{"id":19557,"date":"2025-01-28T09:49:28","date_gmt":"2025-01-28T01:49:28","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=19557"},"modified":"2025-02-06T23:08:51","modified_gmt":"2025-02-06T15:08:51","slug":"deepseekfabutongyi","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/deepseekfabutongyi\/","title":{"rendered":"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro"},"content":{"rendered":"<h2>JanusFlow \u901f\u8bfb<\/h2>\n<p><a href=\"https:\/\/www.kdjingpai.com\/de\/deepseek-chatshena\/\">DeepSeek<\/a> \u56e2\u961f\u53c8\u53d1\u65b0\u6a21\u578b\uff0c28\u65e5\u51cc\u6668\u63a8\u51fa\u4e86\u521b\u65b0\u578b\u591a\u6a21\u6001\u6846\u67b6 Janus-Pro\uff0c\u8fd9\u662f\u4e00\u4e2a\u80fd\u540c\u65f6\u5904\u7406\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u7684\u7edf\u4e00\u6a21\u578b\u3002\u8be5\u6a21\u578b\u57fa\u4e8e DeepSeek-LLM-1.5b-base\/DeepSeek-LLM-7b-base \u6784\u5efa\uff0c\u652f\u6301 384 x 384 \u7684\u56fe\u50cf\u8f93\u5165\uff0c\u5e76\u4f7f\u7528\u7279\u5b9a\u7684 tokenizer \u8fdb\u884c\u56fe\u50cf\u751f\u6210\u3002\u6700\u5927\u7279\u70b9\u662f\u5c06\u89c6\u89c9\u7f16\u7801\u5206\u4e3a\u72ec\u7acb\u901a\u9053\uff0c\u540c\u65f6\u4fdd\u6301\u5355\u4e00 <a href=\"https:\/\/www.kdjingpai.com\/de\/transformer\/\">transformer<\/a> \u67b6\u6784\u8fdb\u884c\u5904\u7406\u3002<\/p>\n<p>\u8fd9\u79cd\u521b\u65b0\u8bbe\u8ba1\u4e0d\u4ec5\u89e3\u51b3\u4e86\u4f20\u7edf\u6a21\u578b\u5728\u89c6\u89c9\u7f16\u7801\u5668\u89d2\u8272\u4e0a\u7684\u51b2\u7a81\u95ee\u9898\uff0c\u8fd8\u8ba9\u6574\u4e2a\u7cfb\u7edf\u53d8\u5f97\u66f4\u52a0\u7075\u6d3b\u3002\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0cJanus-Pro \u7684\u8868\u73b0\u8d85\u8d8a\u4e86\u4e4b\u524d\u7684\u7edf\u4e00\u6a21\u578b\uff0c\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u751a\u81f3\u53ef\u4ee5\u5ab2\u7f8e\u4e13\u95e8\u7684\u4efb\u52a1\u578b\u6a21\u578b\u3002\u5728GenEval \u548c DPG-Bench \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u51fb\u8d25\u4e86 OpenAI \u7684 DALL-E 3 \u548c Stable Diffusion\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>Janus \u6a21\u578b\u7cfb\u5217\uff0c\u8d77\u59cb\u4e8e <strong>JanusFlow<\/strong>\uff0c\u65e8\u5728\u6784\u5efa\u4e00\u4e2a<strong>\u7edf\u4e00\u7684\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6<\/strong>\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u81ea\u56de\u5f52\u8bed\u8a00\u6a21\u578b (LLM) \u4e0e\u4fee\u6b63\u6d41 (Rectified Flow) \u751f\u6210\u6a21\u578b\u76f8\u7ed3\u5408\uff0c\u529b\u6c42\u5728\u5355\u4e2a\u6a21\u578b\u5185\u540c\u65f6\u5b9e\u73b0\u5353\u8d8a\u7684\u89c6\u89c9\u7406\u89e3\u548c\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u751f\u6210\u80fd\u529b\u3002<strong>Janus-Pro<\/strong> \u4f5c\u4e3a Janus \u7684\u8fdb\u9636\u7248\u672c\uff0c\u901a\u8fc7\u5728\u8bad\u7ec3\u7b56\u7565\u3001\u6570\u636e\u89c4\u6a21\u548c\u6a21\u578b\u5c3a\u5bf8\u4e0a\u7684\u5168\u9762\u4f18\u5316\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86 Janus \u6a21\u578b\u7684\u6027\u80fd\uff0c\u5e76\u5728\u591a\u9879\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53d6\u5f97\u4e86\u663e\u8457\u7684\u8fdb\u6b65\u3002\u672c\u6587\u5c06\u7cfb\u7edf\u5730\u68b3\u7406 Janus \u6a21\u578b\u4ece JanusFlow \u5230 Janus-Pro \u7684\u6f14\u8fdb\u5386\u7a0b\uff0c\u91cd\u70b9\u9610\u8ff0\u5176\u529f\u80fd\u3001\u53c2\u6570\u7279\u6027\u4ee5\u53ca\u5173\u952e\u6539\u8fdb\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>1. JanusFlow\uff1a\u7edf\u4e00\u67b6\u6784\u7684\u5960\u57fa\u4e4b\u4f5c<\/h3>\n<p><strong>\u8bba\u6587\u5730\u5740<\/strong>\uff1ahttps:\/\/arxiv.org\/pdf\/2411.07975<\/p>\n<p><strong>JanusFlow<\/strong> \u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u5176<strong>\u6781\u7b80\u4e3b\u4e49\u7684\u7edf\u4e00\u67b6\u6784<\/strong>\uff0c\u5b83\u5de7\u5999\u5730\u5c06\u4fee\u6b63\u6d41\u751f\u6210\u6a21\u578b\u65e0\u7f1d\u96c6\u6210\u5230\u81ea\u56de\u5f52 LLM \u6846\u67b6\u4e2d\uff0c\u65e0\u9700\u5bf9 LLM \u7ed3\u6784\u8fdb\u884c\u590d\u6742\u4fee\u6539\u3002\u8fd9\u4e00\u67b6\u6784\u7684\u6838\u5fc3\u529f\u80fd\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u4fee\u6b63\u6d41\u56fe\u50cf\u751f\u6210\uff1a<\/strong> JanusFlow \u5229\u7528\u4fee\u6b63\u6d41\u6a21\u578b\u8fdb\u884c\u56fe\u50cf\u751f\u6210\uff0c\u4ece\u9ad8\u65af\u566a\u58f0\u51fa\u53d1\uff0c\u8fed\u4ee3\u9884\u6d4b\u901f\u5ea6\u5411\u91cf\u4ee5\u66f4\u65b0\u56fe\u50cf\u7684 latent \u7a7a\u95f4\u8868\u793a\uff0c\u6700\u7ec8\u901a\u8fc7\u89e3\u7801\u5668\u751f\u6210\u9ad8\u8d28\u91cf\u56fe\u50cf\u3002\u8fd9\u79cd\u65b9\u6cd5\u907f\u514d\u4e86\u4f20\u7edf\u65b9\u6cd5\u4e2d LLM \u4ec5\u4f5c\u4e3a\u6761\u4ef6\u751f\u6210\u5668\uff0c\u800c\u7f3a\u4e4f\u76f4\u63a5\u751f\u6210\u80fd\u529b\u7684\u5c40\u9650\u6027\u3002<\/li>\n<li><strong>\u89e3\u8026\u89c6\u89c9\u7f16\u7801\u5668\uff1a<\/strong> \u4e3a\u4e86\u4f18\u5316\u7edf\u4e00\u6a21\u578b\u7684\u6027\u80fd\uff0cJanusFlow \u91c7\u7528\u4e86<strong>\u89e3\u8026\u7f16\u7801\u5668<\/strong>\u7b56\u7565\uff0c\u5206\u522b\u4f7f\u7528\u72ec\u7acb\u7684\u89c6\u89c9\u7f16\u7801\u5668\u5904\u7406\u7406\u89e3\u4efb\u52a1\u548c\u751f\u6210\u4efb\u52a1\uff1a\n<ul>\n<li><strong>\u7406\u89e3\u7f16\u7801\u5668 (fenc)\uff1a<\/strong> \u91c7\u7528\u9884\u8bad\u7ec3\u7684 <strong>SigLIP-Large-Patch\/16<\/strong> \u6a21\u578b\uff0c\u8d1f\u8d23\u63d0\u53d6\u56fe\u50cf\u7684\u8bed\u4e49\u7279\u5f81\uff0c\u63d0\u5347\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u3002<\/li>\n<li><strong>\u751f\u6210\u7f16\u7801\u5668 (genc) \u548c\u89e3\u7801\u5668 (gdec)\uff1a<\/strong> \u91c7\u7528\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u7684 <strong>ConvNeXt<\/strong> \u6a21\u5757\uff0c\u4e13\u95e8\u7528\u4e8e\u56fe\u50cf\u751f\u6210\u4efb\u52a1\uff0c\u4f18\u5316\u751f\u6210\u8d28\u91cf\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8868\u5f81\u5bf9\u9f50\u673a\u5236\uff1a<\/strong> \u5728\u7edf\u4e00\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cJanusFlow \u5f15\u5165<strong>\u8868\u5f81\u5bf9\u9f50<\/strong>\u673a\u5236\uff0c\u5c06\u751f\u6210\u6a21\u5757\u548c\u7406\u89e3\u6a21\u5757\u7684\u4e2d\u95f4\u8868\u5f81\u8fdb\u884c\u5bf9\u9f50\uff0c\u4ece\u800c\u52a0\u5f3a\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u8bed\u4e49\u8fde\u8d2f\u6027\u548c\u4e00\u81f4\u6027\u3002<\/li>\n<li><strong>\u4e09\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565\uff1a<\/strong> JanusFlow \u8bbe\u8ba1\u4e86\u7cbe\u7ec6\u7684\u4e09\u9636\u6bb5\u8bad\u7ec3\u65b9\u6848\uff1a\n<ol>\n<li><strong>Stage 1\uff1a\u968f\u673a\u521d\u59cb\u5316\u7ec4\u4ef6\u9002\u914d<\/strong> &#8211; \u8bad\u7ec3\u7ebf\u6027\u5c42\u3001\u751f\u6210\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\uff0c\u4f7f\u5176\u4e0e\u9884\u8bad\u7ec3\u7684 LLM \u548c SigLIP \u7f16\u7801\u5668\u534f\u540c\u5de5\u4f5c\uff0c\u4f5c\u4e3a\u540e\u7eed\u8bad\u7ec3\u7684\u521d\u59cb\u5316\u9636\u6bb5\u3002<\/li>\n<li><strong>Stage 2\uff1a\u7edf\u4e00\u9884\u8bad\u7ec3<\/strong> &#8211; \u8bad\u7ec3\u9664\u89c6\u89c9\u7f16\u7801\u5668\u5916\u7684\u6574\u4e2a\u6a21\u578b\uff0c\u878d\u5408\u591a\u6a21\u6001\u7406\u89e3\u3001\u56fe\u50cf\u751f\u6210\u548c\u7eaf\u6587\u672c\u6570\u636e\uff0c\u521d\u6b65\u5efa\u7acb\u6a21\u578b\u7684\u7edf\u4e00\u80fd\u529b\u3002<\/li>\n<li><strong>Stage 3\uff1a\u76d1\u7763\u5fae\u8c03<\/strong> &#8211; \u4f7f\u7528\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u8fdb\u4e00\u6b65\u8bad\u7ec3\u6a21\u578b\uff0c\u63d0\u5347\u5bf9\u7528\u6237\u6307\u4ee4\u7684\u54cd\u5e94\u80fd\u529b\uff0c\u5e76\u5728\u6b64\u9636\u6bb5\u89e3\u51bb SigLIP \u7f16\u7801\u5668\u7684\u53c2\u6570\u3002<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n<p><strong>\u53c2\u6570\u7279\u6027\uff1a<\/strong><\/p>\n<ul>\n<li><strong>\u57fa\u7840 LLM\uff1a<\/strong> 1.3B \u53c2\u6570\u7684\u8f7b\u91cf\u7ea7 LLM \u67b6\u6784\u3002<\/li>\n<li><strong>\u89c6\u89c9\u7f16\u7801\u5668\uff1a<\/strong> SigLIP-Large-Patch\/16 (\u7406\u89e3)\uff0cConvNeXt (\u751f\u6210\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668)\u3002<\/li>\n<li><strong>\u56fe\u50cf\u5206\u8fa8\u7387\uff1a<\/strong> 384 \u00d7 384 \u50cf\u7d20\u3002<\/li>\n<\/ul>\n<p><strong>\u6027\u80fd\u8868\u73b0\uff1a<\/strong> JanusFlow \u5728\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u548c\u591a\u6a21\u6001\u7406\u89e3\u4efb\u52a1\u4e0a\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\uff0c\u8d85\u8d8a\u4e86\u8bb8\u591a\u4e13\u4e1a\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u5176\u7edf\u4e00\u67b6\u6784\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>2. Janus-Pro\uff1a\u6570\u636e\u3001\u6a21\u578b\u4e0e\u7b56\u7565\u7684\u5168\u9762\u5347\u7ea7<\/h3>\n<p><strong>\u8bba\u6587\u5730\u5740<\/strong>\uff1ahttps:\/\/github.com\/deepseek-ai\/Janus\/blob\/main\/janus_pro_tech_report.pdf<\/p>\n<p><strong>Janus-Pro<\/strong> \u4f5c\u4e3a