{"id":15466,"date":"2023-11-01T10:39:39","date_gmt":"2023-11-01T02:39:39","guid":{"rendered":"http:\/\/139.9.1.231\/?p=15466"},"modified":"2023-11-01T10:39:41","modified_gmt":"2023-11-01T02:39:41","slug":"real-time-radiance-fields-for-single-image-portrait-view-synthesi","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/11\/01\/real-time-radiance-fields-for-single-image-portrait-view-synthesi\/","title":{"rendered":"Real-Time Radiance Fields for Single-Image Portrait View Synthesis \u8bba\u6587"},"content":{"rendered":"\n<p>\u9009\u81ea<strong>SIGGRAPH 2023<\/strong>\u5b9e\u65f6\u6e32\u67d3\u9886\u57df\u8bba\u6587\u3002\u6587\u7ae0\u5b9e\u73b0\u4e86\u6700\u65b0\u7684\u5355\u56fe\u50cf\u5b9e\u65f6\u5408\u6210\u4e09\u7ef4\u89c6\u89d2\u7684\u6280\u672f\u3002<\/p>\n\n\n\n<p><strong>\u8bba\u6587\u9898\u76ee<\/strong>\uff1aLive 3D Portrait: Real-Time Radiance Fields for Single-Image Portrait View Synthesis<br><strong>\u8bba\u6587\u94fe\u63a5<\/strong>\uff1ahttps:\/\/research.nvidia.com\/labs\/nxp\/lp3d\/<\/p>\n\n\n\n<p><a rel=\"noreferrer noopener\" href=\"https:\/\/research.nvidia.com\/labs\/nxp\/lp3d\/media\/paper.pdf\" target=\"_blank\">Paper PDF<\/a><\/p>\n\n\n\n<p>   \u672c\u6587\u63d0\u51fa\u4e86\u4ece\u5355\u5f20\u56fe\u50cf\u5b9e\u65f6\u63a8\u7406\u6e32\u67d3\u7167\u7247\u7ea7 3D \u8868\u793a\u7684\u5355\u6837\u672c\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u7ed9\u5b9a\u5355\u5f20 RGB \u8f93\u5165\u56fe\u50cf\u540e\uff0c\u7f16\u7801\u5668\u76f4\u63a5\u9884\u6d4b\u795e\u7ecf\u8f90\u5c04\u573a\u7684\u89c4\u8303\u5316\u4e09\u5e73\u9762\u8868\u793a\uff0c\u4ece\u800c\u901a\u8fc7\u4f53\u6e32\u67d3\u5b9e\u73b0 3D \u611f\u77e5\u7684\u65b0\u89c6\u56fe\u5408\u6210\u3002\u8be5\u65b9\u6cd5\u4ec5\u4f7f\u7528\u5408\u6210\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u901a\u8fc7\u7ed3\u5408\u57fa\u4e8e Transformer \u7684\u7f16\u7801\u5668\u548c\u6570\u636e\u589e\u5f3a\u7b56\u7565\uff0c\u53ef\u4ee5\u5904\u7406\u73b0\u5b9e\u4e16\u754c\u4e2d\u5177\u6709\u6311\u6218\u6027\u7684\u8f93\u5165\u56fe\u50cf\uff0c\u5e76\u4e14\u65e0\u9700\u4efb\u4f55\u7279\u6b8a\u5904\u7406\u5373\u53ef\u9010\u5e27\u5e94\u7528\u4e8e\u89c6\u9891\u3002<\/p>\n\n\n\n<h2 