{"id":2198,"date":"2021-04-23T15:11:36","date_gmt":"2021-04-23T07:11:36","guid":{"rendered":"https:\/\/www.deepfaker.xyz\/?page_id=2198"},"modified":"2022-11-11T10:21:33","modified_gmt":"2022-11-11T02:21:33","slug":"deepfacelab-2021-%e7%89%88%e6%9c%ac%e8%ae%b0%e5%bd%95","status":"publish","type":"page","link":"https:\/\/www.deepfaker.xyz\/?page_id=2198","title":{"rendered":"DeepFaceLab 2021 \u7248\u672c\u8bb0\u5f55"},"content":{"rendered":"<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e7411\u670820\u65e5<\/h2>\n<p>\u4fee\u590d\u4e86\u5408\u5e76\u4e2d\u7684 rct \u989c\u8272\u8f6c\u79fb<br \/>\n\u56fa\u5b9a\u6a21\u578b\u5bfc\u51fa\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e7410\u670820\u65e5<\/h2>\n<p>SAEHD\u3001AMP\uff1a\u968f\u673a\u7f29\u653e\u6bd4\u4f8b\u589e\u52a0\u5230 -0.15+0.15\u3002 \u63d0\u5347lr_dropout \u80fd\u529b\u4ee5\u8fbe\u5230\u8f83\u4f4e\u7684\u635f\u5931\u503c\u3002<\/p>\n<p>SAEHD\uff1a\u66f4\u6539\u4e86 bg_style_power \u7684\u7b97\u6cd5\u3002\u53ef\u4ee5\u66f4\u597d\u7684\u878d\u5408\u4eba\u8138<\/p>\n<p>\u6dfb\u52a0\u7684\u9009\u9879<br \/>\n\u968f\u673a\u8272\u8c03\uff08Random hue\uff09<br \/>\n\u9971\u548c\u5ea6\uff08saturation\uff09<br \/>\n\u5149\u5f3a\u5ea6\uff08light \uff09<br \/>\n\u8be5\u53c2\u6570\u4ec5\u5728\u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u5165\u5904\u5e94\u7528\u4e8e src \u4eba\u8138\u96c6\u3002<\/p>\n<p>\u5728\u9762\u90e8\u4ea4\u6362\u671f\u95f4\u7a33\u5b9a\u989c\u8272\u6270\u52a8\u3002 \u901a\u8fc7\u9009\u62e9 src faceset \u4e2d\u6700\u63a5\u8fd1\u7684\u989c\u8272\u6765\u964d\u4f4e\u989c\u8272\u8f6c\u79fb\u7684\u8d28\u91cf\u3002 \u56e0\u6b64 src faceset \u5fc5\u987b\u8db3\u591f\u591a\u6837\u5316\u3002 \u5178\u578b\u7684\u503c\u4e3a 0.05\u3002<\/p>\n<p>Liae\u7ed3\u6784\uff1a\u5f53 random_warp \u5173\u95ed\u65f6\uff0cinter_AB \u7f51\u7edc\u4e0d\u518d\u8bad\u7ec3\u4ee5\u4fdd\u6301\u9762\u90e8\u66f4\u50cf src\u3002<\/p>\n<h2>2021\u5e7410\u670809\u65e5<\/h2>\n<p>SAEHD\uff1a\u6dfb\u52a0\u4e86 -t arhi \u9009\u9879\u3002 \u4f7f\u8138\u90e8\u66f4\u50cf src\u3002<br \/>\nSAEHD\u548c AMP\uff1a<\/p>\n<p>\u53bb\u9664\u4e86\u5468\u671f\u6027\u8bad\u7ec3\u641e\u635f\u5931\u6837\u672c\u7684\u9690\u85cf\u51fd\u6570\u3002<\/p>\n<p>\u4fee\u590d\u4e86DirectX12\u7248\u672c\u5bfc\u51fa.dfm\u7684\u95ee\u9898\u3002<\/p>\n<p>\u5728\u6837\u672c\u751f\u6210\u5668\u4e2d\uff0c\u968f\u673a\u7f29\u653e\u4ece-0.05+0.05 \u589e\u52a0\u5230-0.125+0.125\uff0c\u63d0\u9ad8\u4e86\u4eba\u8138\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<h2>2021\u5e749\u670806\u65e5<\/h2>\n<p>\u4fee\u590d\u4fdd\u5b58\u9519\u8bef<br \/>\nFixed error in model saving.