Janus \u7684\u6539\u8fdb\u7248\u672c\uff0c\u5728 JanusFlow \u7684\u57fa\u7840\u4e0a\uff0c\u4e3b\u8981\u5728\u4ee5\u4e0b\u4e09\u4e2a\u65b9\u9762\u8fdb\u884c\u4e86\u5347\u7ea7\uff1a<\/p>\n<ul>\n<li><strong>\u4f18\u5316\u7684\u8bad\u7ec3\u7b56\u7565\uff1a<\/strong> Janus-Pro \u5bf9 JanusFlow \u7684\u4e09\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565\u8fdb\u884c\u4e86\u7cbe\u7ec6\u8c03\u6574\uff0c\u65e8\u5728\u89e3\u51b3\u8bad\u7ec3\u6548\u7387\u548c\u6027\u80fd\u74f6\u9888\uff1a\n<ul>\n<li><strong>Stage 1\uff1a\u5ef6\u957f ImageNet \u6570\u636e\u96c6\u8bad\u7ec3<\/strong> &#8211; \u589e\u52a0\u5728 ImageNet \u6570\u636e\u96c6\u4e0a\u7684\u8bad\u7ec3\u6b65\u6570\uff0c\u4f7f\u6a21\u578b\u66f4\u5145\u5206\u5730\u5b66\u4e60\u50cf\u7d20\u4f9d\u8d56\u6027\uff0c\u63d0\u5347\u56fe\u50cf\u751f\u6210\u7684\u57fa\u7840\u80fd\u529b\u3002<\/li>\n<li><strong>Stage 2\uff1a\u805a\u7126\u6587\u672c\u5230\u56fe\u50cf\u6570\u636e\u8bad\u7ec3<\/strong> &#8211; \u5728 Stage 2 \u8bad\u7ec3\u4e2d\uff0c\u79fb\u9664 ImageNet \u6570\u636e\uff0c\u76f4\u63a5\u4f7f\u7528\u5e38\u89c4\u7684\u6587\u672c\u5230\u56fe\u50cf\u6570\u636e\u96c6\uff0c\u4f7f\u6a21\u578b\u66f4\u9ad8\u6548\u5730\u5b66\u4e60\u57fa\u4e8e\u5bc6\u96c6\u6587\u672c\u63cf\u8ff0\u751f\u6210\u9ad8\u8d28\u91cf\u56fe\u50cf\u3002<\/li>\n<li><strong>Stage 3\uff1a\u8c03\u6574\u6570\u636e\u6bd4\u4f8b<\/strong> &#8211; \u5728\u76d1\u7763\u5fae\u8c03\u9636\u6bb5\uff0c\u5fae\u8c03\u4e86\u591a\u6a21\u6001\u7406\u89e3\u6570\u636e\u3001\u7eaf\u6587\u672c\u6570\u636e\u548c\u6587\u672c\u5230\u56fe\u50cf\u6570\u636e\u7684\u6bd4\u4f8b\uff08\u4ece 7:3:10 \u8c03\u6574\u4e3a 5:1:4\uff09\uff0c\u5728\u4fdd\u8bc1\u89c6\u89c9\u751f\u6210\u80fd\u529b\u7684\u540c\u65f6\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u591a\u6a21\u6001\u7406\u89e3\u6027\u80fd\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6269\u5c55\u7684\u8bad\u7ec3\u6570\u636e\uff1a<\/strong> Janus-Pro \u5927\u5e45\u6269\u5c55\u4e86\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u548c\u591a\u6837\u6027\uff0c\u4ee5\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u548c\u751f\u6210\u8d28\u91cf\uff1a\n<ul>\n<li><strong>\u591a\u6a21\u6001\u7406\u89e3\u6570\u636e\uff1a<\/strong> \u5728 Stage 2 \u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u65b0\u589e\u7ea6 9000 \u4e07\u6837\u672c\uff0c\u6db5\u76d6\u66f4\u5e7f\u6cdb\u7684\u56fe\u50cf\u5b57\u5e55\u6570\u636e\uff08\u5982 YFCC\uff09\u4ee5\u53ca\u8868\u683c\u3001\u56fe\u8868\u3001\u6587\u6863\u7406\u89e3\u6570\u636e\uff08\u5982 Docmatix\uff09\u3002\u5728 Stage 3 \u5fae\u8c03\u9636\u6bb5\uff0c\u8fdb\u4e00\u6b65\u5f15\u5165 <a href=\"https:\/\/www.kdjingpai.com\/de\/deepseek-vl2\/\">DeepSeek-VL2<\/a> \u6570\u636e\u96c6\u4ee5\u53ca MEME \u7406\u89e3\u3001\u4e2d\u6587\u5bf9\u8bdd\u6570\u636e\u7b49\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u7684\u5bf9\u8bdd\u80fd\u529b\u548c\u591a\u4efb\u52a1\u5904\u7406\u80fd\u529b\u3002<\/li>\n<li><strong>\u89c6\u89c9\u751f\u6210\u6570\u636e\uff1a<\/strong> \u4e3a\u4e86\u63d0\u5347\u751f\u6210\u56fe\u50cf\u7684\u7f8e\u5b66\u8d28\u91cf\u548c\u7a33\u5b9a\u6027\uff0cJanus-Pro \u5f15\u5165\u4e86\u7ea6 7200 \u4e07\u9ad8\u8d28\u91cf\u7684\u5408\u6210\u7f8e\u5b66\u6570\u636e\uff0c\u5e76\u8c03\u6574\u4e86\u771f\u5b9e\u6570\u636e\u4e0e\u5408\u6210\u6570\u636e\u7684\u6bd4\u4f8b\u81f3 1:1\uff0c\u5b9e\u9a8c\u8868\u660e\uff0c\u5408\u6210\u6570\u636e\u7684\u52a0\u5165\u52a0\u901f\u4e86\u6a21\u578b\u6536\u655b\uff0c\u5e76\u663e\u8457\u63d0\u5347\u4e86\u751f\u6210\u56fe\u50cf\u7684\u7f8e\u5b66\u8d28\u91cf\u548c\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6269\u5c55\u7684\u6a21\u578b\u5c3a\u5bf8\uff1a<\/strong> Janus-Pro \u4e0d\u4ec5\u4fdd\u7559\u4e86 JanusFlow \u7684 1.5B \u53c2\u6570\u6a21\u578b\uff0c\u8fd8\u8fdb\u4e00\u6b65\u6269\u5c55\u5230\u4e86 <strong>7B \u53c2\u6570<\/strong>\uff0c\u5e76\u63d0\u4f9b\u4e86 <strong>Janus-Pro \u7cfb\u5217\u7684 1.5B \u548c 7B \u4e24\u79cd\u6a21\u578b\u5c3a\u5bf8<\/strong>\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u66f4\u5927\u89c4\u6a21\u7684 LLM \u80fd\u591f\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\uff0c\u5e76\u52a0\u5feb\u6536\u655b\u901f\u5ea6\uff0c\u9a8c\u8bc1\u4e86 Janus \u6a21\u578b\u67b6\u6784\u7684\u53ef\u6269\u5c55\u6027\u3002<\/li>\n<\/ul>\n<p><strong>\u53c2\u6570\u7279\u6027\uff1a<\/strong><\/p>\n<ul>\n<li><strong>\u6a21\u578b\u5c3a\u5bf8\uff1a<\/strong> \u63d0\u4f9b 1.5B \u548c 7B \u4e24\u79cd\u6a21\u578b\u5c3a\u5bf8\u9009\u62e9\u3002<\/li>\n<li><strong>\u67b6\u6784\uff1a<\/strong> \u6cbf\u7528 JanusFlow \u7684\u89e3\u8026\u89c6\u89c9\u7f16\u7801\u5668\u67b6\u6784\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u6570\u636e\uff1a<\/strong> \u5927\u5e45\u6269\u5c55\u548c\u4f18\u5316\u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u8bad\u7ec3\u6570\u636e\u3002<\/li>\n<li><strong>\u56fe\u50cf\u5206\u8fa8\u7387\uff1a<\/strong> \u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u56fe\u50cf\u5206\u8fa8\u7387\u4ecd\u4e3a 384 \u00d7 384 \u50cf\u7d20\u3002<\/li>\n<\/ul>\n<p><strong>\u6027\u80fd\u8868\u73b0\uff1a<\/strong> Janus-Pro \u5728\u5404\u9879\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5747\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u5c24\u5176\u662f\u5728\u591a\u6a21\u6001\u7406\u89e3\u57fa\u51c6 MMBench \u548c\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u57fa\u51c6 GenEval\u3001DPG-Bench \u4e0a\uff0c\u5747\u8d85\u8d8a\u4e86 JanusFlow \u4ee5\u53ca\u5176\u4ed6\u5148\u8fdb\u7684\u7edf\u4e00\u6a21\u578b\u548c\u4e13\u4e1a\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u6570\u636e\u3001\u6a21\u578b\u548c\u7b56\u7565\u5347\u7ea7\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3.Janus \u771f\u5b9e\u4f7f\u7528\u573a\u666f<\/h3>\n<p><strong>\u89c6\u89c9\u7406\u89e3\u529f\u80fd\uff1a<\/strong><\/p>\n<ul>\n<li>\u56fe\u50cf\u63cf\u8ff0 (Image Description\/Captioning):\n<ul>\n<li>\u8be6\u7ec6\u573a\u666f\u63cf\u8ff0:\u00a0\u80fd\u591f\u6839\u636e\u56fe\u50cf\u5185\u5bb9\u751f\u6210\u8be6\u5c3d\u7684\u6587\u5b57\u63cf\u8ff0\uff0c\u5305\u62ec\u573a\u666f\u5143\u7d20\u3001\u7269\u4f53\u3001\u73af\u5883\u6c1b\u56f4\u7b49\u3002\uff08\u793a\u4f8b\uff1a\u63cf\u8ff0\u897f\u6e56\u4e09\u6f6d\u5370\u6708\u3001\u6d77\u8fb9\u666f\u89c2\u7b49\uff09<\/li>\n<li>\u56fe\u8868\u63cf\u8ff0:\u00a0\u80fd\u591f\u7406\u89e3\u5e76\u63cf\u8ff0\u56fe\u8868\u4fe1\u606f\uff0c\u4f8b\u5982\u6761\u5f62\u56fe\u7684\u6570\u636e\u8868\u793a\u548c\u8d8b\u52bf\u5206\u6790\u3002\uff08\u793a\u4f8b\uff1a\u89e3\u8bfb &#8220;Kid&#8217;s Favourite Fruits&#8221; \u6761\u5f62\u56fe\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u7269\u4f53\u8bc6\u522b\u4e0e\u5206\u7c7b (Object Recognition\/Classification):\n<ul>\n<li>\u8bc6\u522b\u56fe\u50cf\u4e2d\u7684\u7269\u4f53\u79cd\u7c7b:\u00a0\u80fd\u591f\u8bc6\u522b\u5e76\u5217\u4e3e\u56fe\u50cf\u4e2d\u51fa\u73b0\u7684\u7269\u4f53\u7c7b\u522b\u3002\uff08\u793a\u4f8b\uff1a\u8bc6\u522b\u56fe\u4e2d\u6c34\u679c\u7684\u79cd\u7c7b\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u7269\u4f53\u8ba1\u6570 (Object Counting):\n<ul>\n<li>\u7cbe\u786e\u8ba1\u6570\u56fe\u50cf\u4e2d\u7684\u7269\u4f53\u6570\u91cf:\u00a0\u80fd\u591f\u51c6\u786e\u5730\u7edf\u8ba1\u56fe\u50cf\u4e2d\u7279\u5b9a\u7269\u4f53\u7684\u6570\u91cf\u3002\uff08\u793a\u4f8b\uff1a\u6570\u51fa\u56fe\u4e2d\u4f01\u9e45\u7684\u6570\u91cf\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u5730\u6807\u8bc6\u522b (Landmark Recognition):\n<ul>\n<li>\u8bc6\u522b\u56fe\u50cf\u4e2d\u7684\u8457\u540d\u5730\u6807:\u00a0\u80fd\u591f\u8bc6\u522b\u56fe\u50cf\u4e2d\u51fa\u73b0\u7684\u5730\u6807\u6027\u5efa\u7b51\u6216\u5730\u70b9\u3002\uff08\u793a\u4f8b\uff1a\u8bc6\u522b\u51fa\u897f\u6e56\u4e09\u6f6d\u5370\u6708\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u6587\u672c\u8bc6\u522b (Text Recognition\/OCR):\n<ul>\n<li>\u8bc6\u522b\u56fe\u50cf\u4e2d\u7684\u6587\u5b57\u5185\u5bb9:\u00a0\u80fd\u591f\u8bc6\u522b\u56fe\u50cf\u4e2d\u51fa\u73b0\u7684\u6587\u5b57\uff0c\u5e76\u63d0\u53d6\u6587\u672c\u4fe1\u606f\u3002\uff08\u793a\u4f8b\uff1a\u8bc6\u522b\u9ed1\u677f\u4e0a\u7684 &#8220;Serving Soul since Twenty Twelve&#8221;\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u89c6\u89c9\u95ee\u7b54 (Visual Question Answering):\n<ul>\n<li>\u56de\u7b54\u57fa\u4e8e\u56fe\u50cf\u5185\u5bb9\u7684\u63d0\u95ee:\u00a0\u80fd\u591f\u7406\u89e3\u56fe\u50cf\u5185\u5bb9\uff0c\u5e76\u6839\u636e\u7528\u6237\u7684\u63d0\u95ee\u7ed9\u51fa\u5408\u7406\u7684\u7b54\u6848\u3002\uff08\u793a\u4f8b\uff1a\u9488\u5bf9\u56fe\u50cf\u63d0\u95ee &#8220;\u56fe\u4e2d\u6709\u4ec0\u4e48\u6c34\u679c\uff1f&#8221;\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u89c6\u89c9\u63a8\u7406\u4e0e\u77e5\u8bc6\u6574\u5408 (Visual Reasoning\/Knowledge