636\u6bb5 1: \u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u5668<\/strong><\/p>\n

    \u4e3a\u4e86\u80fd\u591f\u50cf\u81ea\u7136\u8bed\u8a00\u4e00\u6837\u7406\u89e3\u548c\u751f\u6210\u89c6\u89c9\u5185\u5bb9\uff0cLaVIT \u5f15\u5165\u4e86\u4e00\u4e2a\u8bbe\u8ba1\u826f\u597d\u7684\u89c6\u89c9\u5206\u8bcd\u5668\uff0c\u7528\u4e8e\u5c06\u89c6\u89c9\u5185\u5bb9\uff08\u8fde\u7eed\u4fe1\u53f7\uff09\u8f6c\u6362\u4e3a\u50cf\u6587\u672c\u4e00\u6837\u7684 token \u5e8f\u5217\uff0c\u5c31\u50cf LLM \u80fd\u591f\u7406\u89e3\u7684\u5916\u8bed\u4e00\u6837\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u4e3a\u4e86\u5b9e\u73b0\u7edf\u4e00\u89c6\u89c9\u548c\u8bed\u8a00\u7684\u5efa\u6a21\uff0c\u8be5\u89c6\u89c9\u5206\u8bcd\u5668 (Tokenizer) \u5e94\u8be5\u5177\u6709\u4ee5\u4e0b\u4e24\u4e2a\u7279\u6027\uff1a<\/p>\n

      \n
    1. \u79bb\u6563\u5316<\/strong>\uff1a\u89c6\u89c9 token \u5e94\u8be5\u88ab\u8868\u793a\u4e3a\u50cf\u6587\u672c\u4e00\u6837\u7684\u79bb\u6563\u5316\u5f62\u5f0f\u3002\u8fd9\u6837\u5bf9\u4e8e\u4e24\u79cd\u6a21\u6001\u91c7\u7528\u7edf\u4e00\u7684\u8868\u793a\u5f62\u5f0f\uff0c\u6709\u5229\u4e8e LaVIT \u5728\u4e00\u4e2a\u7edf\u4e00\u7684\u81ea\u56de\u5f52\u751f\u6210\u5f0f\u8bad\u7ec3\u6846\u67b6\u4e0b\uff0c\u4f7f\u7528\u76f8\u540c\u7684\u5206\u7c7b\u635f\u5931\u8fdb\u884c\u591a\u6a21\u6001\u5efa\u6a21\u4f18\u5316\u3002<\/li>\n
    2. \u52a8\u6001<\/strong>\u5316<\/strong>\uff1a\u4e0e\u6587\u672c token \u4e0d\u540c\u7684\u662f\uff0c\u56fe\u50cf patch \u4e4b\u95f4\u6709\u7740\u663e\u8457\u7684\u76f8\u4e92\u4f9d\u8d56\u6027\uff0c\u8fd9\u4f7f\u5f97\u4ece\u5176\u4ed6\u56fe\u50cf patch \u4e2d\u63a8\u65ad\u53e6\u4e00\u4e2a patch \u76f8\u5bf9\u7b80\u5355\u3002\u56e0\u6b64\uff0c\u8fd9\u79cd\u4f9d\u8d56\u6027\u4f1a\u964d\u4f4e\u539f\u672c LLM \u7684 next-token prediction \u4f18\u5316\u76ee\u6807\u7684\u6709\u6548\u6027\u3002LaVIT \u63d0\u51fa\u901a\u8fc7\u4f7f\u7528 token merging \u6765\u964d\u4f4e\u89c6\u89c9 patch \u4e4b\u95f4\u7684\u5197\u4f59\u6027\uff0c\u5176\u6839\u636e\u4e0d\u540c\u56fe\u50cf\u8bed\u4e49\u590d\u6742\u5ea6\u7684\u4e0d\u540c\uff0c\u7f16\u7801\u51fa\u52a8\u6001\u7684\u89c6\u89c9 token \u6570\u91cf\u3002\u8fd9\u6837\u5bf9\u4e8e\u590d\u6742\u7a0b\u5ea6\u4e0d\u540c\u7684\u56fe\u50cf\uff0c\u91c7\u7528\u52a8\u6001\u7684 token \u7f16\u7801\u4e5f\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u9884\u8bad\u7ec3\u7684\u6548\u7387\uff0c\u907f\u514d\u4e86\u5197\u4f59\u7684 token \u8ba1\u7b97\u3002<\/li>\n<\/ol>\n