<\/p>\n<p>AMP\u548cSAEHD\u6dfb\u52a0 blur out mask \u53c2\u6570<br \/>\nAMP, SAEHD: added option \u2018blur out mask\u2019<\/p>\n<p>\u8bad\u7ec3\u906e\u7f69\u8fb9\u7f18\u7684\u865a\u5316<br \/>\nBlurs nearby area outside of applied face mask of training samples.<\/p>\n<p>\u53ef\u4f7f\u8138\u90e8\u9644\u8fd1\u80cc\u666f\u53d8\u5f97\u5e73\u6ed1\u5e76\u4e14\u5728\u4ea4\u6362\u7684\u8138\u90e8\u4e0a\u4e0d\u592a\u660e\u663e\u3002<br \/>\nThe result is the background near the face is smoothed and less noticeable on swapped face.<\/p>\n<p>The exact xseg mask in src and dst faceset is required.<\/p>\n<p>AMP\u548cSAEHD\uff1a\u793a\u4f8b\u5904\u7406\u5668\u6570\u91cf\u4e0d\u518d\u9650\u5236\u4e3a 8 \u4e2a\uff0c\u56e0\u6b64\u5982\u679c\u4f60\u7528\u7684\u662f16+\u6838\u7684AMD\u5904\u7406\u5668\uff0c\u8bf7\u589e\u52a0\u865a\u62df\u5185\u5b58\u3002<br \/>\nAMP, SAEHD: Sample processors count are no more limited to 8, thus if you have AMD processor with 16+ cores, increase paging file size.<\/p>\n<p>DirectX12 \u7248\u672cml\u6846\u67b6\u5347\u7ea7\u5230\u4e861.15.5\u3002<br \/>\nDirectX12 build: update tensorflow-directml to 1.15.5 version.<\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e7408\u670812\u65e5<\/h2>\n<p>\u906e\u7f69\u6a21\u578b\uff1a\u6539\u8fdb\u4e86\u9884\u8bad\u7ec3\u6a21\u578b<br \/>\nXSeg model: improved pretrain option<\/p>\n<p>\u901a\u7528\u906e\u7f69\uff1a\u6dfb\u52a0\u4e86\u66f4\u591a\u4eba\u8138\uff08\u8be5\u6570\u636e\u96c6\u6ca1\u6709\u516c\u5f00\uff09\u5e76\u4f7f\u7528\u9884\u8bad\u7ec3\u9009\u9879\u91cd\u65b0\u8bad\u7ec3\u3002\u73b0\u5728\u8d28\u91cf\u66f4\u9ad8\u4e86\u3002<br \/>\nGeneric XSeg: added more faces (the faceset is not publicly available) and retrained with pretrain option. The quality is now higher.<\/p>\n<p>\u4f7f\u7528\u65b0\u7684\u901a\u7528\u906e\u7f69\uff0c\u66f4\u65b0\u4e86RTMWF\u6570\u636e\u96c6\uff0c\u6dfb\u52a0\u4e86490\u5f20\u95ed\u773c\u56fe\u3002<br \/>\nUpdated RTM WF Dataset with the new Generic XSeg mask applied, also added 490 faces with closed eyes.<\/p>\n<h2>2021\u5e747\u670830\u65e5<\/h2>\n<p>\u5bfc\u51faAMP\/SAEHD : \u6dfb\u52a0\u4e86\u201cExport quantized\u201d \u9009\u9879\uff08\u4e4b\u524d\u6709\u7528\u8fc7\uff09<br \/>\nExport AMP\/SAEHD: added &#8220;Export quantized&#8221; option. (was enabled before)<\/p>\n<p>\u53ef\u4ee5\u4f7f\u6a21\u578b\u5bfc\u51fa\u66f4\u5feb\uff0c\u5982\u679c\u4f60\u53d1\u73b0\u6709\u6709\u95ee\u9898\uff0c\u5173\u95ed\u8fd9\u4e2a\u9009\u9879\u3002<br \/>\nMakes the exported model faster. If you have problems, disable this option.