id=\"h_651287716_0\">INTRODUCTION<\/h2>\n\n\n\n<p>\u968f\u7740<strong>NeRF<\/strong>\u7684\u63d0\u51fa\uff0c\u4e09\u7ef4\u89c6\u89c9\u6280\u672f\u5f97\u5230\u5feb\u901f\u7684\u53d1\u5c55\u3002\u4e09\u7ef4\u91cd\u5efa\u4e5f\u662f\u975e\u5e38\u6709\u610f\u4e49\u7684\u5de5\u4f5c\uff0c\u5176\u4e2d\uff0c\u5355\u5f20\u8096\u50cf\u5b9e\u73b0\u5b9e\u65f6\u4e09\u7ef4\u89c6\u89d2\u7684\u5408\u6210\u5c06\u63a8\u52a8AR\u3001VR\u30013D\u8fdc\u7a0b\u4f1a\u8bae\u7684\u53d1\u5c55\u3002<\/p>\n\n\n\n<p>\u57fa\u4e8e\u6b64\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u8be5\u6280\u672f\u7684\u6700\u65b0\u65b9\u6cd5\uff0c\u8be5\u6280\u672f\u7684\u539f\u6587\u8868\u8ff0\u662f<em>infer and render a photorealistic 3D representation from a single unposed image (e.g., face portrait) in real-time.<\/em><\/p>\n\n\n\n<p>\u5148\u6765\u770b\u73b0\u6709\u7684\u65b9\u6cd5\uff0c\u4e00\u822c\u7528<strong>NeRF+GANs<\/strong>\u5b9e\u73b03D\u611f\u77e5\u56fe\u50cf\u751f\u6210\u3002\u5176\u4e2d\u6bd4\u8f83\u6709\u540d\u7684\u4e00\u9879\u6280\u672f\u662f<strong>EG3D<\/strong>\uff0cEG3D\u7684\u63d0\u51fa\u8005\u4e5f\u662f\u672c\u8bba\u6587\u7684\u5171\u540c\u4f5c\u8005\u4e4b\u4e00\uff0c\u672c\u6587\u7684\u5de5\u4f5c\u662f\u5728EG3D\u7684\u57fa\u7840\u4e0a\u5c55\u5f00\u7684\u3002<\/p>\n\n\n\n<p>EG3D\u63d0\u51fa\u4e86\u4e00\u79cd\u9ad8\u6548\u7684\u4e09\u5e73\u97623D\u8868\u793a<em>(triplane 3D representation)<\/em>\uff08\u5177\u4f53\u7ec6\u8282\u4f1a\u5728\u540e\u7eed\u7ed9\u51fa\uff09\uff0c\u5e76\u4e14\u80fd\u591f\u8fbe\u5230\u4e0e<em>2D GANs\u76f8\u540c\u7684<\/em>\u5b9e\u65f6\u6e32\u67d3\u8d28\u91cf\u3002\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u6d4b\u8bd5\u65f6\u5fae\u8c03<em>(test-time fine tuning)<\/em>\u5b8c\u6210\u5355\u56fe\u50cf\u4e09\u7ef4\u91cd\u5efa\u3002\u4f46\u8fd9\u79cd\u65b9\u6cd5\u4f1a\u6709\u4e00\u4e9b\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul><li>NeRF\u7684\u8bad\u7ec3\u901a\u5e38\u9700\u8981\u4f18\u5316\u76ee\u6807<em>(careful optimization objectives)<\/em>\u548c3D\u5148\u9a8c<em>(additional 3D priors)<\/em><\/li><li>\u6d4b\u8bd5\u65f6\u4f18\u5316\u9700\u8981\u51c6\u786e\u7684\u76f8\u673a\u59ff\u6001\u4f5c\u4e3a\u8f93\u5165\u6216\u4f18\u5316\u76f8\u673a\u59ff\u6001<\/li><li>\u4e0a\u8ff0\u4e24\u70b9\u4f18\u5316\u65f6\u8017\u65f6\u7684\uff0c\u9650\u5236\u4e86\u5b9e\u65f6\u5e94\u7528<\/li><\/ul>\n\n\n\n<p>\u4e0e\u4ee5\u5f80\u91cd\u590d\u4f7f\u7528\u9884\u8bad\u7ec3\u7684generator\u4e0d\u540c\uff0c\u672c\u7bc7\u8bba\u6587\u8bad\u7ec3\u4e86\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u7f16\u7801\u5668<em>\uff08encoder