{"id":1604,"date":"2020-02-28T18:29:18","date_gmt":"2020-02-28T10:29:18","guid":{"rendered":"https:\/\/www.deepfaker.xyz\/?page_id=1604"},"modified":"2021-01-04T11:01:05","modified_gmt":"2021-01-04T03:01:05","slug":"deepfacelab-2020-dfl2-%e7%89%88%e6%9c%ac%e8%ae%b0%e5%bd%95","status":"publish","type":"page","link":"https:\/\/www.deepfaker.xyz\/?page_id=1604","title":{"rendered":"DeepFaceLab 2020\/DFL2 \u7248\u672c\u8bb0\u5f55"},"content":{"rendered":"
\n

2021\u5e741\u67084\u65e5<\/h2>\n

SAEHD: GAN\u5347\u7ea7\uff0c\u4f7f\u5f97\u9884\u89c8\u56fe\u51cf\u5c11\u4e86\u751f\u786c\u611f\uff0c\u53d8\u5f97\u66f4\u52a0\u5e72\u51c0\uff01
\nSAEHD: GAN is improved. Now produces less artifacts and more cleaner preview.<\/p>\n

GAN\u7684\u9009\u9879
\nAll GAN options:<\/p>\n

GAN\u5f3a\u5ea6
\nGAN power<\/p>\n

\u5f3a\u8feb\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u8138\u90e8\u7684\u5c0f\u7ec6\u8282\u3002
\nForces the neural network to learn small details of the face.<\/p>\n

\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
\nEnable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don’t disable.<\/p>\n

\u6570\u5b57\u8d8a\u9ad8\uff0c\u4f1a\u8d8a\u751f\u786c\u3002\u6bd4\u8f83\u5e38\u7528\u7684\u503c\u4e3a0.1
\nThe higher the value, the higher the chances of artifacts. Typical fine value is 0.1<\/p>\n

GAN Patch\u5927\u5c0f
\nGAN patch size (3-640)<\/p>\n

\u6570\u5b57\u8d8a\u5230\uff0c\u8d28\u91cf\u8d8a\u597d\uff0c\u540c\u65f6\u4e5f\u9700\u8981\u8d8a\u591a\u7684\u663e\u5b58
\nThe higher patch size, the higher the quality, the more VRAM is required.<\/p>\n

\u5373\u4f7f\u5728\u6700\u4f4e\u8bbe\u7f6e\u4e0b\uff0c\u60a8\u4e5f\u53ef\u4ee5\u83b7\u5f97\u66f4\u9510\u5229\u7684\u8fb9\u7f18\u3002
\nYou can get sharper edges even at the lowest setting.<\/p>\n

\u5178\u578b\u503c\u4e3a 8
\nTypical fine value is resolution \/ 8.<\/p>\n

GAN \u7f51\u7edc\u7ef4\u5ea6
\nGAN dimensions (4-64)<\/p>\n

GAN\u7f51\u7edc\u7684\u5c3a\u5bf8
\nThe dimensions of the GAN network.<\/p>\n

\u5c3a\u5bf8\u8d8a\u9ad8\uff0c\u5bf9VRAM\u7684\u8981\u6c42\u5c31\u8d8a\u9ad8\u3002
\nThe higher dimensions, the more VRAM is required.<\/p>\n

\u5373\u4f7f\u5728\u6700\u4f4e\u8bbe\u7f6e\u4e0b\uff0c\u60a8\u4e5f\u53ef\u4ee5\u83b7\u5f97\u66f4\u9510\u5229\u7684\u8fb9\u7f18\u3002
\nYou can get sharper edges even at the lowest setting.<\/p>\n

\u5178\u578b\u503c\u4e3a 16
\nTypical fine value is 16.<\/p>\n

\u4e0d\u540c\u8bbe\u7f6e\u7684\u6bd4\u8f83\u56fe
\nComparison of different settings: https:\/\/i.imgur.com\/6IgvsLN.png<\/p>\n

2020\u5e7412\u670822\u65e5<\/h2>\n

\u7f29\u77ed\u8bad\u7ec3\u6570\u636e\u7684\u52a0\u8f7d\u65f6\u95f4
\nThe load time of training data has been reduced significantly.<\/p>\n

2020\u5e7412\u670820\u65e5<\/h2>\n

SAEHD:<\/p>\n

\u73b0\u5728lr_dropout\u548cAdaBelief\u53ef\u4ee5\u540c\u65f6\u4f7f\u7528\u4e86\u3002
\nlr_dropout now can be used with AdaBelief<\/p>\n

\u773c\u90e8\u4f18\u5148\u88ab\u66ff\u6362\u4e3a\u773c\u90e8\u548c\u5634\u90e8\u4f18\u5148
\nEyes priority is replaced with Eyes and mouth priority<\/p>\n

