{"id":1599,"date":"2020-02-28T18:27:14","date_gmt":"2020-02-28T10:27:14","guid":{"rendered":"https:\/\/www.deepfaker.xyz\/?page_id=1599"},"modified":"2020-02-28T18:27:14","modified_gmt":"2020-02-28T10:27:14","slug":"deepfacelab-2019%e5%b9%b4%e5%8e%86%e5%8f%b2%e7%89%88%e6%9c%ac%e6%9b%b4%e6%96%b0%e8%ae%b0%e5%bd%95%ef%bc%81","status":"publish","type":"page","link":"https:\/\/www.deepfaker.xyz\/?page_id=1599","title":{"rendered":"DeepFaceLab 2019\u5e74\u5386\u53f2\u7248\u672c\u66f4\u65b0\u8bb0\u5f55\uff01"},"content":{"rendered":"<h2>2019\u5e7412\u670829\u65e5<\/h2>\n<p>fix faceset enhancer for faces that contain edited mask<br \/>\n\u4eba\u8138\u589e\u5f3a\u9002\u7528\u4e8e\u5e26\u8499\u7248\u4fe1\u606f\u7684\u56fe\u7247<\/p>\n<p>fix long load when using various gpus in the same DFL folder<br \/>\n\u4fee\u590d\u5728\u540c\u4e00DFL\u6587\u4ef6\u5939\u4e2d\u4f7f\u7528\u4e0d\u540cgpu\u65f6\u7684\u957f\u8d1f\u8f7d<\/p>\n<p>fix extract unaligned faces<br \/>\n\u4fee\u590d\u63d0\u53d6\u672a\u5bf9\u9f50\u7684\u4eba\u8138\u529f\u80fd<\/p>\n<p>avatar: avatar_type is now only head by default<br \/>\n\u963f\u51e1\u8fbe\uff1a\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u963f\u51e1\u8fbe\u7c7b\u578b\u73b0\u5728\u53ea\u6709\u5934\u90e8<\/p>\n<h2>2019\u5e7412\u670828\u65e5<\/h2>\n<p>FacesetEnhancer now asks to merge aligned_enhanced\/ to aligned\/<br \/>\nFacesetEnhancer \u73b0\u5728\u5c06aligned_enhanced\u5408\u5e76\u5230\u4e86aligned\/<\/p>\n<p>fix 0 faces detected in manual extractor<br \/>\n\u4fee\u590d\u624b\u52a8\u63d0\u53d6\u4e3a\u96f6\u7684\u95ee\u9898\u3002<\/p>\n<p>Quick96, SAEHD: optimized architecture. You have to restart training.<br \/>\nQuick96\uff0cSAEHD\uff1a\u4f18\u5316\u7ed3\u6784\uff0c\u4f60\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u6a21\u578b\uff08\u91cd\u8981\uff0c\u5c06\u4e0d\u5728\u517c\u5bb9\u4ee5\u524d\u7684\u6a21\u578b\uff01\uff09<\/p>\n<p>Now there are only two builds: CUDA (based on 9.2) and Opencl.<br \/>\n\u73b0\u5728\u53ea\u6709\u4e24\u4e2a\u7248\u4e86\uff1a\u57fa\u4e8e9.2\u7684CUDA\u548c\u652f\u6301AMD\u7684Opencl<\/p>\n<h2>2019\u5e7412\u670826\u65e5<\/h2>\n<p>fixed mask editor<br \/>\n\u4fee\u590d\u906e\u7f69\u7f16\u8f91\u5668<\/p>\n<p>added FacesetEnhancer<br \/>\n\u6dfb\u52a0\u4eba\u8138\u589e\u5f3a\u811a\u672c<br \/>\n4.2.other) data_src util faceset enhance best GPU.bat<br \/>\n4.2.other) data_src util faceset enhance multi GPU.bat<\/p>\n<p>FacesetEnhancer greatly increases details in your source face set,<br \/>\nFacesetEnhancer\u53ef\u4ee5\u5f88\u597d\u6539\u5584\u56fe\u7247\u8d28\u91cf\u589e\u5f3a\u7ec6\u8282\u3002<\/p>\n<p>same as Gigapixel enhancer, but in fully automatic mode.<br \/>\n\u7c7b\u4f3cGigapixel\u6b63\u786e\uff0c\u4f46\u662f\u73b0\u5728\u662f\u5168\u81ea\u52a8\uff0c\u7701\u53bb\u4e86\u5f88\u591a\u6b65\u9aa4\u3002<\/p>\n<p>In OpenCL build works on CPU only.<br \/>\nOpenCL\u7248\u672c\u53ea\u80fd\u7528CPU<\/p>\n<p>before\/after https:\/\/i.imgur.com\/TAMoVs6.png<br \/>\n\u524d\u540e\u5bf9\u6bd4\u56fe\u7247<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7412\u670823\u65e5<\/h2>\n<p>Extractor: 2nd pass now faster on frames where faces are not found<br \/>\n\u63d0\u53d6\u5668\uff1a\u7b2c\u4e8c\u9636\u6bb5\u5904\u7406\u6ca1\u6709\u4eba\u8138\u7684\u56fe\u7247\u66f4\u5feb\u3002<\/p>\n<p>all models: removed options \u2018src_scale_mod\u2019, and \u2018sort samples by yaw as target\u2019<br \/>\n\u6240\u6709\u6a21\u578b\u79fb\u9664src_scale_mod \u548c sort samples by yaw \u53c2\u6570<\/p>\n<p>If you want, you can manually remove unnecessary angles from src faceset after sort by yaw.<br \/>\n\u5982\u679c\u4f60\u60f3\u8981\uff0c\u90a3\u4e48\u6839\u636eyaw\u6392\u5e8f\u540e\u624b\u52a8\u5220\u9664\u4e0d\u9700\u8981\u7684\u89d2\u5ea6<\/p>\n<p>Optimized sample generators (CPU workers). Now they consume less amount of RAM and work faster.<br \/>\n\u4f18\u5316\u6837\u672c\u751f\u6210\u5668\uff08CPU\u73a9\u5bb6\uff09 \u73b0\u5728\u4ed6\u4eec\u6d88\u8017\u66f4\u5c11\u7684RAM\uff0c\u5de5\u4f5c\u66f4\u5feb\u3002<\/p>\n<p>added<br \/>\n\u65b0\u589e<\/p>\n<p>4.2.other) data_src\/dst util faceset pack.bat<br \/>\nPacks \/aligned\/ samples into one \/aligned\/samples.pak file.<br \/>\nAfter that, all faces will be deleted.<br \/>\nsrc\u6253\u5305\u811a\u672c\uff0c\u53ef\u4ee5\u628a\u4eba\u7c7b\u538b\u7f29\u5230\u4e00\u4e2a\u6587\u4ef6\u91cc\u9762<\/p>\n<p>4.2.other) data_src\/dst util faceset unpack.bat<br \/>\nunpacks faces from \/aligned\/samples.pak to \/aligned\/ dir.<br \/>\nAfter that, samples.pak will be deleted.<br \/>\nsrc\u89e3\u538b\u811a\u672c\uff0c\u53ef\u4ee5\u628a\u538b\u7f29\u6587\u4ef6\u89e3\u538b\u6210\u56fe\u7247<\/p>\n<p>Packed faceset load and work faster.<br \/>\n\u6253\u5305\u540e\u7684\u6570\u636e\u96c6\u52a0\u8f7d\u548c\u5de5\u4f5c\u66f4\u5feb\u3002<\/p>\n<h2>2019\u5e7412\u670820\u65e5<\/h2>\n<p>fix 3rd pass of extractor for some systems<br \/>\n\u9488\u5bf9\u67d0\u4e9b\u7cfb\u7edf\uff0c\u4fee\u590d\u63d0\u53d6\u5668\u7684\u7b2c\u4e09\u9636\u6bb5\u63d0\u53d6\u3002<\/p>\n<p>More stable and precise version of the face transformation matrix<br \/>\n\u4eba\u8138\u8f6c\u6362\u77e9\u9635\u66f4\u7a33\u5b9a\uff0c\u66f4\u7cbe\u786e\u7684\u7248\u672c<\/p>\n<p>SAEHD: lr_dropout now as an option, and disabled by default<br \/>\nWhen the face is trained enough, you can enable this option to get extra sharpness for less amount of iterations<br \/>\nSAEHD\uff1alr_dropout\u73b0\u5728\u662f\u4e00\u4e2a\u9009\u9879\uff0c\u9ed8\u8ba4\u60c5\u51b5\u4e0b\u5904\u4e8e\u7981\u7528\u72b6\u6001<br \/>\n\u5f53\u9762\u90e8\u7ecf\u8fc7\u8db3\u591f\u7684\u8bad\u7ec3\u540e\uff0c\u60a8\u53ef\u4ee5\u542f\u7528\u6b64\u9009\u9879\u4ee5\u51cf\u5c11\u91cd\u590d\u6b21\u6570\uff0c\u4ece\u800c\u83b7\u5f97\u989d\u5916\u7684\u6e05\u6670\u5ea6\u3002<\/p>\n<p>added<br \/>\n4.2.other) data_src util faceset metadata save.bat<br \/>\nsaves metadata of data_src\\aligned\\ faces into data_src\\aligned\\meta.dat<br \/>\nsrc \u5143\u6570\u636e\u4fdd\u5b58<\/p>\n<p>4.2.other) data_src util faceset metadata restore.bat<br \/>\nrestore metadata from \u2018meta.dat\u2019 to images<br \/>\nif image size different from original, then it will be automatically resized<br \/>\nsrc \u5143\u6570\u636e\u6062\u590d,\u5982\u679c\u56fe\u50cf\u5c3a\u5bf8\u4e0e\u539f\u59cb\u5c3a\u5bf8\u4e0d\u540c\uff0c\u5219\u4f1a\u81ea\u52a8\u8c03\u6574\u5c3a\u5bf8<\/p>\n<p>You can greatly enhance face details of src faceset by using Topaz Gigapixel software.<br \/>\n\u60a8\u53ef\u4ee5\u4f7f\u7528Topaz Gigapixel\u8f6f\u4ef6\u6781\u5927\u5730\u589e\u5f3asrc\u9762\u90e8\u8868\u60c5\u7684\u9762\u90e8\u7ec6\u8282\u3002<\/p>\n<p>example before\/after https:\/\/i.imgur.com\/Gwee99L.jpg<br \/>\n\u6f14\u793a\u56fe\uff1a<\/p>\n<p>Download it from torrent https:\/\/rutracker.org\/forum\/viewtopic.php?t=5757118<br \/>\n\u4e0b\u8f7d\u5730\u5740<\/p>\n<p>Example of workflow:<br \/>\n\u4f7f\u7528\u6d41\u7a0b<\/p>\n<p>1) run \u2018data_src util faceset metadata save.bat\u2019<br \/>\n\u8fd0\u884c\u4fdd\u5b58\u811a\u672c<\/p>\n<p>2) launch Topaz Gigapixel<br \/>\n\u8fd0\u884c\u653e\u5927\u8f6f\u4ef6<br \/>\n3) open \u2018data_src\\aligned\\\u2019 and select all images<br \/>\n\u9009\u4e2daligned\u6240\u6709\u56fe\u7247<\/p>\n<p>4) set output folder to \u2018data_src\\aligned_topaz\u2019 (create folder in save dialog)<br \/>\n\u521b\u5efa\u4e00\u4e2a\u540d\u4e3aaligned_topaz\u8f93\u51fa\u76ee\u5f55<\/p>\n<p>5) set settings as on screenshot https:\/\/i.imgur.com\/kAVWMQG.jpg<br \/>\nyou can choose 2x, 4x, or 6x upscale rate<br \/>\n\u8bbe\u7f6e\u53c2\u6570\uff0c\u53c2\u8003\u56fe\u7247\uff0c\u53ef\u4ee5\u9009\u62e9\u653e\u5927\u500d\u6570\u3002<\/p>\n<p>6) start process images and wait full process<br \/>\n\u5f00\u59cb\u5904\u7406\u56fe\u7247<\/p>\n<p>7) rename folders:<br \/>\n\u91cd\u547d\u540d\u6587\u4ef6\u5939<br \/>\ndata_src\\aligned -&gt; data_src\\aligned_original<br \/>\ndata_src\\aligned_topaz -&gt; data_src\\aligned<\/p>\n<p>8) copy \u2018data_src\\aligned_original\\meta.dat\u2019 to \u2018data_src\\aligned\\\u2019<br \/>\n\u628ameta.dat\u62f7\u8d1d\u5230aligned\u4e0b\u9762<\/p>\n<p>9) run \u2018data_src util faceset metadata restore.bat\u2019<br \/>\n\u8fd0\u884crestore\u811a\u672c<\/p>\n<p>images will be downscaled back to original size (256\u00d7256) preserving details<br \/>\nmetadata will be restored<br \/>\n\u56fe\u7247\u4f1a\u7f29\u5c0f\u5230\u539f\u59cb\u5c3a\u5bf8\uff0c\u5143\u6570\u636e\u4f1a\u6062\u590d\u3002<\/p>\n<p>10) now your new enhanced faceset is ready to use !