end-to-end\uff09<\/em>\u7528\u4e8e\u76f4\u63a5\u4ece\u5355\u4e2a\u8f93\u5165\u56fe\u50cf\u9884\u6d4b\u4e09\u5e73\u97623D\u7279\u5f81\u3002\u4e0e\u4ee5\u5f80\u4f9d\u8d56\u4e8e\u591a\u89c6\u56fe\u771f\u5b9e\u56fe\u50cf\u7684\u91c7\u96c6\u76f8\u6bd4\uff0c\u672c\u6587\u4e0d\u9700\u8981\u83b7\u53d6\u771f\u5b9e\u56fe\u50cf\uff0c\u4e5f\u4e0d\u9700\u8981PBR<em>(physically-based rendering)<\/em>\u7ed8\u5236\u90a3\u6837\u8017\u65f6\u3002\u76f8\u53cd\uff0c\u4f5c\u8005\u4f7f\u7528\u9884\u8bad\u7ec3\u76843D GAN\u751f\u6210\u7684\u591a\u89c6\u56fe\u4e00\u81f4\u7684\u5408\u6210\u6570\u636e\u6765\u76d1\u7763\u4e09\u5e73\u9762\u7f16\u7801\u5668\uff0c\u4ee5\u4fbf\u8fdb\u884c\u65b0\u89c6\u56fe\u5408\u6210\uff0c\u518d\u7ed3\u5408\u6570\u636e\u589e\u5f3a\u7b56\u7565\u548c\u57fa\u4e8eTransformer\u7684\u7f16\u7801\u5668\u642d\u5efa\u597d\u6a21\u578b\u3002\u5728\u6587\u7ae0\u4e2d\u4f5c\u8005\u5c55\u793a\u4e86\u5bf9\u4eba\u8138\u548c\u732b\u8138\u4e09\u7ef4\u91cd\u5efa\u7684\u7ed3\u679c\uff0c\u4f46\u4f5c\u8005\u8868\u793a\u4efb\u4f553D\u611f\u77e5\u56fe\u50cfgenerator\u9002\u7528\u7684\u7c7b\u522b\uff0c\u8be5\u6a21\u578b\u540c\u6837\u9002\u7528\u3002<\/p>\n\n\n\n<p>\u6982\u62ec\u4e0b\u6587\u7ae0\u7684\u5de5\u4f5c\u8d21\u732e\uff1a<\/p>\n\n\n\n<ol><li>\u63d0\u51fa\u4e86\u4e00\u79cd\u524d\u9988\u7f16\u7801\u5668\u6a21\u578b\uff0c\u76f4\u63a5\u4ece\u8f93\u5165\u56fe\u50cf\u63a8\u65ad\u4e09\u5e73\u97623D\u8868\u793a\u3002\u4e0d\u9700\u8981\u6d4b\u8bd5\u65f6\u4f18\u5316\u3002<\/li><li>\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u7b56\u7565\uff0c\u4ec5\u4f7f\u7528\u4ece\u9884\u8bad\u7ec3\u76843D\u611f\u77e5\u56fe\u50cf\u751f\u6210\u5668\u751f\u6210\u7684\u5408\u6210\u6570\u636e<\/li><li>\u7ed3\u5408\u57fa\u4e8eTransformer\u7684\u7f16\u7801\u5668\u548c\u5b9e\u65f6\u589e\u5f3a\u7b56\u7565\uff0c\u8be5\u65b9\u6cd5\u53ef\u4ee5\u5904\u7406\u5177\u6709\u6311\u6218\u6027\u7684\u8f93\u5165\u56fe\u50cf\u3002<\/li><\/ol>\n\n\n\n<h2 id=\"h_651287716_1\">2. RELATED WORK<\/h2>\n\n\n\n<h3 id=\"h_651287716_2\">2.1 Light Fields and Image-Based Rendering<\/h3>\n\n\n\n<p>\u4f20\u7edf\u7684\u65b9\u6cd5\u8981\u4e48\u9700\u8981\u8bb8\u591a\u89c6\u56fe\u6837\u672c\uff0c\u8981\u4e48\u9700\u8981\u5149\u573a\u76f8\u673a\u4f5c\u4e3a\u8bad\u7ec3\u6570\u636e\u3002\u6700\u8fd1\u63d0\u51fa\u7684NeRF\u7ed3\u54083D\u9690\u5f0f\u8868\u793a\uff0c\u8fd0\u7528\u4f53\u6e32\u67d3\u7684\u65b9\u5f0f\u5408\u6210\u89c6\u56fe\uff0c\u4f46\u4ecd\u9700\u8981\u5927\u91cf\u8f93\u5165\u7167\u7247\u3002<\/p>\n\n\n\n<h3 