\u4e3b\u8981\u662f\u4e3a\u4e86\u4fee\u590d\u5f02\u5f62\u773c\u548c\u773c\u795e\u4e71\u98d8\u7684\u95ee\u9898\uff0c\u540c\u65f6\u4e5f\u80fd\u8ba9\u7259\u9f7f\u7684\u7ec6\u8282\u66f4\u9ad8\u3002
\nHelps to fix eye problems during training like “alien eyes” and wrong eyes direction.<\/p>\n

\u540c\u65f6\u4e5f\u80fd\u8ba9\u7259\u9f7f\u7684\u7ec6\u8282\u66f4\u9ad8
\nAlso makes the detail of the teeth higher.<\/p>\n

\u65b0\u6a21\u578b\u7684\u9ed8\u8ba4\u503c\u6539\u54af
\nNew default values with new model:<\/p>\n

\u6ca1\u5565\u597d\u7ffb\u8bd1
\nArchi : \u2018liae-ud\u2019<\/p>\n

\u6ca1\u5565\u597d\u7ffb\u8bd1
\nAdaBelief : enabled<\/p>\n<\/div>\n

2020\u5e7412\u670816\u65e5 \uff08\u63d0\u901f\uff09<\/h2>\n

 <\/p>\n

Windows 10 \u6bd4\u8f83\u65b0\u7684\u7248\u672c\u8bb0\u5f97\u4fee\u6539\u56fe\u5f62\u8bbe\u7f6e\uff0c\u542f\u7528GPU\u52a0\u901f\u3002<\/p>\n

\"\"<\/p>\n

\u9488\u5bf9\u6240\u6709\u82f1\u4f1f\u8fbe\u663e\u5361\u7684\u96c6\u6210\u7248
\nNow single build for all video cards.<\/p>\n

\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6TensorFlow\u5347\u7ea7\u5230\u4e862.4.0,\u540c\u6b65\u5347\u7ea7 CUDA 11.2,CuDNN 8.0.5
\nUpgraded to Tensorflow 2.4.0, CUDA 11.2, CuDNN 8.0.5.<\/p>\n

\u4f60\u4e0d\u9700\u8981\u5b89\u88c5\u4efb\u4f55\u4f9d\u8d56\uff08\u9664\u4e86\u5347\u7ea7\u663e\u5361\u9a71\u52a8\uff09
\nYou don\u2019t need to install anything.<\/p>\n

\n

2020\u5e7412\u670811\u65e5<\/h2>\n

\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6tf\u66f4\u65b0\u52302.4.0rc4
\nUpgrade to Tensorflow 2.4.0rc4<\/p>\n

\u652f\u6301\u82f1\u4f1f\u8fbe3000\u7cfb\u5217
\nNow support RTX 3000 series.<\/p>\n

\u7b97\u529b\u53ea\u67093.0\u663e\u5361\u5c06\u4e0d\u518d\u652f\u6301\u3002
\nVideocards with Compute Capability 3.0 are no longer supported.<\/p>\n

\u4e0d\u652f\u6301AVX\u7684CPU\u4e5f\u4e0d\u518d\u652f\u6301
\nCPUs without AVX are no longer supported.<\/p>\n

==========3000\u7cfb\u5217\u5206\u5272\u7ebf=============<\/strong><\/p>\n

\u00a02020\u5e748\u67082\u65e5<\/h2>\n

SAEHD: now random_warp is disabled for pretraining mode by default
\nSAEHD: \u9488\u5bf9\u9884\u8bad\u7ec3\uff0c\u6a21\u578b\u9009\u9879random_warp\u9ed8\u8ba4\u4e3a\u4e0d\u542f\u7528\u3002<\/p>\n

Merger: fix load time of xseg if it has no model files
\n\u5408\u6210\uff1a\u4fee\u590d\u6ca1\u6709\u6a21\u578b\u6587\u4ef6\u662f\u52a0\u8f7dXseg\u65f6\u95f4\u7684\u95ee\u9898\u3002<\/p>\n

2020\u5e747\u670818\u65e5<\/h2>\n

Fixes
\n\u4fee\u6b63<\/p>\n

SAEHD: write_preview_history now works faster
\n\u73b0\u5728\u4fdd\u5b58\u5386\u53f2\u56fe\u7247\u53d8\u5feb\u4e86<\/p>\n

The frequency at which the preview is saved now depends on the resolution.
\n\u4fdd\u5b58\u7684\u9891\u7387\u53d6\u51b3\u4e8e\u6a21\u578b\u7684\u50cf\u7d20<\/p>\n