<br \/>\n\u73b0\u5728\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u589e\u52a0\u540e\u7684\u4eba\u8138\u6570\u636e\u4e86\u3002<\/p>\n<h2><\/h2>\n<h2><\/h2>\n<h2><\/h2>\n<h2>2019\u5e7412\u670815\u65e5<\/h2>\n<p>SAEHD,Quick96:<br \/>\nimproved model generalization, overall accuracy and sharpness<br \/>\n\u63d0\u793a\u6a21\u578b\u6cdb\u5316\u80fd\u529b\uff0c\u603b\u7edf\u51c6\u786e\u6027\u548c\u6e05\u6670\u5ea6\u3002<\/p>\n<p>by using new \u2018Learning rate dropout\u2019 technique from the paper https:\/\/arxiv.org\/abs\/1912.00144<br \/>\n\u4f7f\u7528\u4e86\u65b0\u7684\u201c\u5b66\u4e60\u7387\u4e0b\u964d\u201d\u95ee\u9898\uff0c\u53c2\u8003\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1912.00144<\/p>\n<p>An example of a loss histogram where this function is enabled after the red arrow:<br \/>\n\u4e0b\u56fe\u6f14\u793a\u4e86\u542f\u7528\u65b0\u7684\u5b66\u4e60\u7387\u4e4b\u540e\u7684\u6548\u679c\u3002<\/p>\n<p><iframe id=\"imgur-embed-iframe-pub-3olskOd\" class=\"imgur-embed-iframe-pub imgur-embed-iframe-pub-3olskOd-true-540\" src=\"https:\/\/imgur.com\/3olskOd\/embed?ref=https%3A%2F%2Fwww.deepfaker.xyz%2F%3Fp%3D62&amp;w=540\" scrolling=\"no\" allowfullscreen=\"allowfullscreen\" data-mce-fragment=\"1\"><\/iframe><\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7412\u670812\u65e5<\/h2>\n<p>removed FacesetRelighter due to low quality of the result<br \/>\n\u56e0\u4e3a\u7ed3\u679c\u4e0d\u662f\u592a\u7406\u60f3\uff0c\u56e0\u6b64\u79fb\u9664\u4e86\u4eba\u9020\u9634\u5f71\u529f\u80fdFacesetRelighter<\/p>\n<p>added sort by absdiff<br \/>\n\u6dfb\u52a0\u57fa\u4e8eabsdiff\u7684\u6392\u5e8f<\/p>\n<p>This is sort method by absolute per pixel difference between all faces.<br \/>\n\u8fd9\u662f\u901a\u8fc7\u6240\u6709\u8138\u56fe\u4e4b\u95f4\u7684\u7edd\u5bf9\u6bcf\u50cf\u7d20\u5dee\u8fdb\u884c\u6392\u5e8f\u7684\u65b9\u6cd5<\/p>\n<p>options:<br \/>\n\u9009\u9879<\/p>\n<p>Sort by similar? ( y\/n ?:help skip:y ) :<br \/>\n\u6839\u636e\u76f8\u4f3c\u5ea6\u6392\u5e8f\uff1f<\/p>\n<p>if you choose \u2018n\u2019, then most dissimilar faces will be placed first.<br \/>\n\u5982\u679c\u4f60\u9009\u62e9\u4e86\u5426\uff08N\uff09, \u90a3\u4e48\u4f1a\u5148\u653e\u7f6e\u4e0d\u76f8\u540c\u7684\u8138<\/p>\n<p>\u2018sort by final\u2019 renamed to \u2018sort by best\u2019<br \/>\n\u6700\u7ec8\u6392\u5e8f\u6539\u540d\u4e3a\u6700\u4f73\u6392\u5e8f\u3002<\/p>\n<p>OpenCL: fix extractor for some amd cards<br \/>\nOpenCL(\u9488\u5bf9A\u5361) \u4fee\u590d\u4e86\u4e00\u4e9b\u63d0\u53d6\u7684\u95ee\u9898<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7411\u670814\u65e5<\/h2>\n<p>Converter: added new color transfer mode: mix-m<br \/>\n\u8f6c\u6362\u5668\uff1a\u6dfb\u52a0\u4e86\u65b0\u7684\u989c\u8272\u8f6c\u79fb\u6a21\u5f0f\uff1amix-m<\/p>\n<h2>2019\u5e7411\u670813\u65e5<\/h2>\n<p>SAE,SAEHD,Converter:<br \/>\nSAE\uff0cSAEHD\u8f6c\u6362\u5668<\/p>\n<p>added sot-m color transfer<br \/>\n\u6dfb\u52a0\u4e86sot-m\u7684\u989c\u8272\u8f6c\u6362\u9009\u9879\u3002<\/p>\n<p>Converter:<br \/>\n\u8f6c\u6362\u5668<br \/>\nremoved seamless2 mode<br \/>\n\u79fb\u51faseamless2\u6a21\u5f0f<\/p>\n<p>FacesetRelighter:<br \/>\n\u6570\u636e\u96c6\u9634\u5f71<\/p>\n<p>Added intensity parameter to the manual picker.<br \/>\n\u5411\u624b\u52a8\u9009\u62e9\u5668\u6dfb\u52a0\u4e86\u5f3a\u5ea6\u53c2\u6570\u3002<\/p>\n<p>\u2018One random direction\u2019 and \u2018predefined 7 directions\u2019 use random intensity from 0.3 to 0.6.<br \/>\n\u201c\u4e00\u4e2a\u968f\u673a\u65b9\u5411\u201d\u548c\u201c\u9884\u5b9a\u4e497\u4e2a\u65b9\u5411\u201d\u4f7f\u75280.3\u52300.6\u4e4b\u95f4\u7684\u968f\u673a\u5f3a\u5ea6\u3002<\/p>\n<p>== 12.11.2019 ==<\/p>\n<p>FacesetRelighter fixes and improvements:<br \/>\nFacesetRelighter\u4fee\u590d\u548c\u6539\u8fdb<\/p>\n<p>now you have 3 ways:<br \/>\n\u73b0\u5728\u60a8\u67093\u79cd\u65b9\u6cd5\uff1a<\/p>\n<p>1) define light directions manually (not for google colab)<br \/>\n\u624b\u52a8\u5b9a\u4e49\u706f\u5149\u65b9\u5411\uff08\u4e0d\u9002\u7528\u4e8eGoogle Colab\uff09<\/p>\n<p>watch demo https:\/\/youtu.be\/79xz7yEO5Jw<br \/>\n\u67e5\u770b\u6f14\u793a\uff1ahttps:\/\/youtu.be\/79xz7yEO5Jw<\/p>\n<p>2) relight faceset with one random direction<br \/>\n2) \u7528\u4e00\u4e2a\u968f\u673a\u65b9\u5411\u91cd\u65b0\u7167\u660e\u9762\u90e8<\/p>\n<p>3) relight faceset with predefined 7 directions<br \/>\n3\uff09\u4f7f\u7528\u9884\u5b9a\u4e49\u76847\u4e2a\u65b9\u5411\u4e3a\u9762\u90e8\u7167\u660e<\/p>\n<h2>2019\u5e7411\u670811\u65e5<\/h2>\n<p>added FacesetRelighter:<br \/>\n\u6dfb\u52a0\u8138\u90e8\u6570\u636e\u96c6\u5149\u7ebf\u5904\u7406<\/p>\n<p>Synthesize new faces from existing ones by relighting them using DeepPortraitRelighter network.<br \/>\n\u901a\u8fc7DeepPortraitRelighter\u7f51\u7edc\u5bf9\u73b0\u6709\u9762\u5b54\u6dfb\u52a0\u4e0d\u540c\u5149\u7167\u6765\u751f\u6210\u65b0\u7684\u9762\u5b54<\/p>\n<p>With the relighted faces neural network will better reproduce face shadows.<br \/>\n\u8fd9\u4e48\u505a\u4e4b\u540e\u53ef\u4ee5\u66f4\u597d\u7684\u91cd\u73b0\u9762\u90e8\u9634\u5f71\u3002<\/p>\n<p>Therefore you can synthsize shadowed faces from fully lit faceset.<br \/>\n\u8fd9\u4e48\u4e00\u6765\u5373\u4fbf\u4f60\u7684\u6570\u636e\u96c6\u6ca1\u6709\u9634\u5f71\uff0c\u4e5f\u53ef\u4ee5\u5408\u6210\u5f88\u597d\u7684\u9634\u5f71\u6548\u679c\u3002<\/p>\n<p><iframe id=\"imgur-embed-iframe-pub-wxcmQoi\" class=\"imgur-embed-iframe-pub imgur-embed-iframe-pub-wxcmQoi-true-540\" src=\"https:\/\/imgur.com\/wxcmQoi\/embed?ref=https%3A%2F%2Fwww.deepfaker.xyz%2F%3Fp%3D62&amp;w=540\" scrolling=\"no\" allowfullscreen=\"allowfullscreen\" data-mce-fragment=\"1\"><\/iframe><\/p>\n<p>as a result, better fakes on dark faces:<br \/>\n\u4f7f\u7528\u8fd9\u4e2a\u529f\u80fd\u540e\u6548\u679c\u66f4\u597d<\/p>\n<p><iframe id=\"imgur-embed-iframe-pub-5xXIbz5\" class=\"imgur-embed-iframe-pub imgur-embed-iframe-pub-5xXIbz5-true-540\" src=\"https:\/\/imgur.com\/5xXIbz5\/embed?ref=https%3A%2F%2Fwww.deepfaker.xyz%2F%3Fp%3D62&amp;w=540\" scrolling=\"no\" allowfullscreen=\"allowfullscreen\" data-mce-fragment=\"1\"><\/iframe><\/p>\n<p>operate via<br \/>\n\u901a\u8fc7\u64cd\u4f5c<\/p>\n<p>data_x add relighted faces.bat<br \/>\ndata_x delete relighted faces.bat<\/p>\n<p>in OpenCL build Relighter runs on CPU<br \/>\n\u5728OpenCL\u4e2d\uff0c\u6784\u5efaRelighter\u5728CPU\u4e0a\u8fd0\u884c<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7411\u670809\u65e5<\/h2>\n<p>extractor: removed \u201cincreased speed of S3FD\u201d for compatibility reasons<br \/>\n\u63d0\u53d6\uff1a\u51fa\u4e8e\u517c\u5bb9\u6027\u539f\u56e0\uff0c\u5220\u9664\u4e86\u201c S3FD\u63d0\u901f\u201d<\/p>\n<p>converter:<br \/>\n\u8f6c\u6362\uff1a<\/p>\n<p>fixed crashes<br \/>\n\u4fee\u590d\u5d29\u6e83\u7684\u95ee\u9898<\/p>\n<p>removed useless \u2018ebs\u2019 color transfer<br \/>\n\u79fb\u51fa\u989c\u8272\u8f6c\u6362\u4e2d\u7684ebs<\/p>\n<p>changed keys for color degrade<br \/>\n\u4fee\u6539\u4e86color degrade\u7684\u952e<\/p>\n<p>added image degrade via denoise \u2013 same as denoise extracted data_dst.bat ,<br \/>\n\u6dfb\u52a0\u56fe\u7247\u964d\u566a\uff0c\u548c\u964d\u566a\u63d0\u53d6\u5dee\u4e0d\u591a\u3002<\/p>\n<p>but you can control this option directly in the interactive converter<br \/>\n\u4f46\u60a8\u53ef\u4ee5\u76f4\u63a5\u5728\u4ea4\u4e92\u5f0f\u8f6c\u6362\u5668\u4e2d\u63a7\u5236\u6b64\u9009\u9879<\/p>\n<p>added image degrade via bicubic downscale\/upscale<\/p>\n<p>SAEHD:<br \/>\ndefault ae_dims for df now 256. It is safe to train SAEHD on 256 ae_dims and higher resolution.