id=\"h_651287716_3\">2.2 Few-shot novel view synthesis<\/h3>\n\n\n\n<p>\u6700\u8fd1\u4e00\u4e9b\u6269\u5c55NeRF\u7684\u5de5\u4f5c\u75283D\u9690\u5f0f\u8868\u793a\u5b8c\u6210\u4e86\u5355\u56fe\u50cf\u5408\u6210\uff0c\u7528\u52303D\u5377\u79ef\u3001Transformers \u7b49\u65b9\u6cd5\u3002\u4f46\u662f\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u4e0d\u662f\u5b9e\u65f6\u751f\u6210\u65b0\u89c6\u56fe\u7684\uff0c\u5e76\u4e14\u90fd\u9700\u8981\u591a\u89c6\u56fe\u56fe\u50cf\u6765\u8bad\u7ec3\u6a21\u578b\u3002\u800c\u4f5c\u8005\u7684\u65b9\u6cd5\u53ea\u9700\u8981\u4ece\u9884\u5148\u8bad\u7ec3\u76843D GAN\u751f\u6210\u7684\u5408\u6210\u56fe\u50cf\uff0c\u8fd9\u79cd3D GAN\u662f\u7531\u5355\u89c6\u56fe\u56fe\u50cf\u7684\u96c6\u5408\u8bad\u7ec3\u7684\u3002<\/p>\n\n\n\n<h3 id=\"h_651287716_4\">2.3 Learning with synthetic data<\/h3>\n\n\n\n<p>\u5f53\u6ca1\u6709\u57fa\u51c6\u771f\u5b9e\u6570\u636e<em>(ground truth data )<\/em>\u65f6\uff0c\u5408\u6210\u6570\u636e\u4e3a\u8bad\u7ec3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u63d0\u4f9b\u4e86\u6709\u7528\u7684\u76d1\u7763\u3002\u8fd9\u5f80\u5f80\u8fd8\u9700\u8981\u989d\u5916\u7684\u6b65\u9aa4\u6765\u9002\u5e94\u771f\u5b9e\u56fe\u50cf\u3002<\/p>\n\n\n\n<h3 id=\"h_651287716_5\">2.4 3D-aware portrait generation and manipulation<\/h3>\n\n\n\n<p>\u6700\u8fd1\uff0c3D\u611f\u77e5\u56fe\u50cf\u751f\u6210\u65b9\u6cd5\u5f00\u59cb\u89e3\u51b3\u4ece\u5355\u89c6\u56fe2D\u56fe\u50cf\u96c6\u5408\u4e2d\u65e0\u6761\u4ef6\u751f\u6210\u903c\u771f\u76843D\u8868\u793a\u7684\u95ee\u9898\u3002\u7ed3\u5408\u795e\u7ecf\u4f53\u79ef\u6e32\u67d3<em>(neural volumetric rendering)<\/em>\u548c\u751f\u6210\u5bf9\u6297\u7f51\u7edc<em>(GANs)<\/em>\uff0c\u6700\u65b0\u7684<em>3D GAN<\/em>\u65b9\u6cd5\u80fd\u591f\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u591a\u89c6\u56fe\u4e00\u81f4\u56fe\u50cf\u3002\u4f5c\u8005\u91c7\u7528EG3D \u7684\u4e09\u5e73\u97623D\u8868\u793a\uff0c\u5b9e\u73b0\u5355\u89c6\u56fe\u65b0\u89c6\u56fe\u5408\u6210\u3002<\/p>\n\n\n\n<h3 id=\"h_651287716_6\">2.5 3D GAN inversion<\/h3>\n\n\n\n<p><em>GAN inversion<\/em>\u57282D\u9886\u57df\u53d6\u5f97\u5f88\u5927\u8fdb\u5c55\uff0c\u73b0\u6709\u7684<em>3D GAN inversion<\/em>\u65b9\u6cd5\u5c06\u7ed9\u5b9a\u7684\u56fe\u50cf\u6295\u5f71\u5230\u9884\u8bad\u7ec3\u7684<em>StyleGAN2 