      \u4e0b\u56fe\u662f LaVIT \u6240\u63d0\u51fa\u7684\u89c6\u89c9\u5206\u8bcd\u5668\u7ed3\u6784\uff1a<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u56fe\uff1a(a) \u52a8\u6001\u89c6\u89c9 token \u751f\u6210\u5668 (b) token \u5408\u5e76\u5668<\/p>\n

      \u8be5\u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u5668\u5305\u62ec token \u9009\u62e9\u5668\u548c token \u5408\u5e76\u5668\u3002\u5982\u56fe\u6240\u793a\uff0c token \u9009\u62e9\u5668\u7528\u6765\u9009\u62e9\u6700\u5177\u4fe1\u606f\u7684\u56fe\u50cf\u533a\u5757\uff0c\u800c token \u5408\u5e76\u5668\u5219\u5c06\u90a3\u4e9b uninformative \u7684\u89c6\u89c9\u5757\u7684\u4fe1\u606f\u538b\u7f29\u5230\u4fdd\u7559\u4e0b\u7684 token \u4e0a\uff0c\u5b9e\u73b0\u5bf9\u5197\u4f59 token \u7684 merging\u3002\u6574\u4e2a\u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u5668\u5219\u901a\u8fc7\u6700\u5927\u9650\u5ea6\u5730\u91cd\u6784\u8f93\u5165\u56fe\u50cf\u7684\u8bed\u4e49\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n

      Token \u9009\u62e9\u5668<\/strong><\/p>\n

      Token \u9009\u62e9\u5668\u63a5\u6536 N \u4e2a\u56fe\u50cf\u533a\u5757\u7ea7\u7684\u7279\u5f81\u4f5c\u4e3a\u8f93\u5165\uff0c\u5176\u76ee\u6807\u662f\u8bc4\u4f30\u6bcf\u4e2a\u56fe\u50cf\u533a\u5757\u7684\u91cd\u8981\u6027\u5e76\u9009\u62e9\u4fe1\u606f\u91cf\u6700\u9ad8\u7684\u533a\u5757\uff0c\u4ee5\u5145\u5206\u4ee3\u8868\u6574\u4e2a\u56fe\u50cf\u7684\u8bed\u4e49\u3002\u4e3a\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u91c7\u7528\u8f7b\u91cf\u7ea7\u6a21\u5757\uff0c\u7531\u591a\u4e2a MLP \u5c42\u7ec4\u6210\uff0c\u7528\u4e8e\u9884\u6d4b\u5206\u5e03 \u03c0\u3002\u901a\u8fc7\u4ece\u5206\u5e03 \u03c0 \u4e2d\u91c7\u6837\uff0c\u751f\u6210\u4e00\u4e2a\u4e8c\u8fdb\u5236\u51b3\u7b56 mask\uff0c\u7528\u4e8e\u6307\u793a\u662f\u5426\u4fdd\u7559\u76f8\u5e94\u7684\u56fe\u50cf\u533a\u5757\u3002<\/p>\n

      Token \u5408\u5e76\u5668<\/strong><\/p>\n

      Token \u5408\u5e76\u5668\u636e\u751f\u6210\u7684\u51b3\u7b56\u63a9\u7801\uff0c\u5c06 N \u4e2a\u56fe\u50cf\u533a\u5757\u5212\u5206\u4e3a\u4fdd\u7559 X_r \u548c\u820d\u5f03 X_d \u4e24\u7ec4\u3002\u4e0e\u76f4\u63a5\u4e22\u5f03 X_d \u4e0d\u540c\uff0ctoken \u5408\u5e76\u5668\u53ef\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u4fdd\u7559\u8f93\u5165\u56fe\u50cf\u7684\u8be6\u7ec6\u8bed\u4e49\u3002token \u5408\u5e76\u5668\u7531 L \u4e2a\u5806\u53e0\u7684\u5757\u7ec4\u6210\uff0c\u6bcf\u4e2a\u5757\u5305\u62ec\u56e0\u679c\u81ea\u6ce8\u610f\u529b\u5c42\u3001\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u5c42\u3002\u56e0\u679c\u81ea\u6ce8\u610f\u529b\u5c42\u4e2d\uff0c X_r \u4e2d\u7684\u6bcf\u4e2a token \u53ea\u5173\u6ce8\u5176\u524d\u9762\u7684 token\uff0c\u4ee5\u786e\u4fdd\u4e0e LLM \u4e2d\u7684\u6587\u672c token \u5f62\u5f0f\u4e00\u81f4\u3002\u4e0e\u53cc\u5411\u81ea\u6ce8\u610f\u76f8\u6bd4\uff0c\u8fd9\u79cd\u7b56\u7565\u8868\u73b0\u66f4\u597d\u3002\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5c06\u4fdd\u7559\u7684 token X_r \u4f5c\u4e3a query\uff0c\u5e76\u6839\u636e\u5b83\u4eec\u5728\u8bed\u4e49\u4e0a\u7684\u76f8\u4f3c\u6027\u5408\u5e76 X_d \u4e2d\u7684 token\u3002<\/p>\n