<\/p>\n<p>AMP\u6a21\u578b<br \/>\nAMP model:<\/p>\n<p>\u4fee\u6539\u4e86CT\u6a21\u5f0f\u7684\u5e2e\u52a9\u4fe1\u606f<br \/>\nchanged help of ct mode:<\/p>\n<p>\u66f4\u6539\u63a5\u8fd1 dst \u6837\u672c\u7684 src \u6837\u672c\u7684\u989c\u8272\u5206\u5e03\u3002 \u5982\u679c src faceset \u8db3\u591f\u591a\u6837\u5316\uff0c\u90a3\u4e48\u5927\u591a\u6570\u60c5\u51b5\u4e0b lct \u6a21\u5f0f\u5c31\u53ef\u4ee5\u4e86\u3002<br \/>\nChange color distribution of src samples close to dst samples. If src faceset is deverse enough, then lct mode is fine in most cases.<\/p>\n<p>\u9ed8\u8ba4\u7684\u5185\u90e8\u7ef4\u5ea6\u73b0\u5728\u662f1024<br \/>\nDefault inter dims now 1024<\/p>\n<p>\u8fd4\u56de lr_dropout \u9009\u9879<br \/>\nreturn lr_dropout option<\/p>\n<p>\u6700\u540e\u7684\u9ad8\u635f\u5931\u6837\u672c\u884c\u4e3a &#8211; \u4e0e SAEHD \u76f8\u540c<br \/>\nlast high loss samples behaviour &#8211; same as SAEHD<\/p>\n<p>Xseg\u6a21\u578b\uff1a\u6dfb\u52a0\u9884\u8bad\u7ec3\u9009\u9879\u3002<br \/>\nXSeg model: added pretrain option.<\/p>\n<p>\u901a\u7528Xseg\u6a21\u578b\uff1a\u4f7f\u7528\u9884\u8bad\u7ec3\u9009\u9879\u91cd\u65b0\u8bad\u7ec3\uff0c\u8d28\u91cf\u66f4\u597d\u4e86\u3002<br \/>\nGeneric XSeg: retrained with pretrain option. The quality is now higher.<\/p>\n<p>\u4f7f\u7528\u65b0\u7684\u901a\u7528\u906e\u7f69\u66f4\u65b0\u4e86RTM WF\u6570\u636e\u96c6\u3002<br \/>\nUpdated RTM WF Dataset with the new Generic XSeg mask applied.<\/p>\n<h2><\/h2>\n<h2>2021\u5e747\u670817\u65e5<\/h2>\n<p>SAE\/AMP: GAN\u6a21\u578b\u6062\u590d\u5230\u4e8612\u6708\u4efd\u7684\u7248\u672c\uff0c\u81f3\u4e8e\u54ea\u4e00\u4e2a\u597d\uff0c\u53ef\u4ee5\u5728\u9ad8\u6e05\u6570\u636e\u96c6\u4e0a\u505a\u4e00\u4e2a\u6d4b\u8bd5\u3002<br \/>\nSAE\/AMP: GAN model is reverted to December version, which is better, tested on high-res fakes.<\/p>\n<p>AMP \u9ed8\u8ba4\u7684\u53d8\u5f62\u56e0\u5b50\u73b0\u5728\u662f0.5<br \/>\nAMP: default morph factor is now 0.5<\/p>\n<p>\u773c\u90e8\u548c\u5634\u90e8\u4f18\u5148\u9009\u9879\u5df2\u7ecf\u79fb\u9664\uff0c\u6539\u4e3a\u9ed8\u8ba4\u542f\u52a8<br \/>\nRemoved eyes_mouth_prio option, enabled permanently.<\/p>\n<p>\u906e\u7f69\u8bad\u7ec3\u9009\u9879\u5df2\u7ecf\u79fb\u9664\uff0c\u6539\u4e3a\u9ed8\u8ba4\u542f\u52a8<br \/>\nRemoved masked training, enabled permanently.<\/p>\n<p>\u6dfb\u52a0\u811a\u672c<br \/>\nAdded script<\/p>\n<p>\u6e90\u7d20\u6750\u548c\u6e90\u7d20\u6750\u7684\u8bad\u7ec3\u811a\u672c<br \/>\n6) train AMP SRC-SRC.bat<\/p>\n<p>AMP\u9760\u8c31\u7684\u8bad\u7ec3\u65b9\u6cd5<br \/>\nStable approach to train AMP:<\/p>\n<p>\u6574\u7406\u4e00\u4e2a\u8868\u60c5\u5145\u5206\u8db3\u591f\u591a\u6837\u5316\u7684\u4eba\u8138\u96c6\u5408<br \/>\n1) Get fairly diverse src faceset<\/p>\n<p>\u8bbe\u7f6e\u53d8\u5f62\u56e0\u5b50\u4e3a0.5<br \/>\n2) Set morph factor to 