latent space<\/em>\u4e0a\uff0c\u5e76\u4e14\u5728\u6d4b\u8bd5\u65f6\u9700\u8981\u6444\u50cf\u673a\u59ff\u6001<em>( approximate camera pose )<\/em>\u548c\u751f\u6210\u5668\u6743\u91cd\u5fae\u8c03<em>( generator weight tuning)<\/em>\uff0c\u4ee5\u91cd\u5efa\u57df\u5916\u8f93\u5165\u56fe\u50cf\u3002\u4e0e\u540c\u65f6\u671f\u7684\u5de5\u4f5c\u4e0d\u540c\uff0c\u4f5c\u8005\u7684\u524d\u9988\u7f16\u7801\u5668\u5c06\u672a\u5b9a\u4f4d\u7684\u56fe\u50cf\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u4e14\u4e0d\u9700\u8981\u9488\u5bf9\u6444\u50cf\u673a\u59ff\u6001\u7684\u6d4b\u8bd5\u65f6\u4f18\u5316\u3002<\/p>\n\n\n\n<h3 id=\"h_651287716_7\">2.6 Talking-head generators<\/h3>\n\n\n\n<p>\u7ed9\u51fa\u5355\u4e2a\u76ee\u6807\u8096\u50cf\u548c\u9a71\u52a8\u89c6\u9891\uff0c\u8fd9\u79cdTalking-head\u751f\u6210\u65b9\u6cd5\u4e3b\u8981\u901a\u8fc7\u89c6\u9891\u6570\u636e\u96c6\u8bad\u7ec3\uff0c\u4fa7\u91cd\u4e8e\u901a\u8fc7\u64cd\u7eb52D\u8096\u50cf\u4e2d\u7684\u5934\u50cf\u59ff\u52bf\u548c\u8868\u60c5\u6765\u751f\u6210talking-head\u89c6\u9891\u3002\u56e0\u6b64\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u9884\u6d4b\u89c6\u70b9\u6e32\u67d3\u7684\u4f53\u79ef\u8868\u793a\u548c\u4e09\u7ef4\u51e0\u4f55\u4fe1\u606f\u3002\u6240\u4ee5\u4e0d\u4e88\u6bd4\u8f83\u3002<\/p>\n\n\n\n<h2 id=\"h_651287716_8\">3. PRELIMINARIES: TRIPLANE-BASED 3D GAN<\/h2>\n\n\n\n<p>NeRF\u91c7\u7528\u5b8c\u5168\u9690\u5f0f\u7684\u8868\u793a\uff0c\u4f7f\u7528\u795e\u7ecf\u7f51\u7edc\u6765\u8868\u793a\u6574\u4e2a\u4e09\u7ef4\u7a7a\u95f4\u7684\u8f90\u5c04\u573a\uff0c\u4f46\u8ba1\u7b97\u5f80\u5f80\u9700\u8981\u82b1\u8d39\u5927\u91cf\u65f6\u95f4\u3002\u9996\u5148\uff0c\u5bf9\u524d\u6cbf\u76843D GAN\u65b9\u6cd5EG3D\u8fdb\u884c\u6982\u8ff0\u3002EG3D\u4ece\u5355\u89c6\u56fe\u56fe\u50cf\u96c6\u5408\u548c\u76f8\u5e94\u7684\u566a\u58f0\u76f8\u673a\u59ff\u52bf\u4e2d\u5b66\u4e603D\u611f\u77e5\u56fe\u50cf\u751f\u6210\uff0cEG3D\u4f7f\u7528\u6df7\u5408\u4e09\u5e73\u9762\u8868\u793a\u6765\u8c03\u8282\u795e\u7ecf\u4f53\u79ef\u6e32\u67d3\u8fc7\u7a0b\uff0c\u5176\u4e2d\u4e09\u4e2a\u5178\u578b\u5e73\u9762&nbsp;\ufffd\ufffd,\ufffd\ufffd,\ufffd\ufffd&nbsp;\u90fd\u5b58\u50a8\u4e86\u4e09\u4e2a2D\u7279\u5f81\u7f51\u683c&nbsp;<em>\uff08feature