Integration):\n<ul>\n<li>\u7406\u89e3\u56fe\u50cf\u80cc\u540e\u7684\u542b\u4e49\u548c\u5173\u8054:\u00a0\u80fd\u591f\u8fdb\u884c\u66f4\u6df1\u5c42\u6b21\u7684\u89c6\u89c9\u63a8\u7406\uff0c\u5e76\u7ed3\u5408\u77e5\u8bc6\u80cc\u666f\u8fdb\u884c\u7406\u89e3\u3002\uff08\u793a\u4f8b\uff1a\u89e3\u91ca\u8499\u5a1c\u4e3d\u838e\u72d7\u72d7\u56fe\u7684\u5e7d\u9ed8\u4e4b\u5904\u3001\u7406\u89e3\u86cb\u7cd5\u7684\u5361\u901a\u4e3b\u9898\u80cc\u666f\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u4ee3\u7801\u751f\u6210 (Code Generation &#8211; Python for Plotting):\n<ul>\n<li>\u6839\u636e\u7528\u6237\u6307\u4ee4\u751f\u6210\u4ee3\u7801:\u00a0\u80fd\u591f\u7406\u89e3\u7528\u6237\u5bf9\u56fe\u8868\u7ed8\u5236\u7684\u9700\u6c42\uff0c\u5e76\u751f\u6210\u76f8\u5e94\u7684 Python \u4ee3\u7801\u3002\uff08\u793a\u4f8b\uff1a\u751f\u6210\u7ed8\u5236\u6761\u5f62\u56fe\u7684 Python \u4ee3\u7801\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><strong>\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u529f\u80fd\uff1a<\/strong><\/p>\n<ul>\n<li>\u6587\u672c\u5f15\u5bfc\u7684\u56fe\u50cf\u751f\u6210 (Text-Guided Image Generation):\n<ul>\n<li>\u6839\u636e\u6587\u672c\u63cf\u8ff0\u751f\u6210\u56fe\u50cf:\u00a0\u80fd\u591f\u6839\u636e\u7528\u6237\u8f93\u5165\u7684\u6587\u672c\u63d0\u793a\u8bcd\uff0c\u751f\u6210\u4e0e\u4e4b\u8bed\u4e49\u76f8\u5173\u7684\u56fe\u50cf\u3002<\/li>\n<li>\u521b\u610f\u56fe\u50cf\u751f\u6210:\u00a0\u80fd\u591f\u7406\u89e3\u62bd\u8c61\u548c\u5bcc\u6709\u60f3\u8c61\u529b\u7684\u6587\u672c\u63d0\u793a\uff0c\u751f\u6210\u5177\u6709\u521b\u610f\u548c\u827a\u672f\u6027\u7684\u56fe\u50cf\u3002\uff08\u793a\u4f8b\uff1a\u98de\u5929\u9cb8\u9c7c\u3001\u5b87\u5b99\u661f\u4e91\u67ef\u57fa\u72ac\u7b49\uff09<\/li>\n<li>\u98ce\u683c\u5316\u56fe\u50cf\u751f\u6210:\u00a0\u80fd\u591f\u6839\u636e\u6587\u672c\u63d0\u793a\u8bcd\u4e2d\u7684\u98ce\u683c\u63cf\u8ff0\uff0c\u751f\u6210\u5177\u6709\u7279\u5b9a\u827a\u672f\u98ce\u683c\u7684\u56fe\u50cf\u3002\uff08\u793a\u4f8b\uff1a\u6587\u827a\u590d\u5174\u98ce\u683c\u6559\u5802\u3001\u4e2d\u56fd\u6c34\u58a8\u753b\u98ce\u683c\u5c71\u6751\u7b49\uff09<\/li>\n<li>\u751f\u6210\u5305\u542b\u6587\u5b57\u7684\u56fe\u50cf (Simple Text Generation on Images):\u00a0\u80fd\u591f\u751f\u6210\u5305\u542b\u7b80\u5355\u6587\u5b57\u5143\u7d20\u7684\u56fe\u50cf\u3002\uff08\u793a\u4f8b\uff1a\u5728\u9ed1\u677f\u4e0a\u5199\u5b57 &#8220;Hello&#8221;\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><strong>\u5b9e\u9645\u5e94\u7528\u573a\u666f\u793a\u4f8b\uff1a<\/strong><\/p>\n<ul>\n<li>\u667a\u80fd\u52a9\u624b:\u00a0\u4f5c\u4e3a\u591a\u6a21\u6001\u667a\u80fd\u52a9\u624b\uff0c\u7406\u89e3\u7528\u6237\u4e0a\u4f20\u7684\u56fe\u7247\u5e76\u8fdb\u884c\u95ee\u7b54\u3001\u63cf\u8ff0\u3001\u5206\u6790\u7b49\u3002<\/li>\n<li>\u5185\u5bb9\u521b\u4f5c:\u00a0\u8f85\u52a9\u5185\u5bb9\u521b\u4f5c\u8005\u5feb\u901f\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u7d20\u6750\uff0c\u4f8b\u5982\u793e\u4ea4\u5a92\u4f53\u914d\u56fe\u3001\u6587\u7ae0\u63d2\u56fe\u7b49\u3002<\/li>\n<li>\u6559\u80b2\u5e94\u7528:\u00a0\u7528\u4e8e\u56fe\u50cf\u8bc6\u522b\u6559\u5b66\u3001\u56fe\u8868\u89e3\u8bfb\u6559\u5b66\u7b49\uff0c\u5e2e\u52a9\u5b66\u751f\u7406\u89e3\u89c6\u89c9\u4fe1\u606f\u3002<\/li>\n<li>\u4fe1\u606f\u68c0\u7d22:\u00a0\u901a\u8fc7\u56fe\u50cf\u641c\u7d22\uff0c\u5e76\u7ed3\u5408\u6587\u5b57\u7406\u89e3\u548c\u751f\u6210\u80fd\u529b\uff0c\u63d0\u4f9b\u66f4\u4e30\u5bcc\u7684\u68c0\u7d22\u7ed3\u679c\u3002<\/li>\n<li>\u827a\u672f\u521b\u4f5c:\u00a0\u4f5c\u4e3a\u521b\u610f\u5de5\u5177\uff0c\u8f85\u52a9\u827a\u672f\u5bb6\u8fdb\u884c\u56fe\u50cf\u521b\u4f5c\uff0c\u63a2\u7d22\u65b0\u7684\u89c6\u89c9\u8868\u8fbe\u5f62\u5f0f\u3002<\/li>\n<\/ul>\n<p><strong>\u9700\u8981\u6ce8\u610f\u7684\u53c2\u6570\u7279\u6027\u4e0e\u5c40\u9650\u6027\uff1a<\/strong><\/p>\n<ul>\n<li>\u56fe\u50cf\u5206\u8fa8\u7387\u9650\u5236:\u00a0\u5f53\u524d\u6a21\u578b\u8bad\u7ec3\u548c\u6d4b\u8bd5\u4e3b\u8981\u57fa\u4e8e 384&#215;384 \u5206\u8fa8\u7387\u56fe\u50cf\uff0c\u5bf9\u4e8e\u9700\u8981\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u573a\u666f\u53ef\u80fd\u5b58\u5728\u5c40\u9650\u6027\u3002<\/li>\n<li>\u7ec6\u8282\u7cbe\u7ec6\u5ea6:\u00a0\u867d\u7136\u56fe\u50cf\u8bed\u4e49\u5185\u5bb9\u4e30\u5bcc\uff0c\u4f46\u53d7\u9650\u4e8e\u5206\u8fa8\u7387\u548c vision tokenizer \u7684\u91cd\u5efa\u635f\u5931\uff0c\u5728\u7ec6\u8282\u7cbe\u7ec6\u5ea6\u4e0a\u53ef\u80fd\u4ecd\u6709\u63d0\u5347\u7a7a\u95f4\uff0c\u4f8b\u5982\u5c0f\u5c3a\u5bf8\u4eba\u8138\u533a\u57df\u53ef\u80fd\u4e0d\u591f\u7ec6\u817b\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>4. \u603b\u7ed3\u4e0e\u5c55\u671b<\/h3>\n<p>Janus \u6a21\u578b\u7cfb\u5217\uff0c\u4ece JanusFlow \u5230 Janus-Pro\uff0c\u5c55\u73b0\u4e86\u5728\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u9886\u57df\u6301\u7eed\u7a81\u7834\u7684\u6f5c\u529b\u3002JanusFlow \u5960\u5b9a\u4e86\u7edf\u4e00\u67b6\u6784\u7684\u57fa\u7840\uff0c\u800c Janus-Pro \u5219\u901a\u8fc7\u8bad\u7ec3\u7b56\u7565\u4f18\u5316\u3001\u6570\u636e\u89c4\u6a21\u6269\u5c55\u548c\u6a21\u578b\u5c3a\u5bf8\u5347\u7ea7\uff0c\u5b9e\u73b0\u4e86\u6027\u80fd\u7684\u98de\u8dc3\u3002Janus-Pro \u7684\u6210\u529f\u9a8c\u8bc1\u4e86<strong>\u6570\u636e\u9a71\u52a8\u548c\u6a21\u578b\u6269\u5c55\u662f\u63d0\u5347\u7edf\u4e00\u6a21\u578b\u6027\u80fd\u7684\u5173\u952e<\/strong>\uff0c\u540c\u65f6\u4e5f\u4e3a\u672a\u6765\u66f4\u5927\u89c4\u6a21\u3001\u66f4\u9ad8\u6027\u80fd\u7684\u7edf\u4e00\u591a\u6a21\u6001\u6a21\u578b\u7684\u53d1\u5c55\u6307\u660e\u4e86\u65b9\u5411\u3002Janus \u6a21\u578b\u7cfb\u5217\u7684\u6f14\u8fdb\uff0c\u4e0d\u4ec5\u63a8\u52a8\u4e86\u591a\u6a21\u6001\u6a21\u578b\u7684\u8fdb\u6b65\uff0c\u4e5f\u4e3a\u6784\u5efa\u66f4\u901a\u7528\u3001\u66f4\u667a\u80fd\u7684 AI \u7cfb\u7edf\u5960\u5b9a\u4e86\u575a\u5b9e\u7684\u57fa\u7840\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>\u5b8c\u6574\u8bba\u6587\u300aJanus-Pro\uff1a\u7edf\u4e00\u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u6a21\u578b\uff0c\u901a\u8fc7\u6570\u636e\u548c\u6a21\u578b\u6269\u5c55\u5b9e\u73b0\u300b<\/h2>\n<p><strong>\u4f5c\u8005<\/strong>: Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan<br \/>\n<strong>\u9879\u76ee\u9875\u9762<\/strong>:\u00a0https:\/\/github.com\/deepseek-ai\/Janus<\/p>\n<h3>\u6458\u8981<\/h3>\n<p>\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 Janus-Pro\uff0c\u8fd9\u662f\u4e4b\u524d Janus \u6a21\u578b\u7684\u589e\u5f3a\u7248\u672c\u3002\u5177\u4f53\u6765\u8bf4\uff0cJanus-Pro \u96c6\u6210\u4e86 (1) \u4f18\u5316\u7684\u8bad\u7ec3\u7b56\u7565\uff0c(2) \u6269\u5c55\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u4ee5\u53ca (3) \u6269\u5c55\u5230\u66f4\u5927\u7684\u6a21\u578b\u89c4\u6a21\u3002\u901a\u8fc7\u8fd9\u4e9b\u6539\u8fdb\uff0cJanus-Pro \u5728\u591a\u6a21\u6001\u7406\u89e3\u548c\u6587\u672c\u751f\u6210\u56fe\u50cf\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u65b9\u9762\u53d6\u5f97\u4e86\u91cd\u5927\u8fdb\u6b65\uff0c\u540c\u65f6\u589e\u5f3a\u4e86\u6587\u672c\u751f\u6210\u56fe\u50cf\u7684\u7a33\u5b9a\u6027\u3002\u6211\u4eec\u5e0c\u671b\u8fd9\u9879\u5de5\u4f5c\u80fd\u591f\u6fc0\u53d1\u8be5\u9886\u57df\u7684\u8fdb\u4e00\u6b65\u63a2\u7d22\u3002\u4ee3\u7801\u548c\u6a21\u578b\u5df2\u516c\u5f00\u53d1\u5e03\u3002<\/p>\n<h3>1. \u5f15\u8a00<\/h3>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19562\" title=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e.jpg\" alt=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-1\" width=\"1341\" height=\"513\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e.jpg 1341w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e-300x115.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e-1024x392.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e-768x294.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8f2689af4f9fe1e-18x7.jpg 18w\" sizes=\"auto, (max-width: 1341px) 100vw, 1341px\" \/><br \/>\n(a) \u56db\u4e2a\u591a\u6a21\u6001\u7406\u89e3\u57fa\u51c6\u7684\u5e73\u5747\u6027\u80fd\u3002(b) \u6587\u672c\u751f\u6210\u56fe\u50cf\u6307\u4ee4\u9075\u5faa\u57fa\u51c6\u7684\u6027\u80fd\u3002<\/p>\n<p><strong>\u56fe 1 |<\/strong>\u00a0Janus-Pro \u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u7ed3\u679c\u3002\u5bf9\u4e8e\u591a\u6a21\u6001\u7406\u89e3\uff0c\u6211\u4eec\u5bf9 POPE\u3001MME-Perception\u3001GQA \u548c MMMU \u7684\u51c6\u786e\u6027\u8fdb\u884c\u4e86\u5e73\u5747\u3002MME-Perception \u7684\u5206\u6570\u9664\u4ee5 20 \u4ee5\u7f29\u653e\u5230 [0, 100] \u8303\u56f4\u3002\u5bf9\u4e8e\u89c6\u89c9\u751f\u6210\uff0c\u6211\u4eec\u8bc4\u4f30\u4e86 GenEval \u548c DPG-Bench \u4e24\u4e2a\u6307\u4ee4\u9075\u5faa\u57fa\u51c6\u7684\u6027\u80fd\u3002\u603b\u4f53\u800c\u8a00\uff0cJanus-Pro \u4f18\u4e8e\u4e4b\u524d\u7684\u6700\u5148\u8fdb\u7edf\u4e00\u591a\u6a21\u6001\u6a21\u578b\u4ee5\u53ca\u4e00\u4e9b\u7279\u5b9a\u4efb\u52a1\u7684\u6a21\u578b\u3002\u5728\u5c4f\u5e55\u4e0a\u89c2\u770b\u6548\u679c\u6700\u4f73\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19564\" title=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373.jpg\" alt=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-2\" width=\"1350\" height=\"537\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373.jpg 1350w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373-300x119.