      \u9636\u6bb5 2: \u7edf\u4e00\u7684\u751f\u6210\u5f0f\u9884\u8bad\u7ec3<\/strong><\/p>\n

      \u7ecf\u8fc7\u89c6\u89c9\u5206\u8bcd\u5668\u5904\u7406\u540e\u7684\u89c6\u89c9 token \u4e0e\u6587\u672c token \u76f8\u8fde\u63a5\u5f62\u6210\u591a\u6a21\u6001\u5e8f\u5217\u4f5c\u4e3a\u8bad\u7ec3\u65f6\u7684\u8f93\u5165\u3002\u4e3a\u4e86\u533a\u5206\u4e24\u79cd\u6a21\u6001\uff0c\u4f5c\u8005\u5728\u56fe\u50cf token \u5e8f\u5217\u7684\u5f00\u5934\u548c\u7ed3\u5c3e\u63d2\u5165\u4e86\u7279\u6b8a token \uff1a[IMG] \u548c [\/IMG]\uff0c\u7528\u4e8e\u8868\u793a\u89c6\u89c9\u5185\u5bb9\u7684\u5f00\u59cb\u548c\u7ed3\u675f\u3002\u4e3a\u4e86\u80fd\u591f\u751f\u6210\u6587\u672c\u548c\u56fe\u50cf\uff0cLaVIT \u91c7\u7528\u4e24\u79cd\u56fe\u6587\u8fde\u63a5\u5f62\u5f0f\uff1a[image, text] \u548c [text; image]\u3002<\/p>\n

      \u5bf9\u4e8e\u8fd9\u4e9b\u591a\u6a21\u6001\u8f93\u5165\u5e8f\u5217\uff0cLaVIT \u91c7\u7528\u7edf\u4e00\u7684\u3001\u81ea\u56de\u5f52\u65b9\u5f0f\u6765\u76f4\u63a5\u6700\u5927\u5316\u6bcf\u4e2a\u591a\u6a21\u6001\u5e8f\u5217\u7684\u4f3c\u7136\u6027\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u8fd9\u6837\u5728\u8868\u793a\u7a7a\u95f4\u548c\u8bad\u7ec3\u65b9\u5f0f\u4e0a\u7684\u5b8c\u5168\u7edf\u4e00\uff0c\u6709\u52a9\u4e8e LLM \u66f4\u597d\u5730\u5b66\u4e60\u591a\u6a21\u6001\u4ea4\u4e92\u548c\u5bf9\u9f50\u3002\u5728\u9884\u8bad\u7ec3\u5b8c\u6210\u540e\uff0cLaVIT \u5177\u6709\u611f\u77e5\u56fe\u50cf\u7684\u80fd\u529b\uff0c\u53ef\u4ee5\u50cf\u5904\u7406\u6587\u672c\u4e00\u6837\u7406\u89e3\u548c\u751f\u6210\u56fe\u50cf\u3002<\/p>\n

      \u5b9e\u9a8c<\/h4>\n

      \u96f6\u6837\u672c\u591a\u6a21\u6001\u7406\u89e3<\/strong><\/p>\n

      LaVIT \u5728\u56fe\u50cf\u5b57\u5e55\u751f\u6210\uff08NoCaps\u3001Flickr30k\uff09\u548c\u89c6\u89c9\u95ee\u7b54\uff08VQAv2\u3001OKVQA\u3001GQA\u3001VizWiz\uff09\u7b49\u96f6\u6837\u672c\u591a\u6a21\u6001\u7406\u89e3\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u9886\u5148\u7684\u6027\u80fd\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u8868 1 \u96f6\u6837\u672c\u7684\u591a\u6a21\u6001\u7406\u89e3\u4efb\u52a1\u8bc4\u4f30<\/p>\n