0.5<\/p>\n<p>\u8bad\u7ec3\u6e90\u6e90\u6570\u636e\u96c6\u523050\u4e07\u8dcc\u5012\uff08\u8d8a\u591a\u8d8a\u597d\uff09<br \/>\n3) train AMP SRC-SRC for 500k+ iters (more is better)<\/p>\n<p>\u5220\u9664inter_dst\u7684\u6a21\u578b\u6587\u4ef6<br \/>\n4) delete inter_dst from model files<\/p>\n<p>\u6b63\u5e38\u8bad\u7ec3<br \/>\n5) train as usual<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" id=\"_x0000_i1038\" src=\"https:\/\/i.imgur.com\/nNJFpmV.jpg\" width=\"288\" height=\"337\" border=\"0\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e747\u67081\u65e5<\/h2>\n<p>AMP\u6a21\u578b\uff1a \u4fee\u590d\u9884\u89c8\u5386\u53f2<\/p>\n<p>AMP model: fixed preview history<\/p>\n<p>\u6dfb\u52a0\u2018\u5185\u90e8\u7ef4\u5ea6\u2019\u9009\u9879\uff0c\u6a21\u578b\u5df2\u7ecf\u4fee\u6539\u4e86\u3002\u9700\u8981\u7b49\u4e8e\u6216\u8005\u5927\u4e8e\u81ea\u52a8\u7f16\u7801\u5668\u7ef4\u5ea6\u3002<br \/>\nadded \u2018Inter dimensions\u2019 option. The model is not changed. Should be equal or more than AutoEncoder dimensions.<\/p>\n<p>\u7ef4\u5ea6\u8d8a\u591a\u8d8a\u597d\uff0c\u4f46\u662f\u9700\u8981\u66f4\u591a\u663e\u5b58\uff0c\u4f60\u53ef\u4ee5\u6839\u636e\u4f60\u7684\u914d\u7f6e\u6765\u5fae\u8c03\u6a21\u578b\u7684\u5927\u5c0f<br \/>\nMore dims are better, but require more VRAM. You can fine-tune model size to fit your GPU.<\/p>\n<p>\u79fb\u9664\u9884\u8bad\u7ec3\u9009\u9879<br \/>\nRemoved pretrain option.<\/p>\n<p>\u9ed8\u8ba4\u53d8\u5f62\u56e0\u5b50\u8bbe\u7f6e\u4e3a0.1<br \/>\nDefault morph factor is now 0.1<\/p>\n<p>\u5982\u4f55\u53bb\u8bad\u7ec3AMP<br \/>\nHow to train AMP:<\/p>\n<p>\u7528\u5e38\u89c4\u7684\u6e90\u7d20\u6750\u548c\u76ee\u6807\u7d20\u6750\u8bad\u7ec3<br \/>\n1) Train as usual src-dst.<\/p>\n<p>\u5220\u9664\u5185\u90e8\u6a21\u578b\u6587\u4ef6<br \/>\n2) Delete inters model files.<\/p>\n<p>\u8bad\u7ec3\u6e90\u7d20\u6750\u548c\u6e90\u7d20\u6750\uff0c\u610f\u601d\u5c31\u662f\u628a\u6e90\u7d20\u6750\u653e\u5230\u76ee\u6807\u7d20\u6750\u7684\u6587\u4ef6\u5939<br \/>\n3) Train src-src. It\u2019s mean place src aligned to data_dst<\/p>\n<p>\u5220\u9664\u5185\u90e8\u6a21\u578b\u6587\u4ef6<br \/>\n4) Delete inters model files.<\/p>\n<p>\u7528\u5e38\u89c4\u7684\u6e90\u7d20\u6750\u548c\u76ee\u6807\u7d20\u6750\u8bad\u7ec3<br \/>\n5) Train as usual src-dst.<\/p>\n<p>\u6dfb\u52a0\u811a\u672c<br \/>\nAdded scripts<\/p>\n<p>\u4f7f\u7528dfm.bat\u5bfc\u51faAMP\u6a21\u578b<br \/>\n6) export AMP as dfm.bat<\/p>\n<p>\u4f7f\u7528dfm.bat\u5bfc\u51faSAEHD\u6a21\u578b<br \/>\n6) export SAEHD as dfm.bat<\/p>\n<p>\u5bfc\u51fa\u7684dfm\u6a21\u578b\u53ef\u7528\u76f4\u64ad\u6362\u8138\u8f6f\u4ef6\uff08DeepFaceLive\uff09<br \/>\nExport model as .dfm format to work in DeepFaceLive.