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373-1024x407.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373-768x305.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/db763e264116373-18x7.jpg 18w\" sizes=\"auto, (max-width: 1350px) 100vw, 1350px\" \/><br \/>\n1.\u4e00\u5f20\u6a59\u8272\u7684\u6a58\u5b50\u7b80\u7ea6\u7167\u7247 2.\u4e00\u5757\u5e72\u51c0\u7684\u9ed1\u677f\uff0c\u4e0a\u9762\u7528\u767d\u8272\u7c89\u7b14\u7cbe\u786e\u6e05\u6670\u5730\u5199\u7740\u4e00\u4e2a\u7eff\u8272\u7684\u8868\u9762\u548c\u201cHello\u201d\u8fd9\u4e2a\u8bcd 3.\u8c61\u5f81\u7740\u7e41\u8363\u7684\u5411\u65e5\u8475\u7279\u5199\uff0c\u7eff\u8272\u7684\u830e\u548c\u53f6\uff0c\u82b1\u74e3\u76db\u5f00\uff0c\u4e00\u53ea\u871c\u8702\u505c\u5728\u4e0a\u9762\uff0c\u7fc5\u8180\u5728\u9633\u5149\u4e0b\u95ea\u95ea\u53d1\u5149\u3002<\/p>\n<p><strong>\u56fe 2 |<\/strong> Janus-Pro \u4e0e\u5176\u524d\u8eab Janus \u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u5bf9\u6bd4\u3002Janus-Pro \u4e3a\u7b80\u77ed\u63d0\u793a\u63d0\u4f9b\u4e86\u66f4\u7a33\u5b9a\u7684\u8f93\u51fa\uff0c\u5177\u6709\u66f4\u9ad8\u7684\u89c6\u89c9\u8d28\u91cf\u3001\u66f4\u4e30\u5bcc\u7684\u7ec6\u8282\u4ee5\u53ca\u751f\u6210\u7b80\u5355\u6587\u672c\u7684\u80fd\u529b\u3002\u56fe\u50cf\u5206\u8fa8\u7387\u4e3a 384&#215;384\u3002\u5728\u5c4f\u5e55\u4e0a\u89c2\u770b\u6548\u679c\u6700\u4f73\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6700\u8fd1\u5728\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u6a21\u578b\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55 [30, 40, 45, 46, 48, 50, 54, 55]\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5df2\u88ab\u8bc1\u660e\u53ef\u4ee5\u589e\u5f3a\u89c6\u89c9\u751f\u6210\u4efb\u52a1\u4e2d\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\uff0c\u540c\u65f6\u51cf\u5c11\u6a21\u578b\u5197\u4f59\u3002\u5927\u591a\u6570\u8fd9\u4e9b\u65b9\u6cd5\u4f7f\u7528\u76f8\u540c\u7684\u89c6\u89c9\u7f16\u7801\u5668\u6765\u5904\u7406\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u7684\u8f93\u5165\u3002\u7531\u4e8e\u8fd9\u4e24\u9879\u4efb\u52a1\u6240\u9700\u7684\u8868\u793a\u4e0d\u540c\uff0c\u8fd9\u901a\u5e38\u4f1a\u5bfc\u81f4\u591a\u6a21\u6001\u7406\u89e3\u6027\u80fd\u4e0d\u4f73\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cJanus [46] \u63d0\u51fa\u4e86\u89e3\u8026\u89c6\u89c9\u7f16\u7801\uff0c\u8fd9\u7f13\u89e3\u4e86\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u4e4b\u95f4\u7684\u51b2\u7a81\uff0c\u5728\u8fd9\u4e24\u9879\u4efb\u52a1\u4e2d\u90fd\u53d6\u5f97\u4e86\u4f18\u5f02\u7684\u6027\u80fd\u3002<\/p>\n<p>\u4f5c\u4e3a\u5f00\u521b\u6027\u6a21\u578b\uff0cJanus \u5728 1B \u53c2\u6570\u89c4\u6a21\u4e0b\u5f97\u5230\u4e86\u9a8c\u8bc1\u3002\u7136\u800c\uff0c\u7531\u4e8e\u8bad\u7ec3\u6570\u636e\u7684\u9650\u5236\u548c\u76f8\u5bf9\u8f83\u5c0f\u7684\u6a21\u578b\u5bb9\u91cf\uff0c\u5b83\u8868\u73b0\u51fa\u4e86\u4e00\u4e9b\u7f3a\u70b9\uff0c\u4f8b\u5982\u5728\u7b80\u77ed\u63d0\u793a\u56fe\u50cf\u751f\u6210\u65b9\u9762\u7684\u6027\u80fd\u4e0d\u4f73\u4ee5\u53ca\u6587\u672c\u751f\u6210\u56fe\u50cf\u8d28\u91cf\u4e0d\u7a33\u5b9a\u3002\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 Janus-Pro\uff0c\u8fd9\u662f Janus \u7684\u589e\u5f3a\u7248\u672c\uff0c\u5b83\u5728\u4e09\u4e2a\u65b9\u9762\u8fdb\u884c\u4e86\u6539\u8fdb\uff1a\u8bad\u7ec3\u7b56\u7565\u3001\u6570\u636e\u548c\u6a21\u578b\u89c4\u6a21\u3002Janus-Pro \u7cfb\u5217\u5305\u62ec\u4e24\u79cd\u6a21\u578b\u89c4\u6a21\uff1a1B \u548c 7B\uff0c\u5c55\u793a\u4e86\u89c6\u89c9\u7f16\u7801\u89e3\u7801\u65b9\u6cd5\u7684\u53ef\u6269\u5c55\u6027\u3002<\/p>\n<p>\u6211\u4eec\u5728\u591a\u4e2a\u57fa\u51c6\u4e0a\u8bc4\u4f30\u4e86 Janus-Pro\uff0c\u7ed3\u679c\u663e\u793a\u5176\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u4f18\u8d8a\uff0c\u6587\u672c\u751f\u6210\u56fe\u50cf\u6307\u4ee4\u9075\u5faa\u6027\u80fd\u663e\u8457\u63d0\u9ad8\u3002\u5177\u4f53\u6765\u8bf4\uff0cJanus-Pro-7B \u5728\u591a\u6a21\u6001\u7406\u89e3\u57fa\u51c6 MMBench [29] \u4e0a\u83b7\u5f97\u4e86 79.2 \u5206\uff0c\u8d85\u8fc7\u4e86\u4e4b\u524d\u7684\u6700\u5148\u8fdb\u7edf\u4e00\u591a\u6a21\u6001\u6a21\u578b\uff0c\u5982 Janus [46] (69.4)\u3001TokenFlow [34] (68.9) \u548c MetaMorph [42] (75.2)\u3002\u6b64\u5916\uff0c\u5728\u6587\u672c\u751f\u6210\u56fe\u50cf\u6307\u4ee4\u9075\u5faa\u6392\u884c\u699c GenEval [14] \u4e2d\uff0cJanus-Pro-7B \u5f97\u5206\u4e3a 0.80\uff0c\u8d85\u8fc7\u4e86 Janus [46] (0.61)\u3001DALL-E 3 (0.67) \u548c Stable Diffusion 3 Medium [11] (0.74)\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19563\" title=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd.jpg\" alt=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-3\" width=\"1314\" height=\"525\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd.jpg 1314w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd-300x120.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd-1024x409.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd-768x307.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/11f37299e81d2cd-18x7.jpg 18w\" sizes=\"auto, (max-width: 1314px) 100vw, 1314px\" \/><br \/>\n<strong>\u56fe 3 |<\/strong>\u00a0Janus-Pro \u7684\u67b6\u6784\u3002\u6211\u4eec\u5bf9\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u8fdb\u884c\u89c6\u89c9\u7f16\u7801\u89e3\u8026\u3002\u201cUnd. Encoder\u201d\u548c\u201cGen. Encoder\u201d\u5206\u522b\u662f\u201cUnderstanding Encoder\u201d\u548c\u201cGeneration Encoder\u201d\u7684\u7f29\u5199\u3002\u5728\u5c4f\u5e55\u4e0a\u89c2\u770b\u6548\u679c\u6700\u4f73\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>2. \u65b9\u6cd5<\/h3>\n<h4>2.1. \u67b6\u6784<\/h4>\n<p>Janus-Pro \u7684\u67b6\u6784\u5982\u56fe 3 \u6240\u793a\uff0c\u4e0e Janus [46] \u76f8\u540c\u3002\u6574\u4f53\u67b6\u6784\u7684\u6838\u5fc3\u8bbe\u8ba1\u539f\u5219\u662f\u5bf9\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u8fdb\u884c\u89c6\u89c9\u7f16\u7801\u89e3\u8026\u3002\u6211\u4eec\u5e94\u7528\u72ec\u7acb\u7684\u7f16\u7801\u65b9\u6cd5\u5c06\u539f\u59cb\u8f93\u5165\u8f6c\u6362\u4e3a\u7279\u5f81\uff0c\u7136\u540e\u7531\u4e00\u4e2a\u7edf\u4e00\u7684\u81ea\u56de\u5f52\u53d8\u6362\u5668\u8fdb\u884c\u5904\u7406\u3002\u5bf9\u4e8e\u591a\u6a21\u6001\u7406\u89e3\uff0c\u6211\u4eec\u4f7f\u7528 SigLIP [53] \u7f16\u7801\u5668\u4ece\u56fe\u50cf\u4e2d\u63d0\u53d6\u9ad8\u7ef4\u8bed\u4e49\u7279\u5f81\u3002\u8fd9\u4e9b\u7279\u5f81\u4ece 2-D \u7f51\u683c\u5c55\u5e73\u4e3a 1-D \u5e8f\u5217\uff0c\u5e76\u4f7f\u7528\u7406\u89e3\u9002\u914d\u5668\u5c06\u8fd9\u4e9b\u56fe\u50cf\u7279\u5f81\u6620\u5c04\u5230 LLM \u7684\u8f93\u5165\u7a7a\u95f4\u3002\u5bf9\u4e8e\u89c6\u89c9\u751f\u6210\u4efb\u52a1\uff0c\u6211\u4eec\u4f7f\u7528 [38] \u4e2d\u7684 VQ \u6807\u8bb0\u5668\u5c06\u56fe\u50cf\u8f6c\u6362\u4e3a\u79bb\u6563 ID\u3002ID \u5e8f\u5217\u5c55\u5e73\u4e3a 1-D \u540e\uff0c\u6211\u4eec\u4f7f\u7528\u751f\u6210\u9002\u914d\u5668\u5c06\u6bcf\u4e2a ID \u5bf9\u5e94\u7684\u4ee3\u7801\u672c\u5d4c\u5165\u6620\u5c04\u5230 LLM \u7684\u8f93\u5165\u7a7a\u95f4\u3002\u7136\u540e\u6211\u4eec\u5c06\u8fd9\u4e9b\u7279\u5f81\u5e8f\u5217\u8fde\u63a5\u8d77\u6765\uff0c\u5f62\u6210\u4e00\u4e2a\u591a\u6a21\u6001\u7279\u5f81\u5e8f\u5217\uff0c\u968f\u540e\u5c06\u5176\u8f93\u5165\u5230 LLM \u4e2d\u8fdb\u884c\u5904\u7406\u3002\u9664\u4e86 LLM \u4e2d\u5185\u7f6e\u7684\u9884\u6d4b\u5934\u5916\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u4e00\u4e2a\u968f\u673a\u521d\u59cb\u5316\u7684\u9884\u6d4b\u5934\u7528\u4e8e\u89c6\u89c9\u751f\u6210\u4efb\u52a1\u4e2d\u7684\u56fe\u50cf\u9884\u6d4b\u3002\u6574\u4e2a\u6a21\u578b\u9075\u5faa\u81ea\u56de\u5f52\u6846\u67b6\u3002<\/p>\n<h4>2.2. \u4f18\u5316\u7684\u8bad\u7ec3\u7b56\u7565<\/h4>\n<p>Janus \u7684\u524d\u4e00\u4e2a\u7248\u672c\u91c7\u7528\u4e09\u9636\u6bb5\u8bad\u7ec3\u8fc7\u7a0b\u3002\u7b2c\u4e00\u9636\u6bb5\u4fa7\u91cd\u4e8e\u8bad\u7ec3\u9002\u914d\u5668\u548c\u56fe\u50cf\u5934\u3002\u7b2c\u4e8c\u9636\u6bb5\u5904\u7406\u7edf\u4e00\u9884\u8bad\u7ec3\uff0c\u5728\u6b64\u9636\u6bb5\uff0c\u9664\u4e86\u7406\u89e3\u7f16\u7801\u5668\u548c\u751f\u6210\u7f16\u7801\u5668\u4e4b\u5916\u7684\u6240\u6709\u7ec4\u4ef6\u90fd\u4f1a\u66f4\u65b0\u5176\u53c2\u6570\u3002\u7b2c\u4e09\u9636\u6bb5\u662f\u76d1\u7763\u5fae\u8c03\uff0c\u5b83\u5efa\u7acb\u5728\u7b2c\u4e8c\u9636\u6bb5\u7684\u57fa\u7840\u4e0a\uff0c\u901a\u8fc7\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8fdb\u4e00\u6b65\u89e3\u9501\u7406\u89e3\u7f16\u7801\u5668\u7684\u53c2\u6570\u3002\u8fd9\u79cd\u8bad\u7ec3\u7b56\u7565\u5b58\u5728\u4e00\u4e9b\u95ee\u9898\u3002\u5728\u7b2c\u4e8c\u9636\u6bb5\uff0cJanus \u5c06\u6587\u672c\u751f\u6210\u56fe\u50cf\u80fd\u529b\u7684\u8bad\u7ec3\u5206\u4e3a\u4e24\u90e8\u5206\uff0c\u9075\u5faa PixArt [4]\u3002\u7b2c\u4e00\u90e8\u5206\u4f7f\u7528 ImageNet [9] \u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u7528\u56fe\u50cf\u7c7b\u522b\u540d\u79f0\u4f5c\u4e3a\u6587\u672c\u751f\u6210\u56fe\u50cf\u7684\u63d0\u793a\uff0c\u76ee\u6807\u662f\u5efa\u6a21\u50cf\u7d20\u4f9d\u8d56\u6027\u3002\u7b2c\u4e8c\u90e8\u5206\u4f7f\u7528\u666e\u901a\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u3002\u5728\u5b9e\u65bd\u8fc7\u7a0b\u4e2d\uff0c\u7b2c\u4e8c\u9636\u6bb5\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u8bad\u7ec3\u6b65\u9aa4\u4e2d\u6709 66.67% \u5206\u914d\u7ed9\u4e86\u7b2c\u4e00\u90e8\u5206\u3002\u7136\u800c\uff0c\u901a\u8fc7\u8fdb\u4e00\u6b65\u7684\u5b9e\u9a8c\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u79cd\u7b56\u7565\u662f\u6b21\u4f18\u7684\uff0c\u5e76\u4e14\u5bfc\u81f4\u4e86\u663e\u8457\u7684\u8ba1\u7b97\u6548\u7387\u4f4e\u4e0b\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u505a\u4e86\u4e24\u9879\u4fee\u6539\u3002<\/p>\n<p>\u2022\u00a0<strong>\u7b2c\u4e00\u9636\u6bb5\u66f4\u957f\u7684\u8bad\u7ec3<\/strong>: \u6211\u4eec\u589e\u52a0\u4e86\u7b2c\u4e00\u9636\u6bb5\u7684\u8bad\u7ec3\u6b65\u9aa4\uff0c\u5141\u8bb8\u5728 ImageNet \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5145\u5206\u7684\u8bad\u7ec3\u3002\u6211\u4eec\u7684\u53d1\u73b0\u8868\u660e\uff0c\u5373\u4f7f LLM \u53c2\u6570\u56fa\u5b9a\uff0c\u6a21\u578b\u4e5f\u53ef\u4ee5\u6709\u6548\u5730\u5efa\u6a21\u50cf\u7d20\u4f9d\u8d56\u6027\uff0c\u5e76\u6839\u636e\u7c7b\u522b\u540d\u79f0\u751f\u6210\u5408\u7406\u7684\u56fe\u50cf\u3002<br \/>\n\u2022\u00a0<strong>\u7b2c\u4e8c\u9636\u6bb5\u91cd\u70b9\u8bad\u7ec3<\/strong>: \u5728\u7b2c\u4e8c\u9636\u6bb5\uff0c\u6211\u4eec\u653e\u5f03\u4e86 ImageNet \u6570\u636e\uff0c\u76f4\u63a5\u5229\u7528\u666e\u901a\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u6570\u636e\u6765\u8bad\u7ec3\u6a21\u578b\u6839\u636e\u5bc6\u96c6\u63cf\u8ff0\u751f\u6210\u56fe\u50cf\u3002\u8fd9\u79cd\u91cd\u65b0\u8bbe\u8ba1\u7684\u65b9\u6cd5\u4f7f\u7b2c\u4e8c\u9636\u6bb5\u80fd\u591f\u66f4\u6709\u6548\u5730\u5229\u7528\u6587\u672c\u751f\u6210\u56fe\u50cf\u6570\u636e\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\u548c\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u8c03\u6574\u4e86\u7b2c\u4e09\u9636\u6bb5\u76d1\u7763\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u4e0d\u540c\u7c7b\u578b\u6570\u636e\u96c6\u7684\u6570\u636e\u6bd4\u4f8b\uff0c\u5c06\u591a\u6a21\u6001\u6570\u636e\u3001\u7eaf\u6587\u672c\u6570\u636e\u548c\u6587\u672c\u751f\u6210\u56fe\u50cf\u6570\u636e\u7684\u6bd4\u4f8b\u4ece 7:3:10 \u6539\u4e3a 5:1:4\u3002\u901a\u8fc7\u7565\u5fae\u964d\u4f4e\u6587\u672c\u751f\u6210\u56fe\u50cf\u6570\u636e\u7684\u6bd4\u4f8b\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u8fd9\u79cd\u8c03\u6574\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u4fdd\u6301\u5f3a\u5927\u7684\u89c6\u89c9\u751f\u6210\u80fd\u529b\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u591a\u6a21\u6001\u7406\u89e3\u6027\u80fd\u7684\u63d0\u5347\u3002<\/p>\n<h4>2.3. \u6570\u636e\u6269\u5c55<\/h4>\n<p>\u6211\u4eec\u5728\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u65b9\u9762\u6269\u5c55\u4e86 Janus \u4f7f\u7528\u7684\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n<p>\u2022\u00a0<strong>\u591a\u6a21\u6001\u7406\u89e3<\/strong>: \u5bf9\u4e8e\u7b2c\u4e8c\u9636\u6bb5\u9884\u8bad\u7ec3\u6570\u636e\uff0c\u6211\u4eec\u53c2\u8003 DeepSeekVL2 [49]\uff0c\u589e\u52a0\u4e86\u5927\u7ea6 9000 \u4e07\u4e2a\u6837\u672c\u3002\u8fd9\u4e9b\u6570\u636e\u5305\u62ec\u56fe\u50cf\u5b57\u5e55\u6570\u636e\u96c6\uff08\u4f8b\u5982\uff0cYFCC [31]\uff09\uff0c\u4ee5\u53ca\u7528\u4e8e\u8868\u683c\u3001\u56fe\u8868\u548c\u6587\u6863\u7406\u89e3\u7684\u6570\u636e\uff08\u4f8b\u5982\uff0cDocmatix [20]\uff09\u3002\u5bf9\u4e8e\u7b2c\u4e09\u9636\u6bb5\u76d1\u7763\u5fae\u8c03\u6570\u636e\uff0c\u6211\u4eec\u8fd8\u4ece DeepSeek-VL2 \u4e2d\u52a0\u5165\u4e86\u989d\u5916\u7684\u6570\u636e\u96c6\uff0c\u4f8b\u5982 MEME \u7406\u89e3\u3001\u4e2d\u6587\u5bf9\u8bdd\u6570\u636e\uff0c\u4ee5\u53ca\u65e8\u5728\u589e\u5f3a\u5bf9\u8bdd\u4f53\u9a8c\u7684\u6570\u636e\u96c6\u3002\u8fd9\u4e9b\u589e\u52a0\u663e\u8457\u6269\u5c55\u4e86\u6a21\u578b\u7684\u80fd\u529b\uff0c\u4e30\u5bcc\u4e86\u5b83\u5904\u7406\u5404\u79cd\u4efb\u52a1\u7684\u80fd\u529b\uff0c\u540c\u65f6\u63d0\u9ad8\u4e86\u6574\u4f53\u5bf9\u8bdd\u4f53\u9a8c\u3002<\/p>\n<p>\u2022\u00a0<strong>\u89c6\u89c9\u751f\u6210<\/strong>: \u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u524d\u4e00\u4e2a\u7248\u672c Janus \u4e2d\u4f7f\u7528\u7684\u771f\u5b9e\u4e16\u754c\u6570\u636e\u7f3a\u4e4f\u8d28\u91cf\uff0c\u5e76\u4e14\u5305\u542b\u663e\u8457\u7684\u566a\u58f0\uff0c\u8fd9\u901a\u5e38\u4f1a\u5bfc\u81f4\u6587\u672c\u751f\u6210\u56fe\u50cf\u4e0d\u7a33\u5b9a\uff0c\u5bfc\u81f4\u7f8e\u5b66\u6548\u679c\u4e0d\u4f73\u7684\u8f93\u51fa\u3002\u5728 Janus-Pro \u4e2d\uff0c\u6211\u4eec\u52a0\u5165\u4e86\u5927\u7ea6 7200 \u4e07\u4e2a\u5408\u6210\u7f8e\u5b66\u6570\u636e\u6837\u672c\uff0c\u4f7f\u771f\u5b9e\u4e0e\u5408\u6210\u6570\u636e\u7684\u6bd4\u4f8b\u5728\u7edf\u4e00\u9884\u8bad\u7ec3\u9636\u6bb5\u8fbe\u5230 1:1\u3002\u8fd9\u4e9b\u5408\u6210\u6570\u636e\u6837\u672c\u7684\u63d0\u793a\u662f\u516c\u5f00\u53ef\u7528\u7684\uff0c\u4f8b\u5982 [43] \u4e2d\u7684\u63d0\u793a\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u5f53\u5728\u5408\u6210\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\u65f6\uff0c\u6a21\u578b\u6536\u655b\u901f\u5ea6\u66f4\u5feb\uff0c\u751f\u6210\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u8f93\u51fa\u4e0d\u4ec5\u66f4\u7a33\u5b9a\uff0c\u800c\u4e14\u7f8e\u5b66\u8d28\u91cf\u4e5f\u663e\u8457\u63d0\u9ad8\u3002<\/p>\n<h4>2.4. \u6a21\u578b\u6269\u5c55<\/h4>\n<p>Janus \u7684\u524d\u4e00\u4e2a\u7248\u672c\u9a8c\u8bc1\u4e86\u4f7f\u7528 1.5B LLM \u8fdb\u884c\u89c6\u89c9\u7f16\u7801\u89e3\u8026\u7684\u6709\u6548\u6027\u3002\u5728 Janus-Pro \u4e2d\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u6269\u5c55\u5230 7B\uff0c1.5B \u548c 7B LLM \u7684\u8d85\u53c2\u6570\u5728\u8868 1 \u4e2d\u8be6\u7ec6\u8bf4\u660e\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5f53\u4f7f\u7528\u66f4\u5927\u89c4\u6a21\u7684 LLM \u65f6\uff0c\u4e0e\u8f83\u5c0f\u7684\u6a21\u578b\u76f8\u6bd4\uff0c\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u635f\u5931\u7684\u6536\u655b\u901f\u5ea6\u663e\u8457\u63d0\u9ad8\u3002\u8fd9\u4e00\u53d1\u73b0\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u8fd9\u79cd\u65b9\u6cd5\u5f3a\u5927\u7684\u53ef\u6269\u5c55\u6027\u3002<\/p>\n<p><strong>\u8868 1 |<\/strong>\u00a0Janus-Pro \u7684\u67b6\u6784\u914d\u7f6e\u3002\u6211\u4eec\u5217\u51fa\u4e86\u67b6\u6784\u7684\u8d85\u53c2\u6570\u3002<\/p>\n<table>\n<tbody>\n<tr>\n<td><\/td>\n<td>Janus-Pro-1B<\/td>\n<td>Janus-Pro-7B<\/td>\n<\/tr>\n<tr>\n<td>\u8bcd\u6c47\u91cf\u5927\u5c0f<\/td>\n<td>100K<\/td>\n<td>100K<\/td>\n<\/tr>\n<tr>\n<td>\u5d4c\u5165\u5927\u5c0f<\/td>\n<td>2048<\/td>\n<td>4096<\/td>\n<\/tr>\n<tr>\n<td>\u4e0a\u4e0b\u6587\u7a97\u53e3<\/td>\n<td>4096<\/td>\n<td>4096<\/td>\n<\/tr>\n<tr>\n<td>\u6ce8\u610f\u529b\u5934\u6570<\/td>\n<td>16<\/td>\n<td>32<\/td>\n<\/tr>\n<tr>\n<td>\u5c42\u6570<\/td>\n<td>24<\/td>\n<td>30<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>&nbsp;<\/p>\n<p><strong>\u8868 2 |<\/strong>\u00a0Janus-Pro \u8bad\u7ec3\u7684\u8be6\u7ec6\u8d85\u53c2\u6570\u3002\u6570\u636e\u6bd4\u4f8b\u6307\u7684\u662f\u591a\u6a21\u6001\u7406\u89e3\u6570\u636e\u3001\u7eaf\u6587\u672c\u6570\u636e\u548c\u89c6\u89c9\u751f\u6210\u6570\u636e\u7684\u6bd4\u4f8b\u3002<\/p>\n<p><img decoding=\"async\" title=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/58943527c372445.png\" alt=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<h3>3. \u5b9e\u9a8c<\/h3>\n<h4>3.1. \u5b9e\u73b0\u7ec6\u8282<\/h4>\n<p>\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528 DeepSeek-LLM (1.5B \u548c 7B) [3] \u4f5c\u4e3a\u57fa\u7840\u8bed\u8a00\u6a21\u578b\uff0c\u6700\u5927\u652f\u6301\u5e8f\u5217\u957f\u5ea6\u4e3a 4096\u3002\u5bf9\u4e8e\u7406\u89e3\u4efb\u52a1\u4e2d\u4f7f\u7528\u7684\u89c6\u89c9\u7f16\u7801\u5668\uff0c\u6211\u4eec\u9009\u62e9\u4e86 SigLIP-Large-Patch16-384 [53]\u3002\u751f\u6210\u7f16\u7801\u5668\u6709\u4e00\u4e2a\u5927\u5c0f\u4e3a 16,384 \u7684\u4ee3\u7801\u672c\uff0c\u5e76\u5c06\u56fe\u50cf\u4e0b\u91c7\u6837 16 \u500d\u3002\u7406\u89e3\u9002\u914d\u5668\u548c\u751f\u6210\u9002\u914d\u5668\u90fd\u662f\u4e24\u5c42 MLP\u3002\u6bcf\u4e2a\u9636\u6bb5\u7684\u8be6\u7ec6\u8d85\u53c2\u6570\u5728\u8868 2 \u4e2d\u63d0\u4f9b\u3002\u6240\u6709\u56fe\u50cf\u90fd\u8c03\u6574\u4e3a 384&#215;384 \u50cf\u7d20\u3002\u5bf9\u4e8e\u591a\u6a21\u6001\u7406\u89e3\u6570\u636e\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u7684\u957f\u8fb9\u8c03\u6574\u4e3a 384\uff0c\u77ed\u8fb9\u7528\u80cc\u666f\u8272\uff08RGB: 127, 127, 127\uff09\u586b\u5145\u4ee5\u8fbe\u5230 384\u3002\u5bf9\u4e8e\u89c6\u89c9\u751f\u6210\u6570\u636e\uff0c\u77ed\u8fb9\u8c03\u6574\u4e3a 384\uff0c\u957f\u8fb9\u88c1\u526a\u4e3a 384\u3002\u6211\u4eec\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u5e8f\u5217\u6253\u5305\u6765\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002\u6211\u4eec\u6839\u636e\u6307\u5b9a\u7684\u6bd4\u4f8b\u6df7\u5408\u6240\u6709\u6570\u636e\u7c7b\u578b\uff0c\u5728\u5355\u4e2a\u8bad\u7ec3\u6b65\u9aa4\u4e2d\u8fdb\u884c\u3002\u6211\u4eec\u7684 Janus \u4f7f\u7528 HAI-LLM [15] \u8fdb\u884c\u8bad\u7ec3\u548c\u8bc4\u4f30\uff0cHAI-LLM \u662f\u4e00\u4e2a\u57fa\u4e8e PyTorch \u6784\u5efa\u7684\u8f7b\u91cf\u7ea7\u9ad8\u6548\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6\u3002\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u5728\u914d\u5907 8 \u4e2a Nvidia A100 (40GB) GPU \u7684 16\/32 \u8282\u70b9\u96c6\u7fa4\u4e0a\u5927\u7ea6\u82b1\u8d39\u4e86 7\/14 \u5929\uff0c\u7528\u4e8e 1.5B\/7B \u6a21\u578b\u3002<\/p>\n<h4>3.2. \u8bc4\u4f30\u8bbe\u7f6e<\/h4>\n<p><strong>\u591a\u6a21\u6001\u7406\u89e3 (Multimodal Understanding)\uff1a<\/strong>\u00a0\u4e3a\u4e86\u8bc4\u4f30\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\uff0c\u6211\u4eec\u5728\u5e7f\u6cdb\u8ba4\u53ef\u7684\u57fa\u4e8e\u56fe\u50cf\u7684\u89c6\u89c9-\u8bed\u8a00\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u5bf9\u6211\u4eec\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u8fd9\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u5305\u62ec GQA [17], POPE [23], MME [12], SEED [21], MMB [29], MM-Vet [51], \u4ee5\u53ca MMMU [52]\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u8868 3 |<\/strong> \u4e0e\u6700\u5148\u8fdb\u6280\u672f\u5728\u591a\u6a21\u6001\u7406\u89e3\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6bd4\u8f83\u3002\u201cUnd.\u201d \u548c \u201cGen.\u201d \u5206\u522b\u8868\u793a \u201c\u7406\u89e3\u201d \u548c \u201c\u751f\u6210\u201d\u3002 \u4f7f\u7528\u5916\u90e8\u9884\u8bad\u7ec3\u6269\u6563\u6a21\u578b\u7684\u6a21\u578b\u7528 \u2020 \u6807\u8bb0\u3002<\/p>\n<p><img decoding=\"async\" title=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/a609813238bdadc.png\" alt=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u89c6\u89c9\u751f\u6210 (Visual Generation)\uff1a<\/strong>\u00a0\u4e3a\u4e86\u8bc4\u4f30\u89c6\u89c9\u751f\u6210\u80fd\u529b\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 GenEval [14] \u548c DPG-Bench [16]\u3002GenEval \u662f\u4e00\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u56fe\u50cf\u5230\u6587\u672c\u751f\u6210\u57fa\u51c6\u6d4b\u8bd5\uff0c\u65e8\u5728\u901a\u8fc7\u5bf9\u5176\u7ec4\u5408\u80fd\u529b\u7684\u8be6\u7ec6\u5b9e\u4f8b\u7ea7\u5206\u6790\u6765\u53cd\u6620\u89c6\u89c9\u751f\u6210\u6a21\u578b\u7684\u5168\u9762\u751f\u6210\u80fd\u529b\u3002DPG-Bench (\u5bc6\u96c6\u63d0\u793a\u56fe\u57fa\u51c6\u6d4b\u8bd5) \u662f\u4e00\u4e2a\u5305\u542b 1065 \u4e2a\u5197\u957f\u3001\u5bc6\u96c6\u63d0\u793a\u7684\u7efc\u5408\u6570\u636e\u96c6\uff0c\u65e8\u5728\u8bc4\u4f30\u6587\u672c\u5230\u56fe\u50cf\u6a21\u578b\u7684\u590d\u6742\u8bed\u4e49\u5bf9\u9f50\u80fd\u529b\u3002<\/p>\n<h4>3.3. \u4e0e\u6700\u5148\u8fdb\u7684\u6bd4\u8f83<\/h4>\n<p><strong>\u591a\u6a21\u6001\u7406\u89e3\u6027\u80fd (Multimodal Understanding Performance)\uff1a<\/strong>\u00a0\u6211\u4eec\u5728\u8868 3 \u4e2d\u5c06\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u4e0e\u6700\u5148\u8fdb\u7684\u7edf\u4e00\u6a21\u578b\u548c\u4ec5\u652f\u6301\u7406\u89e3\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002Janus-Pro \u53d6\u5f97\u4e86\u603b\u4f53\u6700\u4f73\u7ed3\u679c\u3002\u8fd9\u53ef\u4ee5\u5f52\u56e0\u4e8e\u5bf9\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u7684\u89c6\u89c9\u7f16\u7801\u8fdb\u884c\u4e86\u89e3\u8026\uff0c\u4ece\u800c\u51cf\u8f7b\u4e86\u8fd9\u4e24\u9879\u4efb\u52a1\u4e4b\u95f4\u7684\u51b2\u7a81\u3002\u5f53\u4e0e\u89c4\u6a21\u5927\u5f97\u591a\u7684\u6a21\u578b\u76f8\u6bd4\u65f6\uff0cJanus-Pro \u4ecd\u7136\u5177\u6709\u5f88\u5f3a\u7684\u7ade\u4e89\u529b\u3002\u4f8b\u5982\uff0cJanus-Pro-7B \u5728\u9664 GQA \u4e4b\u5916\u7684\u6240\u6709\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5747\u4f18\u4e8e TokenFlow-XL (13B)\u3002<\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u8868 4 | <\/strong>\u5728 GenEval \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5bf9\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u80fd\u529b\u7684\u8bc4\u4f30\u3002\u201cUnd.\u201d \u548c \u201cGen.\u201d \u5206\u522b\u8868\u793a\u201c\u7406\u89e3\u201d\u548c\u201c\u751f\u6210\u201d\u3002\u4f7f\u7528\u5916\u90e8\u9884\u8bad\u7ec3\u6269\u6563\u6a21\u578b\u7684\u6a21\u578b\u7528 \u2020 \u6807\u8bb0\u3002<\/p>\n<p><img decoding=\"async\" title=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/9530e72c0d02110.png\" alt=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u8868 5 | <\/strong>\u5728 DPG-Bench \u4e0a\u7684\u6027\u80fd\u3002\u9664\u4e86 Janus \u548c Janus-Pro \u4e4b\u5916\uff0c\u6b64\u8868\u4e2d\u7684\u6240\u6709\u65b9\u6cd5\u90fd\u662f\u7279\u5b9a\u4e8e\u751f\u6210\u4efb\u52a1\u7684\u6a21\u578b\u3002<\/p>\n<p><img decoding=\"async\" title=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f2ad4d58b56bde0.png\" alt=\"DeepSeek \u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6846\u67b6\uff1a\u4ece JanusFlow \u5230 Janus-Pro\uff0c-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u89c6\u89c9\u751f\u6210\u6027\u80fd (Visual Generation Performance)\uff1a<\/strong>\u00a0\u6211\u4eec\u62a5\u544a\u4e86\u5728 GenEval \u548c DPG-Bench \u4e0a\u7684\u89c6\u89c9\u751f\u6210\u6027\u80fd\u3002\u5982\u8868 4 \u6240\u793a\uff0c\u6211\u4eec\u7684 Janus-Pro-7B \u5728 GenEval \u4e0a\u83b7\u5f97\u4e86 80% \u7684\u603b\u4f53\u51c6\u786e\u7387\uff0c\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u7edf\u4e00\u6216\u4ec5\u652f\u6301\u751f\u6210\u7684\u65b9\u6cd5\uff0c\u4f8b\u5982\uff0cTransfusion [55] (63%)\uff0cSD3-Medium (74%) \u548c DALL-E 3 (67%)\u3002\u8fd9\u8868\u660e\u6211\u4eec\u7684\u65b9\u6cd5\u5177\u6709\u66f4\u597d\u7684\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u3002\u5982\u8868 5 \u6240\u793a\uff0cJanus-Pro \u5728 DPG-Bench \u4e0a\u53d6\u5f97\u4e86 84.19 \u5206\uff0c\u8d85\u8fc7\u4e86\u6240\u6709\u5176\u4ed6\u65b9\u6cd5\u3002\u8fd9\u8868\u660e Janus-Pro \u64c5\u957f\u9075\u5faa\u5bc6\u96c6\u7684\u6307\u4ee4\u8fdb\u884c\u6587\u672c\u5230\u56fe\u50cf\u7684\u751f\u6210\u3002<\/p>\n<h4>3.4. \u5b9a\u6027\u7ed3\u679c<\/h4>\n<p>\u6211\u4eec\u5728\u56fe 4 \u4e2d\u5c55\u793a\u4e86\u591a\u6a21\u6001\u7406\u89e3\u7684\u7ed3\u679c\u3002Janus-Pro \u5728\u5904\u7406\u6765\u81ea\u4e0d\u540c\u60c5\u5883\u7684\u8f93\u5165\u65f6\u8868\u73b0\u51fa\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u7406\u89e3\u80fd\u529b\uff0c\u5c55\u73b0\u4e86\u5176\u5f3a\u5927\u7684\u80fd\u529b\u3002\u6211\u4eec\u8fd8\u5728\u56fe 4 \u7684\u4e0b\u534a\u90e8\u5206\u5c55\u793a\u4e86\u4e00\u4e9b\u6587\u672c\u751f\u6210\u56fe\u50cf\u751f\u6210\u7ed3\u679c\u3002Janus-Pro-7B \u751f\u6210\u7684\u56fe\u50cf\u975e\u5e38\u903c\u771f\uff0c\u5c3d\u7ba1\u5206\u8fa8\u7387\u4ec5\u4e3a $384\\times384$\uff0c\u4f46\u4ecd\u7136\u5305\u542b\u5927\u91cf\u7ec6\u8282\u3002\u5bf9\u4e8e\u5bcc\u6709\u60f3\u8c61\u529b\u548c\u521b\u9020\u6027\u7684\u573a\u666f\uff0cJanus-Pro-7B \u51c6\u786e\u5730\u6355\u6349\u4e86\u63d0\u793a\u4e2d\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u751f\u6210\u5408\u7406\u4e14\u8fde\u8d2f\u7684\u56fe\u50cf\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19565\" title=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7.jpg\" alt=\"DeepSeek\u53d1\u5e03\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u751f\u6210\u6a21\u578b\uff1a\u4ece JanusFlow \u5230 Janus-Pro-4\" width=\"1347\" height=\"1950\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7.jpg 1347w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7-207x300.jpg 207w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7-707x1024.jpg 707w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7-768x1112.