<br \/>\nAE_dims\u9ed8\u8ba4\u503c\u6539\u4e3a256\uff0c\u8bad\u7ec3SAEHD\u7684\u65f6\u5019\u7528256\u7684AE\u548c\u9ad8\u50cf\u7d20\u662f\u5b89\u5168\u7684\u3002<\/p>\n<p>Example of recent fake: https:\/\/youtu.be\/_lxOGLj-MC8<br \/>\n\u6700\u8fd1\u505a\u7684\u4e00\u4e2a\u6362\u8138\u89c6\u9891\uff1ahttps:\/\/youtu.be\/_lxOGLj-MC8<\/p>\n<p>added Quick96 model.<br \/>\n\u6dfb\u52a0Quick96\u6a21\u578b<\/p>\n<p>This is the fastest model for low-end 2GB+ NVidia and 4GB+ AMD cards.<br \/>\n\u8fd9\u662f\u4e00\u4e2a\u9488\u5bf9\u4f4e\u7aef\u663e\u5361\u7684\u5feb\u901f\u6a21\u578b\u3002\u9002\u54082GB+\u7684N\u5361\u548c4GB+\u7684A\u5361<br \/>\nModel has zero options and trains a 96pix fullface.<br \/>\n\u6a21\u578b\u65e0\u9700\u914d\u7f6e\uff0c\u9ed8\u8ba4\u8bad\u7ec396\u663e\u793a\u7684\u5168\u7ec3\u6a21\u578b\u3002<\/p>\n<p>It is good for quick deepfake demo.<br \/>\n\u8be5\u6a21\u578b\u53ef\u4ee5\u5b9e\u73b0\u5feb\u901f\u6362\u8138\u3002\uff08\u5f53\u7136\u727a\u7272\u4e86\u6e05\u6670\u5ea6\u548c\u8d28\u91cf\uff09<\/p>\n<p>Example of the preview trained in 15 minutes on RTX2080Ti:<br \/>\n\u8fd9\u662f\u7528RTX2080TI\u8bad\u7ec3\u4e8615\u5206\u949f\u7684\u6548\u679c\u3002<\/p>\n<p><iframe id=\"imgur-embed-iframe-pub-oRMvZFP\" class=\"imgur-embed-iframe-pub imgur-embed-iframe-pub-oRMvZFP-true-540\" src=\"https:\/\/imgur.com\/oRMvZFP\/embed?ref=https%3A%2F%2Fwww.deepfaker.xyz%2F%3Fp%3D62&amp;w=540\" scrolling=\"no\" allowfullscreen=\"allowfullscreen\" data-mce-fragment=\"1\"><\/iframe><\/p>\n<h2>2019\u5e7410\u670827\u65e5<\/h2>\n<p>Extractor: fix for AMD cards<br \/>\n\u63d0\u53d6\uff1a\u4fee\u590dA\u5361\u7684\u95ee\u9898<\/p>\n<h2>2019\u5e7410\u670826\u65e5<\/h2>\n<p>red square of face alignment now contains the arrow that shows the up direction of an image<br \/>\n\u9762\u90e8\u5bf9\u9f50\u7684\u7ea2\u8272\u6b63\u65b9\u5f62\u73b0\u5728\u5305\u542b\u663e\u793a\u56fe\u50cf\u5411\u4e0a\u65b9\u5411\u7684\u7bad\u5934<\/p>\n<p>fix alignment of side faces<br \/>\n\u4fee\u590d\u4fa7\u9762\u5bf9\u9f50<\/p>\n<p>Before https:\/\/i.imgur.com\/pEoZ6Mu.mp4<br \/>\nafter https:\/\/i.imgur.com\/wO2Guo7.mp4<\/p>\n<p>fix message when no training data provided<br \/>\n\u4fee\u590d\u6ca1\u6709\u6570\u636e\u662f\u7684\u63d0\u793a\u4fe1\u606f<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7410\u670823\u65e5<\/h2>\n<p>enhanced sort by final: now faces are evenly distributed not only in the direction of yaw,<br \/>\nbut also in pitch<\/p>\n<p>\u589e\u5f3a\u201c\u6700\u7ec8\u6392\u5e8f\u201d\u529f\u80fd\uff0c\u73b0\u5728\u53ef\u4ee5\u7684\u8138\u90e8\u53ef\u4ee5\u5728yaw\uff08\u5de6\u53f3\uff09\u548cpitch\uff08\u4e0a\u4e0b\uff09\u4e0a\u5747\u5300\u5206\u5e03\u3002<\/p>\n<p>added \u2018sort by vggface\u2019: sorting by face similarity using VGGFace model.<br \/>\n\u6dfb\u52a0\u4e00\u4e2avggface\u6392\u5e8f\uff0c \u57fa\u4e8eVGGFace\u7684\u76f8\u4f3c\u5ea6\u6392\u5e8f<\/p>\n<p>Requires 4GB+ VRAM and internet connection for the first run.<br \/>\n\u9700\u89814G\u663e\u5b58\uff0c\u9996\u6b21\u4f7f\u7528\u9700\u8981\u8054\u7f51<\/p>\n<h2>2019\u5e7410\u670819\u65e5<\/h2>\n<p>fix extractor bug for 11GB+ cards<br \/>\n\u4fee\u590d11G\u663e\u5361\u63d0\u53d6\u65f6\u5019\u7684bug<\/p>\n<h2>2019\u5e7410\u670815\u65e5<\/h2>\n<p>removed fix \u201cfixed bug when the same face could be detected twice\u201d<br \/>\n\u79fb\u9664\u4fee\u590d\u201c\u4fee\u590d\u76f8\u540c\u7684\u8138\u88ab\u63d0\u53d6\u4e24\u6b21\u201d<\/p>\n<p>SAE\/SAEHD:<br \/>\nremoved option \u2018apply random ct\u2019<br \/>\n\u79fb\u9664\u968f\u610f\u989c\u8272\u53c2\u6570\u2018apply random ct\u2019<\/p>\n<p>added option<br \/>\n\u6dfb\u52a0\u53c2\u6570<br \/>\nColor transfer mode apply to src faceset. ( none\/rct\/lct\/mkl\/idt, ?:help skip: none )<br \/>\n\u628a\u989c\u8272\u8f6c\u6362\u5e94\u7528\u4e8esrc\u6570\u636e\u96c6\u3002<br \/>\nChange color distribution of src samples close to dst samples. Try all modes to find the best.<br \/>\n\u6539\u53d8\u989c\u8272\u8ba9src\u548cdst\u6837\u672c\u66f4\u52a0\u63a5\u8fd1\uff0c\u5c1d\u8bd5\u6240\u6709\u6a21\u578b\u627e\u5230\u6700\u597d\u7684\u4e00\u79cd\u3002<\/p>\n<p>before was lct mode, but sometime it does not work properly for some facesets.<br \/>\n\u4e4b\u524d\u7248\u672c\u9ed8\u8ba4\u7528\u7684\u662flct\u6a21\u5f0f\uff0c\u4f46\u662f\u5bf9\u4e8e\u6709\u4e9b\u4eba\u8138\u96c6\u5408\u6548\u679c\u4e0d\u597d\u3002<\/p>\n<h2>2019\u5e7410\u670814\u65e5<\/h2>\n<p>fixed bug when the same face could be detected twice<br \/>\n\u4fee\u590d\u76f8\u4f3c\u7684\u8138\u91cd\u590d\u63d0\u53d6\u7684\u95ee\u9898\u3002<\/p>\n<p>Extractor now produces a less shaked face. but second pass is now slower by 25%<br \/>\n\u73b0\u5728\uff0c\u63d0\u53d6\u5668\u63d0\u53d6\u7684\u8138\u6643\u52a8\u66f4\u5c11\u3002 \u4f46\u662f\u7b2c\u4e8c\u9636\u6bb5\u6162\u4e8625\uff05\uff08\u6211\u611f\u89c9\u4e0d\u6b62\uff09<\/p>\n<p>before\/after: https:\/\/imgur.com\/L77puLH<br \/>\n\u4e4b\u524d\/\u4e4b\u540e<\/p>\n<p>SAE, SAEHD: \u2018random flip\u2019 and \u2018learn mask\u2019 options now can be overridden.<br \/>\nSAE\uff0cSAEHD\u6a21\u578b\u7684 \u2018random flip\u2019,\u2019learn mask\u2019 \u9009\u9879\u53ef\u4ee5\u4e2d\u9014\u4fee\u6539\u4e86\u3002<\/p>\n<p>It is recommended to start training for first 20k iters always with \u2018learn_mask\u2019<br \/>\n\u5f3a\u70c8\u5efa\u8bae\u524d2\u4e07\u8fed\u4ee3\u5f00\u542flearn_mask\u53c2\u6570\u3002<\/p>\n<p>SAEHD: added option Enable random warp of samples, default is on<br \/>\nSAEHD\uff1a\u6dfb\u52a0\u53c2\u6570\u201crandom warp\u201d \uff0c\u9ed8\u8ba4\u4e3a\u542f\u7528\u3002<\/p>\n<p>Random warp is required to generalize facial expressions of both faces.<br \/>\n\u201cRandom warp\u201d \u6982\u62ec\u4e24\u5f20\u8138\u7684\u8868\u60c5\u662f\u5fc5\u9700\u7684\u3002(\u4e0d\u77e5\u9053\u548b\u7ffb\u8bd1)<\/p>\n<p>When the face is trained enough, you can disable it to get extra sharpness for less amount of iterations.<br \/>\n\u5f53\u8bad\u7ec3\u8db3\u591f\u7684\u6b21\u6570\u4e4b\u540e\uff0c\u60a8\u53ef\u4ee5\u7981\u7528\u5b83\u4ee5\u51cf\u5c11\u91cd\u590d\u6b21\u6570\uff0c\u8fd9\u6837\u8fbe\u5230\u540c\u7b49\u6e05\u6670\u5ea6\u7684\u8fed\u4ee3\u6570\u6b21\u66f4\u5c11<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7410\u670810\u65e5<\/h2>\n<p>fixed wrong NVIDIA GPU detection in extraction and training processes<br \/>\n\u4fee\u590d\u63d0\u53d6\u548c\u8bad\u7ec3N\u5361\u7684\u9519\u8bef\u8bc6\u522b<\/p>\n<p>increased speed of S3FD 1st pass extraction for GPU with &gt;= 11GB vram.<br \/>\n\u4e3a11G\u663e\u5361\u63d0\u793a\u7b2c\u4e00\u9636\u6bb5\u901f\u5ea6<\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e7410\u67089\u65e5<\/h2>\n<p>fixed wrong NVIDIA GPU indexes in a systems with two or more GPU<br \/>\n\u4fee\u590d\u591a\u5f20N\u5361\u5e8f\u53f7\u9519\u8bef\u7684\u95ee\u9898<\/p>\n<p>fixed wrong NVIDIA GPU detection on the laptops<br \/>\n\u4fee\u590d\u4e86\u7b14\u8bb0\u672c\u7535\u8111\u4e0aN\u5361\u68c0\u6d4b\u95ee\u9898\u3002<\/p>\n<p>removed TrueFace model.<br \/>\n\u79fb\u9664\u4e86TrueFace \u6a21\u578b<\/p>\n<p>added SAEHD model ( High Definition Styled AutoEncoder )<br \/>\n\u6dfb\u52a0SAEHD\u6a21\u578b\uff08\u9ad8\u6e05\u6837\u5f0f\u7684\u81ea\u52a8\u7f16\u7801\u5668\uff09<\/p>\n<p>Compare with SAE: https:\/\/i.imgur.com\/3QJAHj7.jpg<br \/>\n\u4e0eSAE\u6bd4\u8f83https:\/\/i.imgur.com\/3QJAHj7.jpg<\/p>\n<p>This is a new heavyweight model for high-end cards to achieve maximum possible deepfake quality in 2020.