      \u96f6\u6837\u672c\u591a\u6a21\u6001\u751f\u6210<\/strong><\/p>\n

      \u5728\u8fd9\u4e2a\u5b9e\u9a8c\u4e2d\uff0c\u7531\u4e8e\u6240\u63d0\u51fa\u7684\u89c6\u89c9 tokenizer \u80fd\u591f\u5c06\u56fe\u50cf\u8868\u793a\u4e3a\u79bb\u6563\u5316 token\uff0cLaVIT \u5177\u6709\u901a\u8fc7\u81ea\u56de\u5f52\u751f\u6210\u7c7b\u4f3c\u6587\u672c\u7684\u89c6\u89c9 token \u6765\u5408\u6210\u56fe\u50cf\u7684\u80fd\u529b\u3002\u4f5c\u8005\u5bf9\u6a21\u578b\u8fdb\u884c\u4e86\u96f6\u6837\u672c\u6587\u672c\u6761\u4ef6\u4e0b\u7684\u56fe\u50cf\u5408\u6210\u6027\u80fd\u7684\u5b9a\u91cf\u8bc4\u4f30\uff0c\u6bd4\u8f83\u7ed3\u679c\u5982\u8868 2 \u6240\u793a\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u8868 2 \u4e0d\u540c\u6a21\u578b\u7684\u96f6\u6837\u672c\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u6027\u80fd<\/p>\n

      \u4ece\u8868\u4e2d\u53ef\u4ee5\u770b\u51fa\uff0cLaVIT \u7684\u8868\u73b0\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u591a\u6a21\u6001\u8bed\u8a00\u6a21\u578b\u3002\u4e0e Emu \u76f8\u6bd4\uff0cLaVIT \u5728\u66f4\u5c0f\u7684 LLM \u6a21\u578b\u4e0a\u53d6\u5f97\u4e86\u8fdb\u4e00\u6b65\u6539\u8fdb\uff0c\u5c55\u73b0\u4e86\u51fa\u8272\u7684\u89c6\u89c9 – \u8bed\u8a00\u5bf9\u9f50\u80fd\u529b\u3002\u6b64\u5916\uff0cLaVIT \u5728\u4f7f\u7528\u66f4\u5c11\u7684\u8bad\u7ec3\u6570\u636e\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u4e86\u4e0e\u6700\u5148\u8fdb\u7684\u6587\u672c\u5230\u56fe\u50cf\u4e13\u5bb6 Parti \u53ef\u6bd4\u7684\u6027\u80fd\u3002<\/p>\n

      \u591a\u6a21\u6001\u63d0\u793a\u56fe\u50cf\u751f\u6210<\/strong><\/p>\n

      LaVIT \u80fd\u591f\u5728\u65e0\u9700\u8fdb\u884c\u4efb\u4f55\u5fae\u8c03\u7684\u60c5\u51b5\u4e0b\uff0c\u65e0\u7f1d\u5730\u63a5\u53d7\u591a\u79cd\u6a21\u6001\u7ec4\u5408\u4f5c\u4e3a\u63d0\u793a\uff0c\u751f\u6210\u76f8\u5e94\u7684\u56fe\u50cf\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u4efb\u4f55\u5fae\u8c03\u3002LaVIT \u751f\u6210\u7684\u56fe\u50cf\u80fd\u591f\u51c6\u786e\u53cd\u6620\u7ed9\u5b9a\u591a\u6a21\u6001\u63d0\u793a\u7684\u98ce\u683c\u548c\u8bed\u4e49\u3002\u800c\u4e14\u5b83\u53ef\u4ee5\u901a\u8fc7\u8f93\u5165\u7684\u591a\u6a21\u6001\u63d0\u793a\u4fee\u6539\u539f\u59cb\u8f93\u5165\u56fe\u50cf\u3002\u5728\u6ca1\u6709\u989d\u5916\u5fae\u8c03\u7684\u4e0b\u6e38\u6570\u636e\u7684\u60c5\u51b5\u4e0b\uff0c\u4f20\u7edf\u7684\u56fe\u50cf\u751f\u6210\u6a21\u578b\u5982 Stable Diffusion \u65e0\u6cd5\u8fbe\u5230\u8fd9\u79cd\u80fd\u529b\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u591a\u6a21\u6001\u56fe\u50cf\u751f\u6210\u7ed3\u679c\u7684\u793a\u4f8b<\/p>\n