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7-1061x1536.jpg 1061w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8e5687adb44dda7-8x12.jpg 8w\" sizes=\"auto, (max-width: 1347px) 100vw, 1347px\" \/><br \/>\n<strong>\u56fe 4 |<\/strong>\u00a0\u591a\u6a21\u6001\u7406\u89e3\u548c\u89c6\u89c9\u751f\u6210\u80fd\u529b\u7684\u5b9a\u6027\u7ed3\u679c\u3002\u8be5\u6a21\u578b\u662f Janus-Pro-7B\uff0c\u89c6\u89c9\u751f\u6210\u7684\u56fe\u50cf\u8f93\u51fa\u5206\u8fa8\u7387\u4e3a $384\\times384$\u3002\u5728\u5c4f\u5e55\u4e0a\u89c2\u770b\u6548\u679c\u6700\u4f73\u3002<\/p>\n<h3>4. \u7ed3\u8bba<\/h3>\n<p>\u672c\u6587\u4ece\u8bad\u7ec3\u7b56\u7565\u3001\u6570\u636e\u548c\u6a21\u578b\u89c4\u6a21\u4e09\u4e2a\u65b9\u9762\u5bf9 Janus \u8fdb\u884c\u4e86\u6539\u8fdb\u3002\u8fd9\u4e9b\u589e\u5f3a\u63aa\u65bd\u4f7f\u5f97\u591a\u6a21\u6001\u7406\u89e3\u548c\u6587\u672c\u751f\u6210\u56fe\u50cf\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\u3002\u7136\u800c\uff0cJanus-Pro \u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u6027\u3002\u5728\u591a\u6a21\u6001\u7406\u89e3\u65b9\u9762\uff0c\u8f93\u5165\u5206\u8fa8\u7387\u9650\u5236\u4e3a $384\\times384$\uff0c\u8fd9\u5f71\u54cd\u4e86\u5176\u5728 OCR \u7b49\u7ec6\u7c92\u5ea6\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u3002\u5bf9\u4e8e\u6587\u672c\u751f\u6210\u56fe\u50cf\uff0c\u4f4e\u5206\u8fa8\u7387\u52a0\u4e0a\u89c6\u89c9\u6807\u8bb0\u5668\u5f15\u5165\u7684\u91cd\u6784\u635f\u5931\uff0c\u5bfc\u81f4\u56fe\u50cf\u867d\u7136\u8bed\u4e49\u5185\u5bb9\u4e30\u5bcc\uff0c\u4f46\u4ecd\u7136\u7f3a\u4e4f\u7ec6\u8282\u3002\u4f8b\u5982\uff0c\u5360\u636e\u6709\u9650\u56fe\u50cf\u7a7a\u95f4\u7684\u5c0f\u9762\u90e8\u533a\u57df\u53ef\u80fd\u770b\u8d77\u6765\u4e0d\u591f\u8be6\u7ec6\u3002\u63d0\u9ad8\u56fe\u50cf\u5206\u8fa8\u7387\u53ef\u4ee5\u7f13\u89e3\u8fd9\u4e9b\u95ee\u9898\u3002<\/p>\n<h3>\u53c2\u8003\u6587\u732e<\/h3>\n<p>[1] J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, \u548c J. Zhou. Qwen-vl: \u4e00\u4e2a\u5177\u6709\u591a\u529f\u80fd\u7684\u524d\u6cbf\u5927\u578b\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2308.12966, 2023.<br \/>\n[2] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, \u7b49. \u901a\u8fc7\u66f4\u597d\u7684\u5b57\u5e55\u6539\u8fdb\u56fe\u50cf\u751f\u6210\u3002\u8ba1\u7b97\u673a\u79d1\u5b66\u3002https:\/\/cdn.openai.com\/papers\/dall-e-3.pdf, 2(3):8, 2023.<br \/>\n[3] X. Bi, D. Chen, G. Chen, S. Chen, D. Dai, C. Deng, H. Ding, K. Dong, Q. Du, Z. Fu, \u7b49. DeepSeek LLM: \u4f7f\u7528\u957f\u671f\u4e3b\u4e49\u6269\u5c55\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2401.02954, 2024.<br \/>\n[4] J. Chen, J. Yu, C. Ge, L. Yao, E. Xie, Y. Wu, Z. Wang, J. Kwok, P. Luo, H. Lu, \u7b49. PixArt\u210e: \u5feb\u901f\u8bad\u7ec3\u6269\u6563\u53d8\u6362\u5668\u7528\u4e8e\u7167\u7247\u7ea7\u903c\u771f\u7684\u6587\u672c\u751f\u6210\u56fe\u50cf\u5408\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2310.00426, 2023.<br \/>\n[5] J. Chen, C. Ge, E. Xie, Y. Wu, L. Yao, X. Ren, Z. Wang, P. Luo, H. Lu, \u548c Z. Li. PixArt-Sigma: \u5f31\u5230\u5f3a\u7684\u6269\u6563\u53d8\u6362\u5668\u8bad\u7ec3\u7528\u4e8e 4K \u6587\u672c\u751f\u6210\u56fe\u50cf\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2403.04692, 2024.<br \/>\n[6] X. Chu, L. Qiao, X. Lin, S. Xu, Y. Yang, Y. Hu, F. Wei, X. Zhang, B. Zhang, X. Wei, \u7b49. Mobilevlm: \u4e00\u4e2a\u5feb\u901f\u3001\u53ef\u590d\u73b0\u4e14\u5f3a\u5927\u7684\u79fb\u52a8\u8bbe\u5907\u89c6\u89c9\u8bed\u8a00\u52a9\u624b\u3002arXiv \u9884\u5370\u672c arXiv:2312.16886, 2023.<br \/>\n[7] X. Chu, L. Qiao, X. Zhang, S. Xu, F. Wei, Y. Yang, X. Sun, Y. Hu, X. Lin, B. Zhang, \u7b49. Mobilevlm v2: \u66f4\u5feb\u901f\u3001\u66f4\u5f3a\u5927\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u57fa\u7840\u3002arXiv \u9884\u5370\u672c arXiv:2402.03766, 2024.<br \/>\n[8] W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. Fung, \u548c S. Hoi. Instructblip: \u671d\u7740\u5177\u6709\u6307\u4ee4\u5fae\u8c03\u7684\u901a\u7528\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u53d1\u5c55, 2023.<br \/>\n[9] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, \u548c L. Fei-Fei. Imagenet: \u4e00\u4e2a\u5927\u89c4\u6a21\u5206\u5c42\u56fe\u50cf\u6570\u636e\u5e93\u3002\u5728 2009 \u5e74 IEEE \u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u6a21\u5f0f\u8bc6\u522b\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 248\u2013255. \u7535\u6c14\u7535\u5b50\u5de5\u7a0b\u5e08\u5b66\u4f1a, 2009.<br \/>\n[10] R. Dong, C. Han, Y. Peng, Z. Qi, Z. Ge, J. Yang, L. Zhao, J. Sun, H. Zhou, H. Wei, \u7b49. Dreamllm: \u591a\u6a21\u6001\u534f\u540c\u7406\u89e3\u4e0e\u521b\u9020\u3002arXiv \u9884\u5370\u672c arXiv:2309.11499, 2023.<br \/>\n[11] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Mller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y. Marek, \u548c R. Rombach. \u7f29\u653e\u4fee\u6b63\u6d41\u53d8\u6362\u5668\u7528\u4e8e\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5408\u6210, 2024. URL https:\/\/arxiv.org\/abs\/2403.03206.<br \/>\n[12] C. Fu, P. Chen, Y. Shen, Y. Qin, M. Zhang, X. Lin, J. Yang, X. Zheng, K. Li, X. Sun, \u7b49. MME: \u4e00\u4e2a\u7528\u4e8e\u591a\u6a21\u6001\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u7efc\u5408\u8bc4\u4f30\u57fa\u51c6\u3002arXiv \u9884\u5370\u672c arXiv:2306.13394, 2023.<br \/>\n[13] Y. Ge, S. Zhao, J. Zhu, Y. Ge, K. Yi, L. Song, C. Li, X. Ding, \u548c Y. Shan. SEED-X: \u5177\u6709\u7edf\u4e00\u591a\u7c92\u5ea6\u7406\u89e3\u548c\u751f\u6210\u7684\u591a\u6a21\u6001\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2404.14396, 2024.<br \/>\n[14] D. Ghosh, H. Hajishirzi, \u548c L. Schmidt. GenEval: \u4e00\u4e2a\u9762\u5411\u5bf9\u8c61\u7684\u6846\u67b6\uff0c\u7528\u4e8e\u8bc4\u4f30\u6587\u672c\u751f\u6210\u56fe\u50cf\u5bf9\u9f50\u3002\u795e\u7ecf\u4fe1\u606f\u5904\u7406\u7cfb\u7edf\u7684\u8fdb\u5c55, 36, 2024.<br \/>\n[15] High-flyer. HAI-LLM: \u9ad8\u6548\u8f7b\u91cf\u7ea7\u7684\u5927\u578b\u6a21\u578b\u8bad\u7ec3\u5de5\u5177, 2023. URL https:\/\/www.high-flyer.cn\/en\/blog\/hai-llm.<br \/>\n[16] X. Hu, R. Wang, Y. Fang, B. Fu, P. Cheng, \u548c G. Yu. ELLA: \u4e3a\u589e\u5f3a\u8bed\u4e49\u5bf9\u9f50\u914d\u5907\u6269\u6563\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2403.05135, 2024.<br \/>\n[17] D. A. Hudson \u548c C. D. Manning. GQA: \u4e00\u4e2a\u7528\u4e8e\u771f\u5b9e\u4e16\u754c\u89c6\u89c9\u63a8\u7406\u548c\u7ec4\u5408\u95ee\u7b54\u7684\u65b0\u6570\u636e\u96c6\u3002\u5728 IEEE\/CVF \u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u6a21\u5f0f\u8bc6\u522b\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 6700\u20136709, 2019.<br \/>\n[18] Y. Jin, K. Xu, L. Chen, C. Liao, J. Tan, B. Chen, C. Lei, A. Liu, C. Song, X. Lei, \u7b49. \u52a8\u6001\u79bb\u6563\u89c6\u89c9\u6807\u8bb0\u5316\u7684\u7edf\u4e00\u8bed\u8a00\u89c6\u89c9\u9884\u8bad\u7ec3\u3002arXiv \u9884\u5370\u672c arXiv:2309.04669, 2023.<br \/>\n[19] H. Laurenon, D. van Strien, S. Bekman, L. Tronchon, L. Saulnier, T. Wang, S. Karamcheti, A. Singh, G. Pistilli, Y. Jernite, \u548c \u7b49. \u4ecb\u7ecd IDEFICS: \u4e00\u4e2a\u518d\u73b0\u6700\u5148\u8fdb\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u5f00\u653e\u6a21\u578b, 2023. URL https:\/\/huggingface.co\/blog\/id efics.<br \/>\n[20] H. Laurenon, A. Marafioti, V. Sanh, \u548c L. Tronchon. \u6784\u5efa\u5e76\u66f4\u597d\u5730\u7406\u89e3\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff1a\u6d1e\u5bdf\u4e0e\u672a\u6765\u65b9\u5411, 2024.<br \/>\n[21] B. Li, R. Wang, G. Wang, Y. Ge, Y. Ge, \u548c Y. Shan. SEED-Bench: \u4f7f\u7528\u751f\u6210\u6027\u7406\u89e3\u5bf9\u591a\u6a21\u6001 LLMs \u8fdb\u884c\u57fa\u51c6\u6d4b\u8bd5\u3002arXiv \u9884\u5370\u672c arXiv:2307.16125, 2023.<br \/>\n[22] D. Li, A. Kamko, E. Akhgari, A. Sabet, L. Xu, \u548c S. Doshi. <a href=\"https:\/\/www.kdjingpai.com\/de\/playground\/\">Playground<\/a> v2.5: \u4e09\u70b9\u6d1e\u5bdf\u7528\u4e8e\u589e\u5f3a\u6587\u672c\u751f\u6210\u56fe\u50cf\u751f\u6210\u7684\u7f8e\u5b66\u8d28\u91cf\u3002arXiv \u9884\u5370\u672c arXiv:2402.17245, 2024.<br \/>\n[23] Y. Li, Y. Du, K. Zhou, J. Wang, W. X. Zhao, \u548c J.-R. Wen. \u5728\u5927\u578b\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u4e2d\u8bc4\u4f30\u5bf9\u8c61\u5e7b\u89c9\u3002arXiv \u9884\u5370\u672c arXiv:2305.10355, 2023.<br \/>\n[24] Z. Li, H. Li, Y. Shi, A. B. Farimani, Y. Kluger, L. Yang, \u548c P. Wang. \u53cc\u6269\u6563\u7528\u4e8e\u7edf\u4e00\u56fe\u50cf\u751f\u6210\u548c\u7406\u89e3\u3002arXiv \u9884\u5370\u672c arXiv:2501.00289, 2024.<br \/>\n[25] Z. Li, J. Zhang, Q. Lin, J. Xiong, Y. Long, X. Deng, Y. Zhang, X. Liu, M. Huang, Z. Xiao, \u7b49. Hunyuan-DiT: \u4e00\u4e2a\u5f3a\u5927\u7684\u591a\u5206\u8fa8\u7387\u6269\u6563\u53d8\u6362\u5668\uff0c\u5177\u6709\u7cbe\u7ec6\u7684\u4e2d\u6587\u7406\u89e3\u3002arXiv \u9884\u5370\u672c arXiv:2405.08748, 2024.<br \/>\n[26] H. Liu, C. Li, Y. Li, \u548c Y. J. Lee. \u6539\u8fdb\u7684\u89c6\u89c9\u6307\u4ee4\u5fae\u8c03\u57fa\u7ebf\u3002\u5728 IEEE\/CVF \u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u6a21\u5f0f\u8bc6\u522b\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 26296\u201326306, 2024.<br \/>\n[27] H. Liu, C. Li, Q. Wu, \u548c Y. J. Lee. \u89c6\u89c9\u6307\u4ee4\u5fae\u8c03\u3002\u795e\u7ecf\u4fe1\u606f\u5904\u7406\u7cfb\u7edf\u7684\u8fdb\u5c55, 36, 2024.