<\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e745\u670830\u65e5<\/h2>\n<p>\u6dfb\u52a0\u4e86\u5b9e\u9a8c\u6027\u6a21\u578bAMP\uff08\u4f5c\u4e3a\u653e\u5927\u5668\uff0c\u56e0\u4e3adst\u7684\u9762\u90e8\u8868\u60c5\u653e\u5927\u5230\u4e86src\uff09<br \/>\nAdded new experimental model \u2018AMP\u2019 (as amplifier, because dst facial expressions are amplified to src)<\/p>\n<blockquote class=\"imgur-embed-pub\" lang=\"en\" data-id=\"3eLL8oT\"><p><a href=\"https:\/\/imgur.com\/3eLL8oT\">View post on imgur.com<\/a><\/p><\/blockquote>\n<p><script async src=\"\/\/s.imgur.com\/min\/embed.js\" charset=\"utf-8\"><\/script><\/p>\n<p>\u8fd9\u4e2a\u6a21\u578b\u5177\u6709\u53d8\u5f62\u56e0\u5b50\uff0c\u53ef\u4ee5\u5728\u5408\u6210\u4e4b\u524d\u8bbe\u7f6e0.0-0.1\u4e4b\u95f4\u7684\u503c<\/p>\n<p>It has controllable \u2018morph factor\u2019, you can specify the value (0.0 .. 1.0) in the console before merging process.<\/p>\n<p>\u4e0d\u540c\u7684\u8138\u90e8\u8f6e\u5ed3\uff0c\u4f1a\u6d3b\u7684\u4e0d\u540c\u7684\u4e0b\u5df4\u7ebf<br \/>\nIf the shapes of the faces are different, you will get different jaw line<\/p>\n<blockquote class=\"imgur-embed-pub\" lang=\"en\" data-id=\"xR9MSTr\"><p><a href=\"https:\/\/imgur.com\/xR9MSTr\">View post on imgur.com<\/a><\/p><\/blockquote>\n<p><script async src=\"\/\/s.imgur.com\/min\/embed.js\" charset=\"utf-8\"><\/script><\/p>\n<p>\u8fd9\u5c31\u9700\u8981\u975e\u5e38\u786c\u6838\u7684\u540e\u671f\u5904\u7406<br \/>\nwhich requires a hard post process.<\/p>\n<p>\u4f46\u662f\u4f60\u53ef\u4ee5\u901a\u8fc7\u4e00\u4e2a\u6025\u4e8e\u5927\u4f6c\u4eba\u8138\u7684\u9884\u8bad\u7ec3\u906e\u7f69\u6a21\u578b\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff08\u5305\u542b\u5728BT\u79cd\u5b50\u91cc\u9762\uff09\u3002\u7136\u540e\u4f60\u53ef\u4ee5\u5728\u8fd9\u4e2a\u57fa\u7840\u4e0a\u7ee7\u7eed\u8bad\u7ec3dst<br \/>\nBut you can pretrain a celeb on large dst faceset with applied Generic XSeg mask (included in torrent). Then continue train with dst of the fake.<\/p>\n<p>\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u4f60\u53ef\u4ee5\u6d3b\u7684\u5f88\u597d\u7684\u878d\u5408\u6548\u679c\u3002<br \/>\nIn this case you will get more \u2018sewed\u2019 face.<\/p>\n<blockquote class=\"imgur-embed-pub\" lang=\"en\" data-id=\"jE8kXN7\"><p><a href=\"https:\/\/imgur.com\/jE8kXN7\">View post on imgur.com<\/a><\/p><\/blockquote>\n<p><script async src=\"\/\/s.imgur.com\/min\/embed.js\" charset=\"utf-8\"><\/script><\/p>\n<p>\u5408\u6210\u540e\u7684\u8138\u770b\u8d77\u6765\u4e0d\u9519<br \/>\nAnd merged face looks fine:<\/p>\n<blockquote class=\"imgur-embed-pub\" lang=\"en\" data-id=\"ASl7vco\"><p><a href=\"https:\/\/imgur.com\/ASl7vco\">View post on imgur.com<\/a><\/p><\/blockquote>\n<p><script async src=\"\/\/s.imgur.com\/min\/embed.js\" charset=\"utf-8\"><\/script><\/p>\n<p>\u57fa\u4e8e\u5927\u4f6c\u4eba\u8138\u8bad\u7ec3\u7684\u901a\u7528\u906e\u7f69\u6a21\u578b\u5df2\u7ecf\u5305\u542b\u5728\u79cd\u5b50\u6587\u4ef6\u91cc\u9762\u3002<br \/>\nLarge dst WF faceset with applied Generic XSeg mask is now included in torrent file.