<br \/>\n\u8fd9\u662f\u9ad8\u7aef\u5361\u7684\u65b0\u578b\u91cd\u91cf\u7ea7\u6a21\u578b\uff0c\u53ef\u4ee5\u79f0\u97382020\u5e74\uff08\u8fd8\u80fd\u4e24\u4e2a\u6708~^_^\uff09\u3002<\/p>\n<p>Differences from SAE:<br \/>\n\u4e0eSAE\u7684\u5dee\u522b\uff1a<\/p>\n<p>+ new encoder produces more stable face and less scale jitter<br \/>\n\u65b0\u7684\u7f16\u7801\u5668\u751f\u6210\u7684\u4eba\u8138\u66f4\u65b0\u7a33\u5b9a\uff0c\u6296\u52a8\u66f4\u5c11\u3002<\/p>\n<p>+ new decoder produces subpixel clear result<br \/>\n\u65b0\u7684\u89e3\u7801\u5668\u53ef\u751f\u6210\u4e9a\u50cf\u7d20\u6e05\u6670\u5ea6\u7684\u89c6\u9891<\/p>\n<p>+ pixel loss and dssim loss are merged together to achieve both training speed and pixel trueness<br \/>\npixel loss \u548cdssim loss \u5df2\u7ecf\u88ab\u5408\u5e76\u8fdb\u53bb\uff0c\u8fd9\u6837\u53ef\u4ee5\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u548c\u50cf\u7d20\u771f\u5b9e\u6027<\/p>\n<p>+ by default networks will be initialized with CA weights, but only after first successful iteration<br \/>\n\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u7f51\u7edc\u5c06\u4f7f\u7528CA\u6743\u91cd\u8fdb\u884c\u521d\u59cb\u5316\uff0c\u4f46\u4ec5\u5728\u9996\u6b21\u6210\u529f\u8fed\u4ee3\u4e4b\u540e<\/p>\n<p>therefore you can test network size and batch size before weights initialization process<br \/>\n\u56e0\u6b64\uff0c\u60a8\u53ef\u4ee5\u5728\u6743\u91cd\u521d\u59cb\u5316\u8fc7\u7a0b\u4e4b\u524d\u6d4b\u8bd5\u7f51\u7edc\u5927\u5c0f\u548c\u6279\u6b21\u5927\u5c0f<\/p>\n<p>+ new neural network optimizer consumes less VRAM than before<br \/>\n\u65b0\u7684\u795e\u7ecf\u7f51\u7edc\u4f18\u5316\u5668\u6d88\u8017\u7684\u663e\u5b58\u6709\u6240\u964d\u4f4e<\/p>\n<p>+ added option &lt;Enable \u2018true face\u2019 training&gt;<br \/>\n\u6dfb\u52a0\u65b0\u7684\u8bad\u7ec3\u53c2\u6570 trueface\u3002<\/p>\n<p>The result face will be more like src and will get extra sharpness.<br \/>\n\u542f\u7528\u540e\u7ed3\u679c\u4f1a\u66f4\u50cfsrc\uff0c\u5e76\u4e14\u53ef\u4ee5\u63d0\u5347\u6e05\u6670\u5ea6\u3002<\/p>\n<p>Enable it for last 30k iterations before conversion.<br \/>\n\u542f\u7528\u8fd9\u4e2a\u53c2\u6570\u6700\u597d\u5728\u6700\u540e\u76843\u4e07\u8fed\u4ee3\u4e2d\u3002<\/p>\n<p>+ encoder and decoder dims are merged to one parameter encoder\/decoder dims<br \/>\n\u7f16\u7801\u548c\u754c\u9762\u7684dims\u5df2\u7ecf\u5408\u5e76\u6210\u4e00\u4e2a\u53c2\u6570<\/p>\n<p>+ added mid-full face, which covers 30% more area than half face.<br \/>\n\u6dfb\u52a0\u4e86mid-full\u8138\u7c7b\u578b\uff0c\u8fd9\u4e2a\u7c7b\u578b\u6bd4\u534a\u8138\u591a\u51fa30%\u7684\u533a\u57df\u3002<\/p>\n<p>example of the preview trained on RTX2080TI, 128 resolution, 512-21 dims, 8 batch size, 700ms per iteration:<br \/>\n\u5b9e\u4f8b\u6f14\u793a\uff0c\u57fa\u4e8eRTX2080ti, 128\u50cf\u7d20\uff0c512-21 dims , BS 8 \u6bcf\u4e2a\u8fed\u4ee3700ms<\/p>\n<p>without trueface : https:\/\/i.imgur.com\/MPPKWil.jpg<br \/>\n\u6ca1\u6709\u4f7f\u7528trueface<\/p>\n<p>with trueface +23k iters : https:\/\/i.imgur.com\/dV5Ofo9.jpg<br \/>\n\u4f7f\u7528\u4e86trueface<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e749\u670824\u65e5<\/h2>\n<p>fix TrueFace model, required retraining<br \/>\n\u4fee\u590dTrueFace \u6a21\u578b\uff0c\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e749\u670821\u65e5<\/h2>\n<p>fix avatar model<br \/>\n\u4fee\u590dAvatar\u6a21\u578b<\/p>\n<h2>2019\u5e749\u670819\u65e5<\/h2>\n<p>SAE : WARNING, RETRAIN IS REQUIRED !<br \/>\nSAE : \u8b66\u544a\uff0c\u6a21\u578b\u4e0d\u517c\u5bb9\u4ee5\u5f80\u7248\u672c\u3002<\/p>\n<p>fixed model sizes from previous update.<br \/>\n\u9488\u5bf9\u4e0a\u4e00\u4e2a\u66f4\u65b0\uff0c\u4fee\u590d\u6a21\u578b\u5927\u5c0f\u7684\u95ee\u9898\u3002<\/p>\n<p>avoided bug in ML framework(keras) that forces to train the model on random noise.<br \/>\n\u907f\u514d\u4e86\u4f7f\u7528keras\u65f6\u968f\u673a\u566a\u58f0\u5bf9\u6a21\u578b\u7684\u5f71\u54cd\u3002<\/p>\n<p>Converter: added blur on the same keys as sharpness<br \/>\n\u8f6c\u6362\uff1a\u6dfb\u52a0\u9510\u5316\u6a21\u7cca\u3002<\/p>\n<p>Added new model \u2018TrueFace\u2019. Only for NVIDIA cards.<br \/>\n\u6dfb\u52a0\u4e00\u4e2a\u53ebTrueFace\u7684\u65b0\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4ec5\u9002\u7528\u4e8eN\u5361<\/p>\n<p>This is a GAN model ported from https:\/\/github.com\/NVlabs\/FUNIT<br \/>\n\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8eGAN\u7684\u6a21\u578b\uff0c\u5177\u4f53\u5185\u5bb9\u53ef\u53c2\u8003\u82f1\u4f1f\u8fbe\u7684\u5f00\u6e90\u9879\u76eeFUNIT :https:\/\/github.com\/NVlabs\/FUNIT<\/p>\n<p>Model produces near zero morphing and high detail face.<br \/>\n\u4f18\u70b9\uff1a\u8be5\u6a21\u578b\u53ef\u4ee5\u4ea7\u751f\u9ad8\u76f8\u4f3c\u5ea6\u548c\u9ad8\u6e05\u6670\u5ea6\uff08\u7ec6\u8282\uff09\u7684\u4eba\u8138\u3002<\/p>\n<p>Model has higher failure rate than other models.<br \/>\n\u7f3a\u70b9\uff1a\u6a21\u578b\u7684\u5931\u8d25\u6982\u7387\u8fdc\u9ad8\u4e8e\u5176\u4ed6\u6a21\u578b\u3002<\/p>\n<p>It does not learn the mask, so fan-x mask modes should be used in the converter.<br \/>\n\u5b83\u4e0d\u4f1a\u5b66\u4e60\u906e\u7f69\uff0c\u6240\u4ee5\u5728\u8f6c\u6362\u7684\u65f6\u5019\u5e94\u8be5\u4f7f\u7528fan-x\u7684\u906e\u7f69\u6a21\u5f0f\u3002<\/p>\n<p>Keep src and dst faceset in same lighting conditions.<br \/>\n\u8ba9src\u548cdst\u7684\u4eba\u8138\u96c6\u5408\u4fdd\u6301\u5149\u7167\u4e00\u81f4\u3002<\/p>\n<h2>2019\u5e749\u670813\u65e5<\/h2>\n<p>Converter: added new color transfer modes: mkl, mkl-m, idt, idt-m<br \/>\n\u8f6c\u6362\u5668\uff1a\u6dfb\u52a0\u65b0\u7684\u989c\u8272\u5904\u7406\u65b9\u6848\uff1amkl,mkl-m,idt,idt-m<\/p>\n<p>SAE: removed multiscale decoder, because it\u2019s not effective<br \/>\nSAE\uff1a\u79fb\u9664multiscale\u89e3\u7801\u5668\uff0c\u56e0\u4e3a\u8fd9\u4e1c\u897f\u6ca1\u9e1f\u7528\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e749\u67087\u65e5<\/h2>\n<p>Extractor: fixed bug with grayscale images.<br \/>\n\u63d0\u53d6\u5668\uff1a\u4fee\u590d\u7070\u5ea6\u56fe\u50cf\u3001<\/p>\n<p>Converter:<br \/>\n\u8f6c\u6362\u5668\uff1a<\/p>\n<p>Session is now saved to the model folder.<br \/>\n\u4f1a\u8bdd\u73b0\u5728\u4fdd\u5b58\u5230\u6a21\u578b\u6587\u4ef6\u5939\u4e2d\u4e86\u3002<\/p>\n<p>blur and erode ranges are increased to -400+400<br \/>\n\u6a21\u7cca\u548c\u4fb5\u8680\u8303\u56f4\u589e\u52a0\u5230-400 + 400<\/p>\n<p>hist-match-bw is now replaced with seamless2 mode.<br \/>\nhist-match-bw\u73b0\u5728\u66ff\u6362\u4e3aseamless2\u6a21\u5f0f\u3002<\/p>\n<p>Added \u2018ebs\u2019 color transfer mode (works only on Windows).<br \/>\n\u6dfb\u52a0\u4e86\u2019ebs\u2019\u989c\u8272\u8f6c\u6362\u6a21\u5f0f\uff08\u4ec5\u9002\u7528\u4e8eWindows\uff09\u3002<\/p>\n<p>FANSEG model (used in FAN-x mask modes) is retrained with new model configuration<br \/>\nFANSEG\u6a21\u578b\uff08\u7528\u4e8eFAN-x\u63a9\u6a21\u6a21\u5f0f\uff09\u901a\u8fc7\u65b0\u7684\u6a21\u578b\u914d\u7f6e\u8fdb\u884c\u91cd\u65b0\u8bad\u7ec3<\/p>\n<p>and now produces better precision and less jitter<br \/>\n\u73b0\u5728\u53ef\u4ee5\u4ea7\u751f\u66f4\u597d\u7684\u7cbe\u5ea6\u548c\u66f4\u5c11\u7684\u6296\u52a8<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e748\u670830\u65e5<\/h2>\n<p>interactive converter now saves the session.<br \/>\n\u4ea4\u4e92\u5f0f\u8f6c\u6362\u5668\u73b0\u5728\u4f1a\u4fdd\u5b58\u4f1a\u8bdd\u3002<\/p>\n<p>if input frames are changed (amount or filenames)<br \/>\n\u5982\u679c\u8f93\u5165\u5e27\u88ab\u66f4\u6539<\/p>\n<p>then interactive converter automatically starts a new session.<br \/>\n\u90a3\u4e48\u4ea4\u4e92\u5f0f\u8f6c\u6362\u5668\u4f1a\u81ea\u52a8\u542f\u52a8\u65b0\u4f1a\u8bdd\u3002<\/p>\n<p>if model is more trained then all frames will be recomputed again with their saved configs.<br \/>\n\u5982\u679c\u6a21\u578b\u8bad\u7ec3\u5f97\u66f4\u597d\uff0c\u90a3\u4e48\u6240\u6709\u7684\u5e27\u90fd\u4f1a\u7528\u4ed6\u4eec\u4fdd\u5b58\u7684\u914d\u7f6e\u518d\u6b21\u91cd\u65b0\u8ba1\u7b97\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e748\u670827\u65e5<\/h2>\n<p>fixed converter navigation logic and output filenames in merge folder<br \/>\n\u4fee\u6b63\u4e86\u8f6c\u6362\u5f15\u5bfc\u903b\u8f91\uff0c\u4ee5\u53ca\u5408\u6210\u540e\u7684\u8f93\u51fa\u7684\u6587\u4ef6\u540d<\/p>\n<p>added EbSynth program. It is located in _internal\\EbSynth\\ folder<br \/>\n\u6dfb\u52a0\u4e86EbSynth\u7a0b\u5e8f\uff0c\u4f4d\u4e8e_internal\\EbSynth\\<\/p>\n<p>Start it via 10) EbSynth.bat<br \/>\n\u901a\u8fc710) EbSynth.bat\u542f\u52a8<\/p>\n<p>It starts with sample project loaded from _internal\\EbSynth\\SampleProject<br \/>\n\u5b83\u4f1a\u4ece_internal\\EbSynth\\SampleProject\u52a0\u8f7d\u793a\u4f8b\u9879\u76ee<\/p>\n<p>EbSynth is mainly used to create painted video, but with EbSynth you can fix some weird frames produced by deepfake process.