grids\uff09<\/em>\u3002\u4f7f\u7528StyleGAN2\u751f\u6210\u5668\uff0cEG3D\u5c06\u566a\u58f0\u5411\u91cf\u548c\u76f8\u673a\u59ff\u52bf\u6620\u5c04\u5230\u4e09\u5e73\u9762\u8868\u793a&nbsp;\ufffd\u2208R256\u00d7256\u00d796\uff0c\u5bf9\u5e94\u4e8e3\u4e2a\u8f74\u5bf9\u9f50\u7684\u5e73\u9762\uff0c\u6bcf\u4e2a\u5e73\u9762\u5177\u670932\u4e2a\u901a\u9053\u3002\u8fd9\u4e9b\u7279\u5f81\u8c03\u8282\u795e\u7ecf\u4f53\u79ef\u6e32\u67d3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic3.zhimg.com\/v2-6f65b00d69302afb05ddd8adff58d656_r.jpg\" alt=\"\"\/><figcaption>Our hybrid explicit\u2013implicit tri-plane representation (c) is fast and scales efficiently with resolution, enabling <br>greater detail for equal capacity.<\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"379\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-6-1024x379.png\" alt=\"\" class=\"wp-image-15484\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-6-1024x379.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-6-300x111.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-6-768x284.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-6.png 1496w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u5c06\u7279\u5f81\u5b58\u50a8\u5728\u6b63\u4ea4\u7684\u4e09\u5e73\u9762<em>(triplane)<\/em>\u8868\u793a\u4e2d\uff0c\u901a\u8fc7\u7279\u5f81\u503c\u53e0\u52a0\u8ba1\u7b97\u51fa\u7279\u5b9a\u7a7a\u95f4\u70b9\u7684\u989c\u8272\u3001\u4f53\u79ef\u5bc6\u5ea6\uff0c\u901a\u8fc7NeRF\u8fdb\u884c\u8bad\u7ec3\uff0c\u8bad\u7ec3\u5f97\u5230\u7684\u53c2\u6570\u4e5f\u4fdd\u5b58\u5728\u4e09\u5e73\u9762\u8868\u793a\u4e2d\u3002<\/p>\n\n\n\n<h2 id=\"h_651287716_9\">4.&nbsp;METHOD<\/h2>\n\n\n\n<p>\u4f5c\u8005\u7684\u76ee\u6807\u662f\u5c06EG3D\u751f\u6210\u6a21\u578b\u7684\u4fe1\u606f\u63d0\u70bc\u5230\u4e00\u4e2a\u524d\u9988\u7f16\u7801\u5668\u7684pipline\u4e2d\uff0c\u8fd9\u53ef\u4ee5\u76f4\u63a5\u5c06\u672a\u5b9a\u4f4d\u7684\u56fe\u50cf\u6620\u5c04\u5230\u4e00\u4e2a\u89c4\u8303\u7684\u4e09\u5e73\u97623D\u8868\u793a\uff0c\u8fd9\u91cc\u7684\u89c4\u8303\u8868\u793a\uff0c\u5bf9\u4e8e\u4eba\u8138\uff0c\u5934\u90e8\u7684\u4e2d\u5fc3\u662f\u539f\u70b9\u3002\u8be5pipline\u4ec5\u9700\u8981\u5355\u6b21\u524d\u9988\u7f51\u7edc\u4f20\u9012\uff0c\u4ece\u800c\u907f\u514d\u4e86\u82b1\u9500\u5927\u7684 GAN