<\/p>\n<p>\u5982\u679c\u4f60\u7684\u4eba\u8138\u96c6\u5408\u8db3\u591f\u591a\u6837\u5316\u8db3\u591f\u5927\uff0c\u9884\u8bad\u7ec3\u7684\u65f6\u5019\u4f7f\u7528lct\u989c\u8272\u8f6c\u6362\u6a21\u5f0f\u6bd4\u8f83\u6709\u7528\u3002<br \/>\nIf your src faceset is diverse and large enough, then \u2018lct\u2019 color transfer mode should be used during pretraining.<\/p>\n<p>\u906e\u7f69\u7f16\u8f91\u5668\uff1a\u73b0\u5728\u5220\u9664\u540e\u7684\u4eba\u8138\u88ab\u79fb\u52a8\u5230\u4e86_trash \u76ee\u5f55\uff0c\u5e76\u4e14\u6309\u94ae\u79fb\u52a8\u5230\u4e86\u7a97\u53e3\u7684\u53f3\u8fb9\u3002<br \/>\nXSegEditor: delete button now moves the face to _trash directory and it has been moved to the right border of the window<\/p>\n<p>\u4eba\u8138\u6253\u5305\u5de5\u5177\u73b0\u5728\u4f1a\u63d0\u793a\u662f\u5426\u5220\u9664\u6e90\u6587\u4ef6\u3002<br \/>\nFaceset packer now asks whether to delete the original files<\/p>\n<p>\u6a21\u578b\u81ea\u52a8\u4fdd\u5b58\u65f6\u95f4\u4ece15\u5206\u949f\u6539\u6210\u4e8625\u5206\u949f\u3002<br \/>\nTrainer now saves every 25 min instead of 15<\/p>\n<p>&nbsp;<\/p>\n<h2>2021\u5e745\u670812\u65e5<\/h2>\n<p>\u6570\u636e\u96c6\u7f29\u653e\u5de5\u5177\u5df2\u7ecf\u652f\u6301\u6539\u53d8\u56fe\u7247\u7684\u8138\u7c7b\u578b<br \/>\nFacesetResizer now supports changing face type<\/p>\n<p>\u906e\u7f69\u7f16\u8f91\u5de5\u5177\uff1a\u6dfb\u52a0\u4e86\u5220\u9664\u6309\u94ae<br \/>\nXSegEditor: added delete button<\/p>\n<p>\u906e\u7f69\u8bad\u7ec3\u63d0\u5347<br \/>\nImproved training sample augmentation for XSeg trainer.<\/p>\n<p>XSeg\u6a21\u578b\u5df2\u7ecf\u88ab\u4fee\u6539\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u5904\u7406\u5927\u91cf\u4e0d\u540c\u7684\u9762\uff0c\u56e0\u6b64\u60a8\u5e94\u8be5\u91cd\u65b0\u8bad\u7ec3\u73b0\u6709\u7684XSeg\u6a21\u578b\u3002<br \/>\nXSeg model has been changed to work better with large amount of various faces, thus you should retrain existing xseg model.<\/p>\n<p>\u6dfb\u52a0\u4e86\u4e00\u4e2a\u57fa\u4e8e\u5927\u91cf\u4eba\u8138\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u5b83\u6700\u9002\u7528\u4e8esrc faceset\uff0c\u56e0\u4e3a\u5b83\u5305\u542b\u5e72\u51c0\u7684\u9762\uff0c\u4e5f\u53ef\u4ee5\u7528\u4e8e\u4e0d\u662f\u5f88\u590d\u6742\u7684dst\u955c\u5934\u3002<br \/>\nAdded Generic XSeg model pretrained on various faces. It is most suitable for src faceset because it contains clean faces, but also can be applied on dst footage without complex face obstructions.