<br \/>\nEbSynth\u4e3b\u8981\u7528\u4e8e\u521b\u5efa\u7ed8\u5236\u89c6\u9891\uff0c\u4f46\u4f7f\u7528EbSynth\uff0c\u60a8\u53ef\u4ee5\u4fee\u590d\u7531deepfake\u8fc7\u7a0b\u4ea7\u751f\u7684\u4e00\u4e9b\u5947\u602a\u7684\u5e27\u3002<\/p>\n<p>\u4f7f\u7528\u4e4b\u524d: https:\/\/i.imgur.com\/9xnLAL4.gifv<br \/>\n\u4f7f\u7528\u4e4b\u540e: https:\/\/i.imgur.com\/f0Lbiwf.gifv<br \/>\nEbSynth\u7684\u5b98\u65b9\u6559\u7a0b\uff1ahttps:\/\/www.youtube.com\/watch?v=0RLtHuu5jV4<\/p>\n<h2>2019\u5e748\u670826\u65e5<\/h2>\n<p>updated pdf manuals for AVATAR model.<br \/>\n\u66f4\u65b0\u4e86AVATAR\u6a21\u578b\u7684PDF\u624b\u518c\u3002<\/p>\n<p>Avatar converter: added super resolution option.<br \/>\nAVATAR\u8f6c\u6362\u5668\uff1a\u6dfb\u52a0\u8d85\u7ea7\u5206\u8fa8\u7387\u9009\u9879\u3002<\/p>\n<p>All converters:<br \/>\n\u6240\u6709\u8f6c\u6362\u5668\uff1a<\/p>\n<p>fixes and optimizations<br \/>\n\u4fee\u590d\u548c\u4f18\u5316<\/p>\n<p>super resolution DCSCN network is now replaced by RankSRGAN<br \/>\n\u8d85\u7ea7\u5206\u8fa8\u7387\u7684\u65b9\u6848\u4eceDCSCN\u7f51\u7edc\u6362\u6210\u4e86RankSRGAN<\/p>\n<p>added new option sharpen_mode and sharpen_amount<br \/>\n\u6dfb\u52a0\u4e86\u65b0\u9009\u9879sharpen_mode\u548csharpen_amount<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e748\u670824\u65e5\uff08\u6a21\u578b\u8bad\u7ec3\u76f4\u63a5\u62a5\u9519\uff09<\/h2>\n<p>Converter: FAN-dst mask mode now works for half face models.<br \/>\n\u8f6c\u6362\u5668\uff1a\u534a\u8138\u6a21\u578b\u4e5f\u53ef\u4ee5\u7528Fan-dst\u3002<\/p>\n<p>Added interactive converter.<br \/>\n\u6dfb\u52a0\u4e86\u4e92\u52a8\u8f6c\u6362\u5668<\/p>\n<p>With interactive converter you can change any parameter of any frame and see the result in real time.<br \/>\n\u4f7f\u7528\u4ea4\u4e92\u5f0f\u8f6c\u6362\u5668\uff0c\u60a8\u53ef\u4ee5\u66f4\u6539\u4efb\u4f55\u5e27\u7684\u4efb\u4f55\u53c2\u6570\u5e76\u5b9e\u65f6\u67e5\u770b\u7ed3\u679c\u3002<\/p>\n<p>Converter: added motion_blur_power param.<br \/>\n\u8f6c\u6362\uff1a\u6dfb\u52a0motion_blur_power\u53c2\u6570<\/p>\n<p>Motion blur is applied by precomputed motion vectors.<br \/>\n\u8fd0\u52a8\u6a21\u7cca\u901a\u8fc7\u5e94\u7528\u9884\u5148\u8ba1\u7b97\u7684\u8fd0\u52a8\u77e2\u91cf\u5b9e\u73b0\u3002<\/p>\n<p>So the moving face will look more realistic.<br \/>\n\u53ef\u4ee5\u4f7f\u79fb\u52a8\u7684\u8138\u770b\u8d77\u6765\u66f4\u903c\u771f\u3002<\/p>\n<p>RecycleGAN model is removed.<br \/>\nRecycleGAN\u6a21\u578b\u5df2\u7ecf\u88ab\u79fb\u51fa<\/p>\n<p>Added experimental AVATAR model. Minimum required VRAM is 6GB for NVIDIA and 12GB for AMD.<br \/>\n\u6dfb\u52a0\u963f\u51e1\u8fbe\u6a21\u578b\u9884\u4f53\u9a8c\uff0c\u914d\u7f6e\u8981\u6c426G N\u5361\uff0c\u621612G A\u5361\u3002<\/p>\n<p>Usage:<br \/>\n\u4f7f\u7528\u65b9\u6cd5\uff1a<\/p>\n<p>1) place data_src.mp4 10-20min square resolution video of news reporter sitting at the table with static background,other faces should not appear in frames.<br \/>\n\u627e\u4e00\u4e2a\u9759\u6001\u80cc\u666f\uff0c\u6bcf\u4e00\u5e27\u90fd\u53ea\u6709\u4e00\u4e2a\u4eba\uff0c\u65b9\u5f62\u7684\u65b0\u95fb\u64ad\u62a5\u7c7b\u89c6\u9891\u3002\u7247\u957f10\u523020\u5206\u949f\u3002<\/p>\n<p>2) process \u201cextract images from video data_src.bat\u201d with FULL fps<br \/>\n\u4f7f\u7528extract images from video data_src.bat\u811a\u672c\u5168\u5e27\u7387\u5206\u5272\u89c6\u9891\u3002<\/p>\n<p>3) place data_dst.mp4 video of face who will control the src face<br \/>\n\u627e\u4e00\u4e2a\u7528\u6765\u63a7\u5236\u539f\u59cb\u7684data_dst\u89c6\u9891\u3002<\/p>\n<p>4) process \u201cextract images from video data_dst FULL FPS.bat\u201d<br \/>\n\u4f7f\u7528extract images from video data_dst FULL FPS.bat \u811a\u672c\u5168\u5e27\u7387\u5206\u5272\u89c6\u9891\u3002<\/p>\n<p>5) process \u201cdata_src mark faces S3FD best GPU.bat\u201d<br \/>\n\u4f7f\u7528data_src mark faces S3FD best GPU.bat\u811a\u672c<\/p>\n<p>6) process \u201cdata_dst extract unaligned faces S3FD best GPU.bat\u201d<br \/>\n\u4f7f\u7528data_dst extract unaligned faces S3FD best GPU.bat\u811a\u672c<\/p>\n<p>7) train AVATAR.bat stage 1, tune batch size to maximum for your card (32 for 6GB), train to 50k+ iters.<br \/>\n\u8bad\u7ec3\u963f\u51e1\u8fbe\u6a21\u578b\u7b2c\u4e00\u9636\u6bb5\uff0c\u628aBS\u8c03\u6574\u6781\u9650\uff086GB\u53ef\u4ee5\u8c0332\uff09\uff0c\u8bad\u7ec3\u52305\u4e07+\u8fed\u4ee3<\/p>\n<p>8) train AVATAR.bat stage 2, tune batch size to maximum for your card (4 for 6GB), train to decent sharpness.<br \/>\n\u8bad\u7ec3\u963f\u51e1\u8fbe\u6a21\u578b\u7b2c\u4e8c\u9636\u6bb5\uff0c\u628aBS\u8c03\u6574\u6781\u9650\uff086GB\u53ef\u4ee5\u8c034\uff09\uff0c\u8bad\u7ec3\u5230\u8db3\u591f\u6e05\u6670<\/p>\n<p>9) convert AVATAR.bat<br \/>\n\u4f7f\u7528\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u8fdb\u884c\u8f6c\u6362<\/p>\n<p>10) converted to mp4.bat<br \/>\n\u5c06\u56fe\u7247\u5408\u6210\u5408\u6210MP4\u683c\u5f0f\u7684\u89c6\u9891<\/p>\n<h2><\/h2>\n<h2>2019\u5e748\u670816\u65e5\uff08\u6709\u5751\u52ff\u7528\uff0c19\u7248\u4ea6\u5982\u662f\uff09<\/h2>\n<p>fixed error \u201cFailed to get convolution algorithm\u201d on some systems<br \/>\n\u4fee\u590d\u90e8\u5206\u7cfb\u7edf\u9519\u8bef\u201cFailed to get convolution algorithm\u201d<\/p>\n<p>fixed error \u201cdll load failed\u201d on some systems<br \/>\n\u4fee\u590d\u90e8\u5206\u7cfb\u7edf\u9519\u8bef\u201cdll load failed\u201d<\/p>\n<p>model summary is now better formatted<br \/>\n\u4fee\u6539\u6a21\u578b\u4e0b\u9762\u7684summary\u6587\u4ef6\u7684\u683c\u5f0f\uff0c\u663e\u793a\u66f4\u591a\u5185\u5bb9\u3002<\/p>\n<p>Expanded eyebrows line of face masks. It does not affect mask of FAN-x converter mode..<br \/>\n\u4fee\u6539\u8138\u90e8\u906e\u7f69\uff0c\u6269\u5c55\u7709\u6bdb\u90e8\u5206\u533a\u57df\uff0c\u8fd9\u4e2a\u4fee\u6539\u4e0d\u5f71\u54cdFan-x<\/p>\n<p>ConverterMasked: added mask gradient of bottom area, same as side gradient<br \/>\n\u8f6c\u6362\u906e\u7f69\uff1a\u4e3a\u5e95\u90e8\u533a\u57df\u6dfb\u52a0\u906e\u7f69\u6e10\u53d8\u3002<\/p>\n<h2>2019\u5e747\u670823\u65e5<\/h2>\n<p>OpenCL : update versions of internal libraries<br \/>\nOpenCL : \u66f4\u65b0\u5185\u90e8\u5e93\u7684\u7248\u672c<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e746\u670820\u65e5<\/h2>\n<div class=\"table-list-cell\">\n<div class=\"commit-desc\">\n<p>&nbsp;<\/p>\n<p>Enable autobackup? (y\/n ?:help skip:%s) :<br \/>\n\u542f\u7528\u81ea\u52a8\u5907\u4efd<\/p>\n<p>Autobackup model files with preview every hour for last 15 hours. Latest backup located in model\/&lt;&gt;_autobackups\/01<br \/>\n\u81ea\u52a8\u5907\u4efd\u6a21\u578b\u6587\u4ef6\uff0c\u8fc7\u53bb15\u5c0f\u65f6\u6bcf\u5c0f\u65f6\u4e00\u6b21\u3002\u5907\u4efd\u6587\u4ef6\u4f4d\u4e8emodel\u76ee\u5f55\u4e0b\u3002<\/p>\n<p>SAE: added option only for CUDA builds:<br \/>\nSAE\uff1a\u6dfb\u52a0\u4e00\u4e2a\u9488\u5bf9CUDA\u7684\u9009\u9879<\/p>\n<p>Enable gradient clipping? (y\/n, ?:help skip:%s) :<br \/>\n\u542f\u7528\u6e10\u53d8\u526a\u88c1<\/p>\n<p>Gradient clipping reduces chance of model collapse, sacrificing speed of training.<br \/>\n\u6e10\u53d8\u88c1\u526a\u51cf\u5c11\u4e86\u6a21\u578b\u5d29\u6e83\u7684\u53ef\u80fd\u6027\uff0c\u727a\u7272\u4e86\u8bad\u7ec3\u7684\u901f\u5ea6\u3002<\/p>\n<\/div>\n<\/div>\n<p>&nbsp;<\/p>\n<h2>2019\u5e745\u670820\u65e5<\/h2>\n<p>OpenCL : fixed bug when analysing ops was repeated after each save of the model<\/p>\n<p>OpenCL\uff1a\u4fee\u590d\u6bcf\u6b21\u4fdd\u5b58\u6a21\u578b\u5c31\u91cd\u590d\u51fa\u73b0analysing ops \u7684\u95ee\u9898\u3002<\/p>\n<h2>2019\u5e745\u670810\u65e5<\/h2>\n<p>fixed work of model pretraining<\/p>\n<p>\u4fee\u590d\u9884\u8bad\u7ec3<\/p>\n<h2>2019\u5e745\u67088\u65e5<\/h2>\n<p>SAE: added new option<br \/>\nSAE:\u6dfb\u52a0\u65b0\u9009\u9879<br \/>\nApply random color transfer to src faceset? (y\/n, ?:help skip:%s) :<br \/>\n\u5bf9Src\u8138\u90e8\u5e94\u7528\u968f\u673a\u989c\u8272\u8f6c\u6362<br \/>\nIncrease variativity of src samples by apply LCT color transfer from random dst samples.