inversion\u8fc7\u7a0b\uff0c\u540c\u65f6\u5141\u8bb8\u5b9e\u65f6\u91cd\u65b0\u6e32\u67d3\u8f93\u5165\u7684\u4efb\u610f\u89c6\u70b9\u3002<\/p>\n\n\n\n<p>\u4f5c\u8005\u7684\u5de5\u4f5c\u4e3b\u8981\u96c6\u4e2d\u5728\u56fe\u50cf\u5230\u4e09\u5e73\u9762\u7f16\u7801\u5668\u548c\u76f8\u5173\u7684\u5408\u6210\u8bad\u7ec3\u65b9\u6cd5\u4e0a\uff0c\u4f7f\u7528EG3D\u7684MLP\u4f53\u79ef\u6e32\u67d3\u5668\u548c\u8d85\u5206\u8fa8\u7387\u67b6\u6784\uff0c\u7aef\u5230\u7aef\u5730\u8bad\u7ec3\u6240\u6709\u7ec4\u4ef6\u3002\u4e0b\u56fe\u662f\u6574\u4e2a\u6a21\u578b\u7684\u63a8\u7406\u548c\u8bad\u7ec3\u90e8\u5206\uff0c\u662f\u6587\u7ae0\u7684\u91cd\u70b9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic4.zhimg.com\/v2-38d952d27f342832dcfab3270950dfc7_r.jpg\" alt=\"\"\/><figcaption>\u56fe3\u56fe 2\uff1a\u63a8\u7406\u548c\u8bad\u7ec3\u7ba1\u7ebf\u3002\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u6211\u4eec\u4ee5\u5355\u5f20\u56fe\u50cf\u4f5c\u4e3a\u8f93\u5165\uff0c\u4f7f\u7528 DeepLabV3 \u63d0\u53d6\u4f4e\u5206\u8fa8\u7387\u7279\u5f81\u3002\u8fd9\u4e9b\u7279\u5f81\u7ecf\u8fc7 ViT \u548c\u5377\u79ef\u8f93\u51fa\uff0c\u4e0e\u9ad8\u5206\u8fa8\u7387\u7279\u5f81\u4e32\u8054\uff0c\u518d\u901a\u8fc7 ViT \u548c\u5377\u79ef\u89e3\u7801\u4e3a\u4e09\u5e73\u9762\u8868\u793a\uff0c\u4ece\u800c\u4e3a\u4f53\u6e32\u67d3\u8fc7\u7a0b\u63d0\u4f9b\u6761\u4ef6\uff0c\u751f\u6210\u6df1\u5ea6\u3001\u7279\u5f81\u3001\u989c\u8272\u548c\u8d85\u5206\u8fa8\u7387\u56fe\u50cf\u3002\u5728\u8bad\u7ec3\u9636\u6bb5\uff0c\u6211\u4eec\u4ece EG3D \u4e2d\u91c7\u6837\u4e00\u4e2a\u8eab\u4efd\uff0c\u6e32\u67d3\u4e24\u4e2a\u76d1\u7763\u89c6\u56fe\u3002\u7b2c\u4e00\u4e2a\u89c6\u56fe\u4f5c\u4e3a\u7f16\u7801\u5668\u8f93\u5165\uff0c\u9884\u6d4b\u4e09\u5e73\u9762\uff0c\u7136\u540e\u6839\u636e\u8fd9\u4e24\u4e2a\u89c6\u89d2\u8fdb\u884c\u4f53\u6e32\u67d3\uff0c\u5e76\u5c06\u6e32\u67d3\u7ed3\u679c\u4e0e EG3D \u7684\u7ed3\u679c\u8fdb\u884c\u6bd4\u8f83\u4f18\u5316\u3002<\/figcaption><\/figure>\n\n\n\n<p>\u6211\u4eec\u7684\u76ee\u6807\u662f\u5c06\u8bad\u7ec3\u597d\u7684 EG3D \u751f\u6210\u6a21\u578b\u77e5\u8bc6\u84b8\u998f\u81f3\u524d\u9988\u7f16\u7801\u7ba1\u7ebf\uff0c\u8be5\u7ba1\u7ebf\u53ea\u9700\u4e00\u6b21\u524d\u9988\u7f51\u7edc\u4f20\u64ad\u5373\u53ef\u5c06\u5355\u5f20\u56fe\u50cf\u76f4\u63a5\u6620\u5c04\u4e3a\u89c4\u8303\u7684\u4e09\u5e73\u9762 3D \u8868\u793a\uff0c\u540c\u65f6\u5141\u8bb8\u5bf9\u8f93\u5165\u5728\u81ea\u7531\u89c6\u89d2\u4e0b\u8fdb\u884c\u5b9e\u65f6\u6e32\u67d3\u3002\u6211\u4eec\u7684\u8d21\u732e\u96c6\u4e2d\u4e8e\u56fe\u50cf\u5230\u4e09\u5e73\u9762\u7f16\u7801\u5668\u548c\u76f8\u5173\u7684\u5408\u6210\u6570\u636e\u8bad\u7ec3\u65b9\u6cd5\u3002\u6211\u4eec\u4f7f\u7528 