<\/p>\n<p>5.XSeg Generic) data_dst whole_face mask &#8211; apply.bat<\/p>\n<p>5.XSeg Generic) data_src whole_face mask &#8211; apply.bat<\/p>\n<h2><\/h2>\n<h2>2021\u5e744\u670822\u65e5<\/h2>\n<p>\u6dfb\u52a0\u4e86\u65b0\u7684\u96c6\u6210\u7248DeepFaceLab_DirectX12\uff0c\u80fd\u5728\u6240\u6709\u652f\u6301DirectX12\u7684Win10\u7cfb\u7edf\u4e0a\u8fd0\u884c\u3002<br \/>\nAdded new build DeepFaceLab_DirectX12, works on all devices that support DirectX12 in Windows 10:<\/p>\n<p>\u6bd4\u5982<br \/>\nAMD Radeon R5\/R7\/R9 2xx series or newer<\/p>\n<p>\u6bd4\u5982<br \/>\nIntel HD Graphics 5xx or newer<\/p>\n<p>\u6bd4\u5982<br \/>\nNVIDIA GeForce GTX 9xx series GPU or newer<\/p>\n<p>\u5728N\u5361\u4e0a\u9762\uff0cDirectX12\u7248\u672c\u8981\u6bd4NVIDIA\u7248\u672c\u616220~80%\u3002<br \/>\nDirectX12 is 20-80% slower on NVIDIA Cards comparing to \u2018NVIDIA\u2019 build.<\/p>\n<p>\u63d0\u5347\u4e86Xseg\u8bad\u7ec3\u65f6\u6837\u4f8b\u751f\u6210\u6548\u7387\u3002<br \/>\nImproved XSeg sample generator in the training process.<\/p>\n<p>== 23.03.2021 ==<\/p>\n<p>SAEHD: random_flip\u5df2\u7ecf\u88ab\u66ff\u4e3a\u65b0\u7684\u53c2\u6570<br \/>\nSAEHD: random_flip option is replaced with<\/p>\n<p>random_src_flip (default OFF)<br \/>\nRandom horizontal flip SRC faceset. Covers more angles, but the face maylook less naturally<br \/>\nrandom_dst_flip (default ON)<br \/>\nRandom horizontal flip DST faceset. Makes generalization of src-&gt;dst better, if src random flip is not enabled.<\/p>\n<p>\u6dfb\u52a0\u4e86\u4eba\u8138\u6570\u636e\u96c6\u56fe\u7247\u5927\u5c0f\u4fee\u6539\u5de5\u5177<br \/>\nAdded faceset resize tool via<\/p>\n<p>\u811a\u672c\u540d\u5982\u4e0b\uff1a<br \/>\n4.2) data_src util faceset resize.bat<br \/>\n5.2) data_dst util faceset resize.bat<\/p>\n<p>\u628a\u4eba\u8138\u6570\u636e\u96c6\u8c03\u6574\u5230\u548c\u6a21\u578b\u50cf\u7d20\u4e00\u6837\uff0c\u53ef\u4ee5\u51cf\u5c11CPU\u8d1f\u8f7d<br \/>\nResize faceset to match model resolution to reduce CPU load during training.<\/p>\n<p>\u522b\u5fd8\u4e86\u5907\u4efd\u539f\u59cb\u6570\u636e\u96c6<br \/>\nDon\u2019t forget to keep original faceset.<\/p>\n<h2>2021\u5e741\u67084\u65e5<\/h2>\n<p>SAEHD: GAN\u5347\u7ea7\uff0c\u4f7f\u5f97\u9884\u89c8\u56fe\u51cf\u5c11\u4e86\u751f\u786c\u611f\uff0c\u53d8\u5f97\u66f4\u52a0\u5e72\u51c0\uff01<br \/>\nSAEHD: GAN is improved. Now produces less artifacts and more cleaner preview.