<br \/>\n\u901a\u8fc7\u5e94\u7528\u968f\u673adst\u6837\u54c1\u7684LCT\u989c\u8272\u8f6c\u79fb\uff0c\u6765\u589e\u52a0src\u6837\u672c\u7684\u53d8\u5f02\u6027\u3002<\/p>\n<p>It is like \u2018face_style\u2019 learning, but more precise color transfer and without risk of model collapse,<br \/>\n\u8fd9\u6709\u70b9\u8c61\u5b66\u4e60\u2019face_style\u2019\uff0c\u4f46\u76f8\u6bd4\u4e4b\u4e0b\u66f4\u51c6\u786e\uff0c\u6ca1\u6709\u6a21\u578b\u5d29\u6e83\u7684\u98ce\u9669<\/p>\n<p>also it does not require additional GPU resources, but the training time may be longer, due to the src faceset is becoming more diverse.<br \/>\n\u540c\u65f6\u5b83\u4e5f\u4e0d\u9700\u8981\u989d\u5916\u7684GPU\u8d44\u6e90\uff0c\u4f46\u7531\u4e8esrc \u53d8\u5f97\u66f4\u52a0\u591a\u6837\u5316\uff0c\u8bad\u7ec3\u65f6\u95f4\u53ef\u80fd\u4f1a\u66f4\u957f\u3002<\/p>\n<h2><\/h2>\n<h2>2019\u5e745\u67081\u65e5<\/h2>\n<p>SAE: added option \u2018Pretrain the model?\u2019<br \/>\nSAE: \u6dfb\u52a0\u6a21\u578b\u9884\u8bad\u7ec3\u53c2\u6570<\/p>\n<p>Pretrain the model with large amount of various faces.<br \/>\n\u7528\u5927\u91cf\u7684\u4eba\u8138\u8fdb\u884c\u9884\u8bad\u7ec3\u3002<br \/>\nThis technique may help to train the fake with overlay different face shapes and light conditions of src\/dst data.<br \/>\n\u901a\u8fc7\u8986\u76d6\u4e0d\u540c\u7684\u8138\u90e8\u5f62\u72b6\u548c\u5149\u7167\u6761\u4ef6\u53ef\u80fd\u4f1a\u63d0\u5347\u6a21\u578b\u6548\u679c\u3002<\/p>\n<p>Face will be look more like a morphed. To reduce the morph effect,<br \/>\n\u7528\u8fd9\u79cd\u65b9\u5f0f\u8138\u90e8\u770b\u8d77\u6765\u4f1a\u6709\u4e9b\u53d8\u5f62\uff0c\u4e3a\u4e86\u51cf\u5c11\u8fd9\u79cd\u53d8\u5f62\u6548\u679c\uff0c<\/p>\n<p>some model files will be initialized but not be updated after pretrain: LIAE: inter_AB.h5 DF: encoder.h5.<br \/>\n\u4e00\u4e9b\u6a21\u578b\u6587\u4ef6\u5c06\u88ab\u521d\u59cb\u5316\u4f46\u5728\u9884\u8bad\u7ec3\u540e\u4e0d\u4f1a\u66f4\u65b0\uff1aLIAE\uff1ainter_AB.h5 DF\uff1aencoder.h5\u3002<\/p>\n<p>The longer you pretrain the mode\uff0clthe more morphed face will look. After that, save and run the training again.<br \/>\n\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u65f6\u95f4\u8d8a\u957f\uff0c\u53d8\u5f62\u7684\u8138\u5c31\u4f1a\u8d8a\u591a\u3002 \u7136\u540e\uff0c\u4fdd\u5b58\u5e76\u518d\u6b21\u8fd0\u884c\u6a21\u578b\u5c31\u4f1a\u8fdb\u5165\u5e38\u89c4\u6a21\u5f0f\u3002<\/p>\n<h2><\/h2>\n<h2>2019\u5e744\u670828\u65e5<\/h2>\n<p>fix 3rd pass extractor hang on AMD 8+ core processors<br \/>\n\u4fee\u590d \u63d0\u53d6\u8138\u90e8\u7b2c\u4e09\u9636\u6bb5\u7684\u95ee\u9898\uff0c\u4e3b\u8981\u9488\u5bf9AMD 8\u6838CPU<\/p>\n<p>Converter: fixed error with degrade color after applying \u2018lct\u2019 color transfer<br \/>\n\u8f6c\u6362\uff1a\u4fee\u590dlct\u4e0b\u4f7f\u7528degrade color\u51fa\u9519\u7684\u95ee\u9898\u3002<\/p>\n<p>added option at first run for all models: Choose image for the preview history? (y\/n skip:n)<br \/>\n\u4e3a\u6bcf\u4e2a\u6a21\u578b\u6dfb\u52a0\u914d\u7f6e\u9879\uff1a \u4e3a\u9884\u89c8\u7a97\u53e3\u9009\u62e9\u56fe\u7247\u3002<\/p>\n<p>Controls: [p] \u2013 next, [enter] \u2013 confirm.<br \/>\n\u63a7\u5236:[p] -\u4e0b\u4e00\u4e2a\uff0c[enter] \u2013 \u786e\u8ba4<\/p>\n<p>fixed error with option sort by yaw. Remember, do not use sort by yaw if the dst face has hair that covers the jaw.<br \/>\n\u4fee\u590dYaw\u6a21\u5f0f\u7684\u6392\u5e8f\u95ee\u9898\u3002\u8bb0\u4f4f\uff0c\u5982\u679cdst\u8138\u90e8\u6709\u8986\u76d6\u4e0b\u5df4\u7684\u5934\u53d1\uff0c\u8bf7\u4e0d\u8981\u4f7f\u7528yaw\u6a21\u5f0f\u3002<\/p>\n<h2>2019\u5e744\u670824\u65e5<\/h2>\n<p>SAE: finally the collapses were fixed<br \/>\nSAE:\u00a0<strong>\u5d29\u6e83\u7684\u95ee\u9898\u7ec8\u4e8e\u89e3\u51b3\u4e86<\/strong><\/p>\n<p>added option \u2018Use CA weights? (y\/n, ?:help skip: %s ) :<br \/>\n\u6dfb\u52a0 Use CA weights\u9009\u9879<\/p>\n<p>Initialize network with \u2018Convolution Aware\u2019 weights from paper https:\/\/arxiv.org\/abs\/1702.06295.<br \/>\n\u4f7f\u7528Convolution Aware\u521d\u59cb\u5316\u6743\u91cd\uff0c\u53c2\u8003\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/1702.06295.<\/p>\n<p>This may help to achieve a higher accuracy model, but consumes a time at first run.<br \/>\n\u8fd9\u4e5f\u8bb8\u80fd\u591f\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u6a21\u578b\uff0c\u4f46\u662f\u9996\u6b21\u8fd0\u884c\u4f1a\u6bd4\u8f83\u6162\u3002<\/p>\n<h2>2019\u5e744\u670823\u65e5<\/h2>\n<p>SAE: training should be restarted<br \/>\nSAE\uff1a\u6a21\u578b\u5fc5\u987b\u91cd\u65b0\u8bad\u7ec3<\/p>\n<p>remove option \u2018Remove gray border\u2019 because it makes the model very resource intensive.<br \/>\n\u79fb\u9664\u4e86Remove gray border\u9009\u9879\uff0c\u56e0\u4e3a\u4ed6\u4f1a\u4f7f\u5f97\u6a21\u578b\u975e\u5e38\u8017\u8d39\u8d44\u6e90\u3002<\/p>\n<h2>2019\u5e744\u670821\u65e5<\/h2>\n<p>SAE:<br \/>\nfix multiscale decoder.<br \/>\n\u4fee\u590d\u591a\u5c3a\u5ea6\u89e3\u7801\u5668<br \/>\ntraining with liae archi should be restarted<br \/>\n\u4f7f\u7528liae\u8bad\u7ec3\u7684\u6a21\u578b\u5fc5\u987b\u91cd\u65b0\u6765\u8fc7<\/p>\n<p>changed help for \u2018sort by yaw\u2019 option:<br \/>\n\u4fee\u6539sort by yaw\u7684\u5e2e\u52a9\u4fe1\u606f<br \/>\nNN will not learn src face directions that don\u2019t match dst face directions. Do not use if the dst face has hair that covers the jaw.<br \/>\n\u795e\u7ecf\u7f51\u7edc\u4e0d\u4f1a\u5b66\u4e60\u4e0edst\u9762\u90e8\u65b9\u5411\u4e0d\u5339\u914d\u7684src\u9762\u90e8\u3002 \u5982\u679cdst\u8138\u90e8\u7684\u5934\u53d1\u8986\u76d6\u4e0b\u989a\uff0c\u8bf7\u4e0d\u8981\u4f7f\u7528\u3002<\/p>\n<h2>2019\u5e744\u670820\u65e5<\/h2>\n<p>fixed work with NVIDIA cards in TCC mode<br \/>\n\u4fee\u590dN\u5361 TCC\u6a21\u5f0f<\/p>\n<p>Converter: improved FAN-x masking mode.<br \/>\n\u8f6c\u6362\u5668\uff1a\u4f18\u5316Fan-x \u906e\u7f69\u6a21\u5f0f<br \/>\nNow it excludes face obstructions such as hair, fingers, glasses, microphones, etc.<br \/>\n\u73b0\u5728\u80fd\u6392\u9664\u90a3\u4e9b\u8138\u90e8\u906e\u6321\u4e86\uff0c\u6bd4\u5982\u5934\u53d1\uff0c\u624b\u6307\uff0c\u773c\u955c\uff0c\u9ea6\u514b\u98ce\uff0cdiao \u7b49\uff01<\/p>\n<p>example https:\/\/i.imgur.com\/x4qroPp.gifv<br \/>\n\u6f14\u793a\uff1ahttps:\/\/i.imgur.com\/x4qroPp.gifv<br \/>\nIt works only for full face models, because there were glitches in half face version.<br \/>\n\u65b0\u529f\u80fd\u4ec5\u9002\u7528\u4e8e\u5168\u8138\u6a21\u578b\uff0c\u56e0\u4e3a\u534a\u8138\u7248\u672c\u6709\u6bdb\u523a\u3002<\/p>\n<p>Fanseg is trained by using manually refined by MaskEditor &gt;3000 various faces with obstructions.<br \/>\nFanseg\u901a\u8fc7\u4f7f\u7528MaskEditor&gt; 3000\u624b\u5de5\u7cbe\u5236\u7684\u5404\u79cd\u9762\u90e8\u969c\u788d\u7269\u8fdb\u884c\u8bad\u7ec3\u800c\u6210\u3002<\/p>\n<p>Accuracy of fanseg to handle complex obstructions can be improved by adding more samples to dataset, but I have no time for that\u00a0\ud83d\ude41<br \/>\n\u901a\u8fc7\u5411\u6570\u636e\u96c6\u6dfb\u52a0\u66f4\u591a\u6837\u672c\u53ef\u4ee5\u63d0\u9ad8fanseg\u5904\u7406\u590d\u6742\u969c\u788d\u7269\u7684\u51c6\u786e\u6027\uff0c\u4f46\u6211\u6ca1\u6709\u65f6\u95f4:(\u6ca1\u6709\u65f6\u95f4\uff1a<\/p>\n<p>Dataset is located in the official mega.nz folder.<br \/>\n\u8bad\u7ec3\u7684\u6570\u636e\u96c6\u5c31\u5728\u5b98\u65b9mega.nz\u6587\u4ef6\u5939\u4e2d\u3002<\/p>\n<p>If your fake has some complex obstructions that incorrectly recognized by fanseg,<br \/>\n\u5982\u679c\u4f60\u6362\u8138\u7684\u65f6\u5019\u6709\u4e00\u4e9b\u590d\u6742\u7684\u969c\u788d\u7269\u88abfanseg\u9519\u8bef\u8bc6\u522b\uff0c<\/p>\n<p>you can add manually masked samples from your fake to the dataset<br \/>\n\u4f60\u53ef\u4ee5\u624b\u52a8\u5c06\u4f60\u7684\u6837\u672c\u6dfb\u52a0\u5230\u6570\u636e\u96c6\u4e2d<\/p>\n<p>and retrain it by using \u2013model FANSEG argument in bat file. Read more info in dataset archive.<br \/>\n\u7136\u540e\u901a\u8fc7\u5728BAT\u6587\u4ef6\u4e2d\u6dfb\u52a0 \u2013moel FANSEG \u53c2\u6570\u91cd\u65b0\u8bad\u7ec3\u3002\u66f4\u591a\u5185\u5bb9\u8bf7\u770b\u6570\u636e\u96c6\u3002<\/p>\n<p>Minimum recommended VRAM is 6GB and batch size 24 to train fanseg.