EG3D \u4e2d\u7684 MLP \u4f53\u6e32\u67d3\u5668\u548c\u8d85\u5206\u8fa8\u7387\u67b6\u6784\uff0c\u5e76\u5bf9\u6240\u6709\u7ec4\u4ef6\u8fdb\u884c\u7aef\u5230\u7aef\u7684\u8bad\u7ec3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"955\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7-1024x955.png\" alt=\"\" class=\"wp-image-15486\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7-1024x955.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7-300x280.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7-768x716.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7-1536x1432.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-7.png 1740w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"493\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8-1024x493.png\" alt=\"\" class=\"wp-image-15488\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8-1024x493.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8-300x144.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8-768x370.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8-1536x739.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-8.png 1762w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u9009\u81eaSIGGRAPH 2023\u5b9e\u65f6\u6e32\u67d3\u9886\u57df\u8bba\u6587\u3002\u6587\u7ae0\u5b9e\u73b0\u4e86\u6700\u65b0\u7684\u5355\u56fe\u50cf\u5b9e\u65f6\u5408\u6210\u4e09\u7ef4\u89c6\u89d2\u7684\u6280\u672f\u3002 \u8bba\u6587\u9898\u76ee\uff1aLi &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2023\/11\/01\/real-time-radiance-fields-for-single-image-portrait-view-synthesi\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">Real-Time Radiance Fields for Single-Image Portrait View Synthesis \u8bba\u6587<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,35],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15466"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=15466"}],"version-history":[{"count":21,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15466\/revisions"}],"predecessor-version":[{"id":15490,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15466\/revisions\/15490"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=15466"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=15466"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=15466"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}