<\/p>\n<p>GAN\u7684\u9009\u9879<br \/>\nAll GAN options:<\/p>\n<p>GAN\u5f3a\u5ea6<br \/>\nGAN power<\/p>\n<p>\u5f3a\u8feb\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u8138\u90e8\u7684\u5c0f\u7ec6\u8282\u3002<br \/>\nForces the neural network to learn small details of the face.<\/p>\n<p>\u5f53\u5f00\u59cblr_dropout\uff0c\u5173\u95edrandom_warp\u8bad\u7ec3\u8db3\u591f\u4e4b\u540e\uff0c\u542f\u7528\u8fd9\u4e2a\u53c2\u6570\uff0c\u5f00\u4e86\u4e4b\u540e\u5c31\u522b\u5173\u4e86\uff01<br \/>\nEnable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don&#8217;t disable.<\/p>\n<p>\u6570\u5b57\u8d8a\u9ad8\uff0c\u4f1a\u8d8a\u751f\u786c\u3002\u6bd4\u8f83\u5e38\u7528\u7684\u503c\u4e3a0.1<br \/>\nThe higher the value, the higher the chances of artifacts. Typical fine value is 0.1<\/p>\n<p>GAN Patch\u5927\u5c0f<br \/>\nGAN patch size (3-640)<\/p>\n<p>\u6570\u5b57\u8d8a\u5230\uff0c\u8d28\u91cf\u8d8a\u597d\uff0c\u540c\u65f6\u4e5f\u9700\u8981\u8d8a\u591a\u7684\u663e\u5b58<br \/>\nThe higher patch size, the higher the quality, the more VRAM is required.<\/p>\n<p>\u5373\u4f7f\u5728\u6700\u4f4e\u8bbe\u7f6e\u4e0b\uff0c\u60a8\u4e5f\u53ef\u4ee5\u83b7\u5f97\u66f4\u9510\u5229\u7684\u8fb9\u7f18\u3002<br \/>\nYou can get sharper edges even at the lowest setting.<\/p>\n<p>\u5178\u578b\u503c\u4e3a 8<br \/>\nTypical fine value is resolution \/ 8.<\/p>\n<p>GAN \u7f51\u7edc\u7ef4\u5ea6<br \/>\nGAN dimensions (4-64)<\/p>\n<p>GAN\u7f51\u7edc\u7684\u5c3a\u5bf8<br \/>\nThe dimensions of the GAN network.<\/p>\n<p>\u5c3a\u5bf8\u8d8a\u9ad8\uff0c\u5bf9VRAM\u7684\u8981\u6c42\u5c31\u8d8a\u9ad8\u3002<br \/>\nThe higher dimensions, the more VRAM is required.<\/p>\n<p>\u5373\u4f7f\u5728\u6700\u4f4e\u8bbe\u7f6e\u4e0b\uff0c\u60a8\u4e5f\u53ef\u4ee5\u83b7\u5f97\u66f4\u9510\u5229\u7684\u8fb9\u7f18\u3002<br \/>\nYou can get sharper edges even at the lowest setting.<\/p>\n<p>\u5178\u578b\u503c\u4e3a 16<br \/>\nTypical fine value is 16.<\/p>\n<p>\u4e0d\u540c\u8bbe\u7f6e\u7684\u6bd4\u8f83\u56fe<br \/>\nComparison of different settings: https:\/\/i.imgur.com\/6IgvsLN.png<\/p>\n","protected":false},"excerpt":{"rendered":"<p>&nbsp; &nbsp; 2021\u5e7411\u670820\u65e5 \u4fee\u590d\u4e86\u5408\u5e76\u4e2d\u7684 rct \u989c\u8272\u8f6c\u79fb \u56fa\u5b9a\u6a21\u578b\u5bfc\u51fa\u3002 &#038;nbs [&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\/2198"}],"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=2198"}],"version-history":[{"count":17,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages\/2198\/revisions"}],"predecessor-version":[{"id":2391,"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=\/wp\/v2\/pages\/2198\/revisions\/2391"}],"wp:attachment":[{"href":"https:\/\/www.deepfaker.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2198"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}