<br \/>\n[28] H. Liu, W. Yan, M. Zaharia, \u548c P. Abbeel. \u4e16\u754c\u6a21\u578b\u5728\u767e\u4e07\u957f\u5ea6\u89c6\u9891\u548c\u8bed\u8a00\u4e0a\u7684\u5e94\u7528\uff0c\u4f7f\u7528\u73af\u6ce8\u610f\u529b\u3002arXiv \u9884\u5370\u672c arXiv:2402.08268, 2024.<br \/>\n[29] Y. Liu, H. Duan, Y. Zhang, B. Li, S. Zhang, W. Zhao, Y. Yuan, J. Wang, C. He, Z. Liu, \u7b49. MMBench: \u4f60\u7684\u591a\u6a21\u6001\u6a21\u578b\u662f\u4e00\u4e2a\u5168\u80fd\u9009\u624b\u5417\uff1farXiv \u9884\u5370\u672c arXiv:2307.06281, 2023.<br \/>\n[30] Y. Ma, X. Liu, X. Chen, W. Liu, C. Wu, Z. Wu, Z. Pan, Z. Xie, H. Zhang, X. yu, L. Zhao, Y. Wang, J. Liu, \u548c C. Ruan. Janusflow: \u8c03\u548c\u81ea\u56de\u5f52\u548c\u4fee\u6b63\u6d41\u4ee5\u5b9e\u73b0\u7edf\u4e00\u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210, 2024.<br \/>\n[31] mehdidc. YFCC-Huggingface. https:\/\/huggingface.co\/datasets\/mehdidc\/yfcc15 m, 2024.<br \/>\n[32] D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Mller, J. Penna, \u548c R. Rombach. SDXL: \u6539\u8fdb\u6f5c\u5728\u6269\u6563\u6a21\u578b\u7528\u4e8e\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5408\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2307.01952, 2023.<br \/>\n[33] D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Mller, J. Penna, \u548c R. Rombach. SDXL: \u6539\u8fdb\u6f5c\u5728\u6269\u6563\u6a21\u578b\u7528\u4e8e\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5408\u6210\u30022024.<br \/>\n[34] L. Qu, H. Zhang, Y. Liu, X. Wang, Y. Jiang, Y. Gao, H. Ye, D. K. Du, Z. Yuan, \u548c X. Wu. Tokenflow: \u7edf\u4e00\u56fe\u50cf\u6807\u8bb0\u5668\u7528\u4e8e\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2412.03069, 2024.<br \/>\n[35] A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, \u548c M. Chen. \u4f7f\u7528 CLIP \u6f5c\u5728\u503c\u8fdb\u884c\u5206\u5c42\u6587\u672c\u6761\u4ef6\u56fe\u50cf\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2204.06125, 1(2):3, 2022.<br \/>\n[36] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, \u548c B. Ommer. \u4f7f\u7528\u6f5c\u5728\u6269\u6563\u6a21\u578b\u8fdb\u884c\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5408\u6210\u30022022.<br \/>\n[37] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, \u548c B. Ommer. \u4f7f\u7528\u6f5c\u5728\u6269\u6563\u6a21\u578b\u8fdb\u884c\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5408\u6210\u3002\u5728 IEEE\/CVF \u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u6a21\u5f0f\u8bc6\u522b\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 10684\u201310695, 2022.<br \/>\n[38] P. Sun, Y. Jiang, S. Chen, S. Zhang, B. Peng, P. Luo, \u548c Z. Yuan. \u81ea\u56de\u5f52\u6a21\u578b\u51fb\u8d25\u6269\u6563\uff1aLLama \u7528\u4e8e\u53ef\u6269\u5c55\u56fe\u50cf\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2406.06525, 2024.<br \/>\n[39] Q. Sun, Q. Yu, Y. Cui, F. Zhang, X. Zhang, Y. Wang, H. Gao, J. Liu, T. Huang, \u548c X. Wang. \u591a\u6a21\u6001\u751f\u6210\u6027\u9884\u8bad\u7ec3\u3002arXiv \u9884\u5370\u672c arXiv:2307.05222, 2023.<br \/>\n[40] C. Team. Chameleon: \u6df7\u5408\u6a21\u6001\u65e9\u671f\u878d\u5408\u57fa\u7840\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2405.09818, 2024.<br \/>\n[41] G. Team, R. Anil, S. Borgeaud, Y. Wu, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, \u7b49. <a href=\"https:\/\/www.kdjingpai.com\/de\/geminibardgubu\/\">Gemini<\/a>: \u4e00\u4e2a\u80fd\u529b\u5f3a\u5927\u7684\u591a\u6a21\u6001\u6a21\u578b\u5bb6\u65cf\u3002arXiv \u9884\u5370\u672c arXiv:2312.11805, 2023.<br \/>\n[42] S. Tong, D. Fan, J. Zhu, Y. Xiong, X. Chen, K. Sinha, M. Rabbat, Y. LeCun, S. Xie, \u548c Z. Liu. Metamorph: \u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\u5b9e\u73b0\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2412.14164, 2024.<br \/>\n[43] Vivym. <a href=\"https:\/\/www.kdjingpai.com\/de\/midjourney\/\">Midjourney<\/a> \u63d0\u793a\u6570\u636e\u96c6\u3002https:\/\/huggingface.co\/datasets\/vivym\/midjourney-prompts, 2023. \u8bbf\u95ee\u65e5\u671f: [\u63d2\u5165\u8bbf\u95ee\u65e5\u671f\uff0c\u4f8b\u5982 2023-10-15].<br \/>\n[44] C. Wang, G. Lu, J. Yang, R. Huang, J. Han, L. Hou, W. Zhang, \u548c H. Xu. Illume: \u7167\u4eae\u4f60\u7684 LLMs \u4ee5\u770b\u5230\u3001\u7ed8\u5236\u548c\u81ea\u6211\u589e\u5f3a\u3002arXiv \u9884\u5370\u672c arXiv:2412.06673, 2024.<br \/>\n[45] X. Wang, X. Zhang, Z. Luo, Q. Sun, Y. Cui, J. Wang, F. Zhang, Y. Wang, Z. Li, Q. Yu, \u7b49. Emu3: \u4e0b\u4e00\u4e2a\u6807\u8bb0\u9884\u6d4b\u662f\u4f60\u6240\u9700\u8981\u7684\u5168\u90e8\u3002arXiv \u9884\u5370\u672c arXiv:2409.18869, 2024.<br \/>\n[46] C. Wu, X. Chen, Z. Wu, Y. Ma, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, C. Ruan, \u7b49. Janus: \u89e3\u8026\u89c6\u89c9\u7f16\u7801\u4ee5\u5b9e\u73b0\u7edf\u4e00\u7684\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2410.13848, 2024.<br \/>\n[47] S. Wu, H. Fei, L. Qu, W. Ji, \u548c T.-S. Chua. Next-gpt: \u4efb\u4f55\u5230\u4efb\u4f55\u7684\u591a\u6a21\u6001 LLM\u3002arXiv \u9884\u5370\u672c arXiv:2309.05519, 2023.<br \/>\n[48] Y. Wu, Z. Zhang, J. Chen, H. Tang, D. Li, Y. Fang, L. Zhu, E. Xie, H. Yin, L. Yi, \u7b49. VILA-U: \u4e00\u4e2a\u6574\u5408\u89c6\u89c9\u7406\u89e3\u548c\u751f\u6210\u7684\u57fa\u7840\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2409.04429, 2024.<br \/>\n[49] Z. Wu, X. Chen, Z. Pan, X. Liu, W. Liu, D. Dai, H. Gao, Y. Ma, C. Wu, B. Wang, \u7b49. DeepSeek-VL2: \u6df7\u5408\u4e13\u5bb6\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7528\u4e8e\u9ad8\u7ea7\u591a\u6a21\u6001\u7406\u89e3\u3002arXiv \u9884\u5370\u672c arXiv:2412.10302, 2024.<br \/>\n[50] J. Xie, W. Mao, Z. Bai, D. J. Zhang, W. Wang, K. Q. Lin, Y. Gu, Z. Chen, Z. Yang, \u548c M. Z. Shou. Show-o: \u4e00\u4e2a\u5355\u4e00\u7684\u53d8\u6362\u5668\u7edf\u4e00\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u3002arXiv \u9884\u5370\u672c arXiv:2408.12528, 2024.<br \/>\n[51] W. Yu, Z. Yang, L. Li, J. Wang, K. Lin, Z. Liu, X. Wang, \u548c L. Wang. MM-Vet: \u8bc4\u4f30\u5927\u578b\u591a\u6a21\u6001\u6a21\u578b\u7684\u7efc\u5408\u80fd\u529b\u3002arXiv \u9884\u5370\u672c arXiv:2308.02490, 2023.<br \/>\n[52] X. Yue, Y. Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y. Sun, \u7b49. MMMU: \u4e00\u4e2a\u5927\u89c4\u6a21\u591a\u5b66\u79d1\u591a\u6a21\u6001\u7406\u89e3\u548c\u63a8\u7406\u57fa\u51c6\uff0c\u7528\u4e8e\u4e13\u5bb6 AGI\u3002\u5728 IEEE\/CVF \u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u6a21\u5f0f\u8bc6\u522b\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 9556\u20139567, 2024.<br \/>\n[53] X. Zhai, B. Mustafa, A. Kolesnikov, \u548c L. Beyer. \u8bed\u8a00\u56fe\u50cf\u9884\u8bad\u7ec3\u7684 Sigmoid \u635f\u5931\u3002\u5728 IEEE\/CVF \u56fd\u9645\u8ba1\u7b97\u673a\u89c6\u89c9\u4f1a\u8bae\u8bba\u6587\u96c6, \u9875\u7801 11975\u201311986, 2023.<br \/>\n[54] C. Zhao, Y. Song, W. Wang, H. Feng, E. Ding, Y. Sun, X. Xiao, \u548c J. Wang. Monoformer: \u4e00\u4e2a\u7528\u4e8e\u6269\u6563\u548c\u81ea\u56de\u5f52\u7684\u5355\u4e00\u53d8\u6362\u5668\u3002arXiv \u9884\u5370\u672c arXiv:2409.16280, 2024.<br \/>\n[55] C. Zhou, L. Yu, A. Babu, K. Tirumala, M. Yasunaga, L. Shamis, J. Kahn, X. Ma, L. Zettlemoyer, \u548c O. Levy. Transfusion: \u4f7f\u7528\u4e00\u4e2a\u591a\u6a21\u6001\u6a21\u578b\u9884\u6d4b\u4e0b\u4e00\u4e2a\u6807\u8bb0\u5e76\u6269\u6563\u56fe\u50cf\u3002arXiv \u9884\u5370\u672c arXiv:2408.11039, 2024.<br \/>\n[56] Y. Zhu, M. Zhu, N. Liu, Z. Ou, X. Mou, \u548c J. Tang. LLAVA-Phi: \u9ad8\u6548\u7684\u591a\u6a21\u6001\u52a9\u624b\uff0c\u5e26\u6709\u5c0f\u8bed\u8a00\u6a21\u578b\u3002arXiv \u9884\u5370\u672c arXiv:2401.02330, 2024.<\/p>\n<p>[57] L. Zhuo, R. Du, H. Xiao, Y. Li, D. Liu, R. Huang, W. Liu, L. Zhao, F.-Y. Wang, Z. Ma, \u7b49. Lumina-Next: \u4f7f\u7528 Next-DiT \u4f7f Lumina-T2X \u66f4\u5f3a\u5927\u3001\u66f4\u5feb\u901f\u3002arXiv \u9884\u5370\u672c arXiv:2406.18583, 2024.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>JanusFlow \u901f\u8bfb DeepSeek \u56e2\u961f\u53c8\u53d1\u65b0\u6a21\u578b\uff0c28\u65e5\u51cc\u6668\u63a8\u51fa\u4e86\u521b\u65b0\u578b\u591a\u6a21\u6001\u6846\u67b6 Janus-Pro\uff0c\u8fd9\u662f\u4e00\u4e2a\u80fd\u540c\u65f6\u5904\u7406\u591a\u6a21\u6001\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u7684\u7edf\u4e00\u6a21\u578b\u3002\u8be5\u6a21\u578b\u57fa\u4e8e DeepSeek-LLM-1.5b-base\/DeepSeek-&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46],"tags":[],"class_list":["post-19557","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/19557","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=19557"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/19557\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=19557"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=19557"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=19557"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}