<br \/>\n\u8bad\u7ec3fanseg\u7684\u63a8\u8350\u914d\u7f6e\u4e3a\uff0c\u663e\u5b586GB\uff0cBS 24<br \/>\nResult model\\FANSeg_256_full_face.h5 should be placed to DeepFacelab\\facelib\\ folder<br \/>\n\u8bad\u7ec3\u5b8c\u540e\u9700\u8981\u5c06model\\FANSeg_256_full_face.h5 \u653e\u5230DeepFacelab\\facelib\\\u66ff\u6362\u540c\u540d\u6587\u4ef6\u3002<\/p>\n<p>Google Colab now works on Tesla T4 16GB.<br \/>\n\u8c37\u6b4c\u4e91\u7b14\u8bb0\u672c\u73b0\u5728\u7684\u914d\u7f6e\u662fTesla T4 16GB<br \/>\nWith Google Colaboratory you can freely train your model for 12 hours per session, then reset session and continue with last save.<br \/>\n\u4f60\u53ef\u4ee5\u4f7f\u7528\u8c37\u6b4c\u4e91\u7b14\u8bb0\u672c\u8bad\u7ec3\u6a21\u578b\uff0c\u4f46\u662f\u56de\u8bdd\u4f1a\u6700\u591a\u53ea\u80fd\u6301\u7eed12\u5c0f\u65f6\uff0c\u7136\u540e\u73af\u5883\u4f1a\u91cd\u7f6e\u3002<\/p>\n<h2>2019\u5e744\u670806\u65e5<\/h2>\n<ul>\n<li>\u6dfb\u52a0\u8bd5\u9a8c\u6027\u7684\u8499\u7248\u7f16\u8f91\u5668<\/li>\n<li>\u66f4\u65b0SAE\uff0c\u66f4\u65b0\u4e4b\u540e\u4e0d\u652f\u6301\u8001\u7684\u6a21\u578b\u3002<\/li>\n<li>\u66f4\u65b0SAE\uff0c\u5927\u5927\u964d\u4f4e\u4e86\u6a21\u578b\u5d29\u6e83\u7684\u53ef\u80fd\u6027\u3002<\/li>\n<li>\u66f4\u65b0SAE\uff0c\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u51c6\u786e\u6027\u3002<\/li>\n<li>\u66f4\u65b0SAE\uff0cResidual blocks \u6210\u4e3a\u9ed8\u8ba4\u914d\u7f6e\uff0c\u79fb\u9664\u4e86\u914d\u7f6e\u9009\u9879\u3002<\/li>\n<li>\u66f4\u65b0SAE\uff0c\u63d0\u5347\u4e86learn mask<\/li>\n<li>\u66f4\u65b0SAE\uff0c\u6dfb\u52a0\u4e86\u906e\u7f69\u9884\u89c8\uff08\u6309\u7a7a\u683c\u952e\u5207\u6362\uff09\u3002<\/li>\n<li>\u66f4\u65b0\u8f6c\u6362\u5668\uff0c\u4fee\u590dseamless \u6a21\u5f0f\u4e0b\u7684rct\/lct\u3002\u6dfb\u52a0\u4e00\u4e2amask mode\u9009\u9879: learned*FAN-prd*FAN-dst<\/li>\n<li>\u4fee\u590d\u5728\u975e\u9884\u89c8\u6a21\u5f0f\u4e0bCtrl+C \u9000\u51fa\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<p><strong>\u8499\u7248\u7f16\u8f91\u5668\u8bf4\u660e<\/strong><\/p>\n<p>\u5b83\u7684\u521b\u5efa\u662f\u4e3a\u4e86\u6539\u8fdbFANSeg\u6a21\u578b\uff0c\u4f46\u4f60\u53ef\u4ee5\u5c1d\u8bd5\u5728\u6362\u8138\u4e2d\u4f7f\u7528\u5b83\u3002<br \/>\n\u4f46\u8bf7\u8bb0\u4f4f\uff1a\u5b83\u5e76\u4e0d\u80fd\u4fdd\u8bc1\u8d28\u91cf\u7684\u63d0\u9ad8\u3002<br \/>\n\u4f7f\u7528\u65b9\u6cd5:<br \/>\n\u8fd0\u884c 5.4) data_dst mask editor.bat<\/p>\n<p>\u7f16\u8f91\u90a3\u4e9bDST\u8138\u90e8\u4e2d\u6709\u969c\u788d\u7269\u7684\u906e\u7f69\u3002<br \/>\n\u4f7f\u7528 \u2018learn mask\u2019 \u6216\u2019style values\u2019\u8bad\u7ec3SAE<br \/>\n\u622a\u56fe\u9884\u89c8: https:\/\/i.imgur.com\/SaVpxVn.jpg<br \/>\n\u4f7f\u7528<strong>\u8499\u7248\u7f16\u8f91<\/strong>\u8bad\u7ec3\u548c\u5408\u5e76\u7684\u7ed3\u679c: https:\/\/i.imgur.com\/QJi9Myd.jpg<br \/>\n\u590d\u6742\u7684\u906e\u7f69\u66f4\u96be\u8bad\u7ec3\u3002<\/p>\n<h2>2019\u5e743\u670831\u65e5<\/h2>\n<ul>\n<li>\u4fee\u590dseamless\u6a21\u5f0f\u7684\u8138\u90e8\u6296\u52a8<\/li>\n<li>\u5220\u9664\u4e24\u4e2a\u914d\u7f6e\u9009\u9879\uff0c\u5206\u522b\u662f Suppress seamless jitter \u548c seamless erode mask modifier.<\/li>\n<li>seamlessed face\u73b0\u5728\u6b63\u786e\u4f7f\u7528\u6a21\u7cca\u4fee\u9970\u7b26\uff08\u4e0d\u61c2\u3002\u3002\u3002\u673a\u7ffb\uff09<\/li>\n<li>\u6dfb\u52a0\u9009\u9879\u2019FAN-prd\uff06dst\u2019 \u2013 \u4f7f\u7528\u76f8\u4e58\u7684FAN prd\u548cdst\u906e\u7f69\uff08\u4e0d\u61c2\u2026.\u673a\u7ffb\uff09<\/li>\n<\/ul>\n<h2>2019\u5e743\u670826\u65e5<\/h2>\n<ul>\n<li>\u4fee\u6539SAE\u6a21\u578b\uff0c\u79fb\u9664\u4e86\u8f7b\u91cf\u7ea7\u7f16\u7801\u5668\uff08lightweight encoder\uff09\u3002<\/li>\n<li>\u4fee\u6539SAE\u6a21\u578b\uff0c\u7ee7\u7eed\u8bad\u7ec3\u7684\u65f6\u5019\u53ef\u4ee5\u4fee\u6539optimizer mode\u8fd9\u4e2a\u9009\u9879\u4e86\u3002<\/li>\n<li>\u4fee\u6539\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u635f\u5931\u7ebf\u73b0\u5728\u663e\u793a\u4fdd\u5b58\u540e\u7684\u5e73\u5747\u635f\u5931\u503c<\/li>\n<li>\u4fee\u590d\u8f6c\u6362\u8fc7\u7a0b\u4e2d\u6ca1\u6709\u6ca1\u6709\u8138\u62f7\u8d1d\u6587\u4ef6\u7684\u95ee\u9898\u3002<\/li>\n<li>\u5347\u7ea7\u56fe\u7247\u9884\u89c8\u5de5\u5177XNViewMP<\/li>\n<li>\u4fee\u590d\u00a0cut video.bat \u8def\u5f84\u5e26\u7a7a\u683c\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h2>2019\u5e743\u670824\u65e5<\/h2>\n<ul>\n<li>\u4fee\u6539SAE\u6a21\u578b\uff0c\u4fee\u6539\u540e\u8001\u5f97SAE\u6a21\u578b\u5c06\u4e0d\u518d\u9002\u7528\u4e8e\u65b0\u7684\u7248\u672c<\/li>\n<li>\u4fee\u590dSAE\u7684BUG\uff0c\u89e3\u51b3\u4e00\u6bb5\u65f6\u95f4\u540e\u5d29\u6e83\u7684\u95ee\u9898\u3002<\/li>\n<li>\u4fee\u6539SAE\u6a21\u578b\uff0c\u79fb\u9664\u4e86 CA weights \uff0c encoder\/decoder dims\uff0c\u8fd9\u4e09\u4e2a\u9009\u9879\u3002<\/li>\n<li>\u65b0\u589e\u6a21\u578b\u53c2\u6570\uff0cEncoder dims per channel\uff0cDecoder dims per channel\uff0cAdd residual blocks to decoder\uff0cRemove gray border<\/li>\n<li>\u65b0\u589e\u63d0\u53d6\u53c2\u6570\uff0cOutput image format?\u00a0 \u8f93\u51fa\u56fe\u7247\u683c\u5f0f\u53ef\u9009JPG \u548cPNG<\/li>\n<\/ul>\n<p>\u53c2\u6570\u7684\u4e00\u4e9b\u8bf4\u660e\uff1a<\/p>\n<p><strong>Encoder dims per channel (21-85 ?:help skip:%d)<\/strong><\/p>\n<p>\u7f16\u7801\u5668\u9009\u9879\uff0c\u6570\u503c\u8d8a\u5927\u5c31\u80fd\u8bc6\u522b\u8d8a\u591a\u7684\u9762\u90e8\u7279\u5f81\u3002\u4f46\u662f\u9700\u8981\u66f4\u591a\u7684\u663e\u5b58\uff0c\u60a8\u53ef\u4ee5\u5fae\u8c03\u6a21\u578b\u5927\u5c0f\u4ee5\u9002\u5408\u60a8\u7684GPU<\/p>\n<p><strong>Decoder dims per channel (11-85 ?:help skip:%d)<\/strong><\/p>\n<p>\u89e3\u7801\u5668\u9009\u9879\uff0c\u6570\u503c\u8d8a\u5927\uff0c\u7ec6\u8282\u8d8a\u597d\u3002\u4f46\u662f\u9700\u8981\u66f4\u591a\u7684\u663e\u5b58\uff0c\u60a8\u53ef\u4ee5\u5fae\u8c03\u6a21\u578b\u5927\u5c0f\u4ee5\u9002\u5408\u60a8\u7684GPU<\/p>\n<p><strong>Add residual blocks to decoder? (y\/n, ?:help skip:n) :<\/strong><\/p>\n<p>\u8fd9\u4e9b<strong>\u6b8b\u4f59\u5757<\/strong>\u52a9\u4e8e\u83b7\u5f97\u66f4\u597d\u7684\u7ec6\u8282\uff0c\u4f46\u9700\u8981\u66f4\u591a\u7684\u8ba1\u7b97\u65f6\u95f4\u3002<\/p>\n<p><strong>Remove gray border? (y\/n, ?:help skip:n) :<\/strong><\/p>\n<p>\u5220\u9664\u9884\u6d4b\u9762\u7684\u7070\u8272\u8fb9\u6846\uff0c\u4f46\u9700\u8981\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n<p><strong>Output image format? ( jpg png ?:help skip:png ) :<\/strong><\/p>\n<p>\u89c6\u9891\u8f6c\u56fe\u7247\uff0c\u8f93\u51fa\u53ef\u9009\u6587\u4ef6\u683c\u5f0fjpg\u6216png.<\/p>\n<div class=\"trans-right\">\n<div class=\"output-wrap small-font\">\n<div class=\"output-mod ordinary-wrap\">\n<div class=\"output-bd\" dir=\"ltr\">\n<p class=\"ordinary-output target-output clearfix\"><span class=\"\">PNG\u662f\u65e0\u635f\u7684\uff0c\u4f46\u5b83\u751f\u6210\u7684\u56fe\u50cf\u5927\u5c0f\u4e3aJPG\u768410\u500d\u3002<\/span><\/p>\n<p>JPG\u63d0\u53d6\u901f\u5ea6\u66f4\u5feb\uff0c\u5c24\u5176\u662f\u5728HDD\u786c\u76d8\u4e0a\u3002<\/p>\n<\/div>\n<p>&nbsp;<\/p>\n<\/div>\n<\/div>\n<\/div>\n<h2>2019\u5e743\u670821\u65e5<\/h2>\n<ul>\n<li>\u4fee\u590dOpenCL\u7248\u672c\uff0c\u652f\u6301\u66f4\u591a\u7684\u663e\u5361\u3002<\/li>\n<li>\u4fee\u590d\u4e86SAE\u5728\u4e00\u6bb5\u65f6\u95f4\u5185\u53ef\u80fd\u51fa\u73b0\u5d29\u6e83\u7684\u95ee\u9898\uff0c\u65b0\u7248\u4e0d\u652f\u6301\u8001\u7248\u672c\u7684model\uff0c\u4e5f\u5c31\u662f\u8001\u4e39\u767d\u7ec3\u4e86\u3002<\/li>\n<li>\u6dfb\u52a0\u9009\u9879Use CA weights\u00a0 \u4f7f\u7528\u201cConvolution Aware\u201d\u6743\u91cd\u521d\u59cb\u5316\u7f51\u7edc\u3002\u8fd9\u53ef\u80fd\u6709\u52a9\u4e8e\u5b9e\u73b0\u66f4\u9ad8\u7cbe\u5ea6\u7684\u6a21\u578b\uff0c\u4f46\u5728\u9996\u6b21\u8fd0\u884c\u65f6\u4f1a\u6d88\u8017\u65f6\u95f4\u3002<\/li>\n<li>\u79fb\u9664\u8fc7\u65f6\u7684DLIB\u63d0\u53d6\u5668\uff0c\u5efa\u8bae\u4f7f\u7528S3FD \u63d0\u53d6\u5668\uff0c\u80fd\u5927\u5927\u63d0\u9ad8\u7cbe\u5ea6\u3002<\/li>\n<li>\u66f4\u65b0\u8f6c\u6362\u5668\uff0c\u6dfb\u52a0\u9488\u5bf9\u5168\u8138\u6a21\u578b\u7684Mask\u3002<\/li>\n<\/ul>\n<p>Mask\u6a21\u5f0f\uff1a(1) learned, (2) dst, (3) FAN-prd, (4) FAN-dst (?) help. Default \u2013 1 :<\/p>\n<p>Learned :\u5982\u679c\u8bad\u7ec3\u6a21\u578b\u7684\u65f6\u5019\u9009Learn maks\u90a3\u4e48\u8f6c\u6362\u7684\u65f6\u5019\u5c31\u9009\u8fd9\u4e2a\uff0c\u8fd9\u4e2a\u9009\u9879\u6362\u8138\u540e\u8f6e\u5ed3\u76f8\u5f53\u5e73\u6ed1\uff0c\u4f46\u53ef\u80fd\u6447\u6446\u4e0d\u5b9a\u3002<br \/>\nDst ?:\u6e90\u8138\u7684\u539f\u59cbMask\uff0c\u6447\u6643\u7684\u8f6e\u5ed3\u3002<br \/>\nFAN-prd : mask from pretrained FAN model from predicted face. \u975e\u5e38\u5149\u6ed1\u800c\u4e0d\u4f1a\u6447\u6643\u7684\u8f6e\u5ed3\u3002<br \/>\nFAN-dst : mask from pretrained FAN model from dst face. \u975e\u5e38\u5149\u6ed1\u800c\u4e0d\u4f1a\u6447\u6643\u7684\u8f6e\u5ed3\u3002<\/p>\n<p>FAN\u4f18\u70b9:\u00a0 \u00a0 \u5149\u6ed1\u800c\u4e0d\u4f1a\u6447\u6643\u7684\u8f6e\u5ed3<\/p>\n<p>FAN\u7f3a\u70b9\uff1a\u5982\u679c\u8138\u90e8\u53d7\u963b\uff0c\u53ef\u80fd\u4f1a\u5728\u8f6e\u5ed3\u4e0a\u4ea7\u751f\u4f2a\u5f71<\/p>\n<p>\u4e0b\u8f7d\uff1amagnet:?xt=urn:btih:A148C2178F87F9195EB3A90FD519B651EDA42C58<\/p>\n<h2>2019\u5e743\u670813\u65e5<\/h2>\n<ul>\n<li>\u65b0\u589eSAE\u6a21\u578b\u914d\u7f6e\u9879Optimizer mode\uff0c\u00a0 \u8fd9\u662f\u4f4e\u914dN\u5361\u4eba\u7fa4\u7684\u798f\u5229\u53c2\u6570\u3002<\/li>\n<li>Epoch term \u540d\u5b57\u6539\u6210 iteration term.<\/li>\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u5728\u63a7\u5236\u53f0\u663e\u793a\u65f6\u95f4\u3002<\/li>\n<\/ul>\n<p><strong>Optimizer mode? ( 1,2,3 ?:help skip:1) :\u00a0 \uff08\u9009\u9879\u8bf4\u660e\uff09<\/strong><br \/>\n1 \u2013 default.\u00a0 \uff08\u9ed8\u8ba4\uff09<br \/>\n2 \u2013 allows you to train x2 bigger network, uses a lot of RAM. \uff08\u7f51\u7edc\u7ffb\u500d, \u4f7f\u7528\u4e00\u90e8\u5206\u5185\u5b58\uff09<br \/>\n3 \u2013 allows you to train x3 bigger network, uses huge amount of RAM and 30% slower.\uff08\u7f51\u7edcx3\uff0c\u4f7f\u7528\u8ddf\u591a\u5185\u5b58\uff0c\u964d\u901f30%\uff09<\/p>\n<p><strong>Optimizer mode\u00a0<\/strong>\u8fd9\u4e2a\u53c2\u6570\u548c\u6838\u5fc3\u7528\u610f\uff0c\u662f\u4e3a\u4e86\u89e3\u51b3\u663e\u5b58\u4e0d\u591fOOM\u7684\u60c5\u51b5\u3002\u5728\u540c\u7b49\u663e\u5361\u914d\u7f6e\u4e0b\uff0c\u53ef\u4ee5\u901a\u8fc7\u5c06\u8fd9\u4e2a\u53c2\u6570\u8bbe\u7f6e\u4e3a2\u62163\u63d0\u9ad8\u795e\u7ecf\u7f51\u7edc\u7684\u590d\u6742\u5ea6\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u4f4e\u914d\u663e\u5361\u80fd\u8fd0\u884c\u76f8\u5bf9\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u3002\u518d\u76f4\u63a5\u70b9\u8bf4\uff0c\u5c31\u662f\u4f60\u4e4b\u524d\u7684\u914d\u7f6e\u8dd1\u4e0d\u4e86\uff0c\u901a\u8fc7\u8fd9\u4e2a\u914d\u7f6e\u5c31\u80fd\u8dd1\u4e86\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4e0b\u8f7d\uff1amagnet:?xt=urn:btih:C363301FCF40D8A3F99B8CC5153603526678B08C<\/p>\n<h2>2019\u5e743\u670811\u65e5<\/h2>\n<ul>\n<li>\u5bf9\u4e8eCUDA10.1\u7528\u6237\uff0c\u9700\u8981\u66f4\u65b0\u4f60\u4eec\u7684\u60f3\u5361\u9a71\u52a8<\/li>\n<li>\u6dfb\u52a0\u65b0\u7684\u8138\u90e8\u63d0\u53d6\u65b9\u6cd5S3FD- \u66f4\u52a0\u7cbe\u786e\uff0c\u51cf\u5c11\u9519\u8bef\uff0c\u63d0\u53d6\u66f4\u5feb<\/li>\n<li>\u4f18\u5316DLib\uff0c\u63d0\u793a\u7b2c\u4e00\u9636\u6bb5\u7684\u901f\u5ea6\u3002<\/li>\n<li>\u4f18\u5316\u6240\u6709\u63d0\u53d6\u5668\uff0c\u51cf\u5c11\u201c\u5047\u8138\u201d\u3002<\/li>\n<li>\u4f18\u5316\u624b\u52a8\u63d0\u53d6\u5668\uff0c\u6dfb\u52a0\u4e86\u201ch\u201d\u6309\u94ae\u4ee5\u9690\u85cf\u5e2e\u52a9\u4fe1\u606f<\/li>\n<li>\u4f18\u5316\u5b89\u88c5\u5305\uff0c\u51cf\u5c0fAPP\u4e0e\u7cfb\u7edfPython\u4e4b\u95f4\u7684\u51b2\u7a81\u3002<\/li>\n<li>\u5220\u9664\u4e0d\u9700\u8981\u7684\u63a7\u5236\u53f0\u4fe1\u606f\uff0c\u770b\u8d77\u6765\u66f4\u52a0\u7b80\u6d01\u660e\u4e86<\/li>\n<\/ul>\n<p>\u4e0b\u8f7d\uff1amagnet:?xt=urn:btih:DBA1FFD2EECE6F9AD672CE1AD1C5BD3F6A0637D1<\/p>\n<p>&nbsp;<\/p>\n<h2>2019\u5e743\u67087\u65e5<\/h2>\n<ul>\n<li>\u53d1\u5e03\u4e86\u4e09\u4e2a\u9884\u7f16\u8bd1\u7248\u7248\u672c\uff08CUDA9.2 SSE\uff0cCUDA10.1AVX,OpenClSSE\uff09<\/li>\n<li>SSE\u7248\u672c\u7684CDUA\u4ece9.0\u5347\u7ea7\u52309.2<\/li>\n<li>AVX\u7248\u672c\u7684CUDA\u4ece10.0\u5347\u7ea7\u523010.1<\/li>\n<li>Python\u5347\u7ea7\u5230python 3.6.8<\/li>\n<li>\u7f29\u51cf\u4e86\u9884\u7f16\u8bd1\u5b89\u88c5\u5305\uff0c\u5220\u9664\u4e00\u4e9b\u65e0\u7528\u6587\u4ef6\uff0c\u5b89\u88c5\u4f53\u79ef\u53d8\u5c0f\u3002<\/li>\n<\/ul>\n<p>\u8bf4\u660e\uff1a \u4e3b\u8981\u662f\u7248\u672c\u4f9d\u8d56\u5347\u7ea7\uff0c\u5e76\u4e0d\u4fee\u6539\u6838\u5fc3\u4ee3\u7801\u3002\u8fd9\u4e09\u4e2a\u9884\u7f16\u8bd1\u7684\u96c6\u6210\u677f\u90fd\u6709\u81ea\u5df1\u9002\u7528\u7684\u8303\u56f4\u3002<\/p>\n<ul>\n<li>SSE\u00a0 \u4f4e\u914d\u7248\uff0c\u9488\u5bf9\u4f4e\u914d\u4f4eCUDA\u7248\u672c\u7684NVIDIA\u5361\uff0c\u6700\u9ad8\u5230\u00a0GTX1080 \u548c64\u4e3aCPU\u3002<\/li>\n<li>AVX \u5b8c\u5168\u7248\uff0c\u652f\u6301\u5b89\u88c5\u6700\u65b0CUDA10.1\u7684NVIDIA\u5361\uff0c\u9002\u7528\u4e8e\u652f\u6301AVX\u6307\u4ee4\u7684NVIDIA RTX\u548cCPU\u3002<\/li>\n<li>OpenCL\u517c\u5bb9\u7248\uff0c\u652f\u6301AMD\/IntelHD \u663e\u5361\u548c\u6240\u670964\u4f4dCPU<\/li>\n<\/ul>\n<p>\u4e0b\u8f7d\u5730\u5740\uff1amagnet:?xt=urn:btih:34528CAEE3077C86C155DE0DA116403785E3EE41<\/p>\n<h2>2019\u5e743\u67084\u65e5<\/h2>\n<ul>\n<li>\u9996\u6b21\u5212\u5206\u6210\u4e09\u4e2a\u7248\u672c\uff08CUDA9 SSE\uff0cCUDA10AVX,OpenClSSE\uff09<\/li>\n<li>\u4ece\u539f\u5148\u56fa\u5b9aCPU\u5230\u4f7f\u7528\u6240\u6709CPU<\/li>\n<li>\u7070\u5ea6\u68af\u5ea6\u6c42\u5747\u503c\uff0c\u4f18\u5316\u5408\u6210\u89c6\u9891\u6548\u679c\u3002<\/li>\n<li>\u66f4\u65b0\u6bcf\u4e2a\u7248\u672crequirements\u6587\u4ef6<\/li>\n<li>\u4fee\u6539multiscale_decoder\u9ed8\u8ba4\u503c\u4e3aNO<\/li>\n<li>\u52a0\u5165Python\u7248ffmpeg<\/li>\n<\/ul>\n<p>\u7248\u672c\u8bf4\u660e\uff1a<\/p>\n<ul>\n<li>SSE\u00a0 \u4f4e\u914d\u7248\uff0c\u9488\u5bf9\u4f4e\u914d\u4f4e\u7248\u672cNVIDIA\u5361<\/li>\n<li>AVX \u5b8c\u5168\u7248\uff0c\u9ad8\u914dNVIDIA\u5361<\/li>\n<li>OpenCL\u517c\u5bb9\u7248\uff0c\u652f\u6301A\u5361\uff0c\u652f\u6301CPU\u3002<\/li>\n<\/ul>\n<p>\u4e0b\u8f7d\u5730\u5740\uff1amagnet:?xt=urn:btih:4CCAA5D3A0E02C5079B28A3567EFFB9A49BAEDCD<\/p>\n<h2>2019\u5e743\u67081\u65e5<\/h2>\n<ul>\n<li>\u4fee\u590dseamless\u6a21\u5f0f\u5408\u6210\u89c6\u9891\u6296\u52a8\u7684\u95ee\u9898<\/li>\n<\/ul>\n<h2>2019\u5e742\u670828\u65e5<\/h2>\n<ul>\n<li>\u4f18\u5316\u63d0\u53d6\u5668<\/li>\n<li>\u4f18\u5316\u6392\u5e8f\u5de5\u5177\uff0c\u53ef\u4ee5\u81ea\u5b9a\u4e49\u6700\u7ec8\u8138\u90e8\u6570\u91cf<\/li>\n<li>\u4f18\u5316\u8f6c\u5316\u5668\uff0c\u4fee\u590dseamless<\/li>\n<li>\u4fee\u590d\u63d0\u53d6\uff0c\u5982\u679c\u4e0d\u5177\u5907\u7cfb\u7edf\u8981\u6c42\uff0c\u7740\u9000\u56deCPU\u5904\u7406\u3002<\/li>\n<li>\u5347\u7ea7\u5230\u00a0cuda 10.0, cudnn 7.4.1, tf 1.13.1, dlib 19.16.0<\/li>\n<\/ul>\n<h2>2019\u5e742\u670822\u65e5<\/h2>\n<ul>\n<li>\u8f6f\u4ef6\u91cd\u6784\uff0c\u6d89\u53ca\u591a\u4e2a\u6587\u4ef6<\/li>\n<li>\u4f18\u5316\u624b\u52a8\u63d0\u53d6\uff0c\u53ef\u83b7\u53d6\u9ad8\u8d28\u91cfjpg\u9762\u90e8\u6570\u636e<\/li>\n<li>\u4f18\u5316\u6392\u5e8f\u5de5\u5177\uff0c\u5bf9\u4e8eN\u5361\uff0cA\u5361\uff0c\u96c6\u6210\u5361\u9700\u8981\u7684\u53ef\u7528\u5185\u5b58\u4ece512\u964d\u5230256\u3002<\/li>\n<li>SAE\u6a21\u5f0f\u4f18\u5316\uff0c\u8138\u90e8\u50cf\u7d20\u8303\u56f4\u6269\u5c55\u81f364~256 \uff0c\u53ef\u4ee5\u8bad\u7ec3\u9ad8\u50cf\u7d20\u56fe\u7247\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>2019\u5e7412\u670829\u65e5 fix faceset enhancer for faces that contain [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":[],"_links":{"self":[{"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages\/1599"}],"collection":[{"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1599"}],"version-history":[{"count":1,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages\/1599\/revisions"}],"predecessor-version":[{"id":1600,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages\/1599\/revisions\/1600"}],"wp:attachment":[{"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1599"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}