      \u5b9a\u6027\u5206\u6790<\/strong><\/p>\n

      \u5982\u4e0b\u56fe\u6240\u793a\uff0cLaVIT \u7684\u52a8\u6001\u5206\u8bcd\u5668\u53ef\u4ee5\u6839\u636e\u56fe\u50cf\u5185\u5bb9\u52a8\u6001\u9009\u62e9\u6700\u5177\u4fe1\u606f\u91cf\u7684\u56fe\u50cf\u5757\uff0c\u5b66\u4e60\u5230\u7684\u4ee3\u7801\u672c\u53ef\u4ee5\u4ea7\u751f\u5177\u6709\u9ad8\u5c42\u8bed\u4e49\u7684\u89c6\u89c9\u7f16\u7801\u3002<\/p>\n

      \"\u56fe\u7247\"<\/p>\n

      \u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u5668\uff08\u5de6\uff09\u548c\u5b66\u4e60\u5230\u7684 codebook\uff08\u53f3\uff09\u7684\u53ef\u89c6\u5316<\/p>\n

      \u603b\u7ed3<\/h4>\n

      LaVIT \u7684\u51fa\u73b0\u4e3a\u591a\u6a21\u6001\u4efb\u52a1\u7684\u5904\u7406\u53c8\u63d0\u4f9b\u4e86\u4e00\u79cd\u521b\u65b0\u8303\u5f0f\uff0c\u901a\u8fc7\u4f7f\u7528\u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u5668\u5c06\u89c6\u89c9\u548c\u8bed\u8a00\u8868\u793a\u4e3a\u7edf\u4e00\u7684\u79bb\u6563 token \u8868\u793a\uff0c\u7ee7\u627f\u4e86 LLM \u6210\u529f\u7684\u81ea\u56de\u5f52\u751f\u6210\u5b66\u4e60\u8303\u5f0f\u3002\u901a\u8fc7\u5728\u7edf\u4e00\u751f\u6210\u76ee\u6807\u4e0b\u8fdb\u884c\u4f18\u5316\uff0cLaVIT \u53ef\u4ee5\u5c06\u56fe\u50cf\u89c6\u4e3a\u4e00\u79cd\u5916\u8bed\uff0c\u50cf\u6587\u672c\u4e00\u6837\u7406\u89e3\u548c\u751f\u6210\u5b83\u4eec\u3002\u8fd9\u4e00\u65b9\u6cd5\u7684\u6210\u529f\u4e3a\u672a\u6765\u591a\u6a21\u6001\u7814\u7a76\u7684\u53d1\u5c55\u65b9\u5411\u63d0\u4f9b\u4e86\u65b0\u7684\u542f\u793a\uff0c\u5229\u7528 LLM \u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5b9e\u73b0\u66f4\u667a\u80fd\u3001\u66f4\u5168\u9762\u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u6253\u5f00\u65b0\u7684\u53ef\u80fd\u6027\u3002<\/p>\n<\/div>\n<\/div>\n

      <\/div>\n

      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

      <\/div>\n","protected":false},"excerpt":{"rendered":"

      \u52a8\u6001\u89c6\u89c9\u5206\u8bcd\u7edf\u4e00\u56fe\u6587\u8868\u793a\uff0c\u5feb\u624b\u4e0e\u5317\u5927\u5408\u4f5c\u63d0\u51fa\u57fa\u5ea7\u6a21\u578b LaVIT \u5237\u699c\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u4efb\u52a1\u3002 \u5f53\u524d\u7684\u5927\u578b\u8bed\u8a00\u6a21 […]<\/p>\n

      <\/div>\n

      <\/g><\/svg><\/i> \"Loading\"<\/p>\n

      <\/div>\n","protected":false},"author":56,"featured_media":18088,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,80],"tags":[124,887],"class_list":["post-18087","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-80","tag-ai","tag-887"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18087","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=18087"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18087\/revisions"}],"predecessor-version":[{"id":18098,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18087\/revisions\/18098"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/18088"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=18087"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=18087"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=18087"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}