For example 64×64 \u2013 every 10 iters. 448×448 \u2013 every 70 iters.
\n\u6bd4\u598264×64\u7684\u50cf\u7d20\uff0c\u6ca110\u4e2a\u8fed\u4ee3\u4fdd\u5b58\u4e00\u6b21\uff0c448×448 \u6ca170\u4e2a\u8fed\u4ee3\u4fdd\u5b58\u4e00\u6b21\u3002<\/p>\n

Merger: added option \u201cNumber of workers?\u201d
\n\u5408\u6210\uff1a\u6dfb\u52a0\u4e86\u7ebf\u7a0b\u6570\u91cf\u7684\u9009\u9879\u3002<\/p>\n

Specify the number of threads to process.
\n\u53ef\u4ee5\u6307\u5b9a\u7ebf\u7a0b\u7684\u6570\u91cf<\/p>\n

A low value may affect performance.
\n\u503c\u592a\u4f4e\u53ef\u80fd\u4f1a\u5f71\u54cd\u6027\u80fd<\/p>\n

A high value may result in memory error.
\n\u503c\u592a\u9ad8\u53ef\u80fd\u4f1a\u5bfc\u81f4\u5185\u5b58\u9519\u8bef\u3002<\/p>\n

The value may not be greater than CPU cores.
\n\u8fd9\u4e2a\u503c\u4e0d\u5e94\u8be5\u5927\u4e8ecpu\u5185\u6838\u6570\u91cf\u3002<\/p>\n

 <\/p>\n

2020\u5e747\u670817\u65e5<\/h2>\n

SAEHD:<\/p>\n

Pretrain dataset is replaced with high quality FFHQ dataset.
\n\u9884\u8bad\u7ec3\u96c6\u6362\u6210\u4e86\u9ad8\u8d28\u91cf\u7684FFHQ\u6570\u636e\u96c6\u3002<\/p>\n

Changed help for \u201cLearning rate dropout\u201d option:
\n\u6539\u53d8\u4e86\u201cLearning rate dropout\u201d \u7684\u5e2e\u52a9\u4fe1\u606f<\/p>\n

When the face is trained enough, you can enable this option to get extra sharpness and reduce subpixel shake for less amount of iterations.
\n\u5f53\u8bad\u7ec3\u8db3\u591f\u65f6\uff0c\u4f60\u53ef\u4ee5\u542f\u7528\u8fd9\u4e2a\u9009\u9879\u83b7\u5f97\u989d\u5916\u7684\u6e05\u6670\u5ea6\u548c\u51cf\u5c11\u5b50\u50cf\u7d20\u7684\u6296\u52a8\uff0c\u4ece\u800c\u51cf\u5c11\u8fed\u4ee3\u6b21\u6570\u3002<\/p>\n

Enabled it before \u201cdisable random warp\u201d and before GAN. n disabled. y enabled
\n\u5728\u7981\u7528random warp \u548cGAN\u4e4b\u524d\u542f\u7528\u5b83\uff0c N \u7981\u7528\uff0cY \u542f\u7528<\/p>\n

cpu enabled on CPU. This allows not to use extra VRAM, sacrificing 20% time of iteration.
\nCPU\u5219\u8bf7\u7528CPU\uff0c\u8fd9\u6837\u53ef\u4ee5\u4e0d\u9002\u7528\u989d\u5916\u7684VRAM\uff0c\u4ece\u800c\u51cf\u5c1120%\u7684\u8fed\u4ee3\u65f6\u95f4\u3002<\/p>\n

Changed help for GAN option:
\n\u4fee\u6539 GAN\u9009\u9879\u7684\u5e2e\u52a9\u4fe1\u606f<\/p>\n

Train the network in Generative Adversarial manner.
\n\u7528GAN\u8bad\u7ec3\u7f51\u7edc<\/p>\n

Forces the neural network to learn small details of the face.
\n\u5f3a\u5236\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u9762\u90e8\u66f4\u5c0f\u7684\u7ec6\u8282\u3002<\/p>\n

Enable it only when the face is trained enough and don’t disable.
\n\u53ea\u6709\u5f53\u5bf9\u8138\u90e8\u8bad\u7ec3\u91cd\u590d\u4e4b\u540e\u542f\u7528\uff0c\u5e76\u4e14\u4e0d\u518d\u5173\u95ed\u3002<\/p>\n

Typical value is 0.1
\n\u5178\u578b\u503c\u4e3a0.1<\/p>\n

improved GAN. Now it produces better skin detail, less patterned aggressive artifacts, works faster.
\n\u6539\u8fdbGAN\uff0c\u73b0\u5728\u53ef\u4ee5\u4ea7\u751f\u66f4\u597d\u7684\u76ae\u80a4\u7ec6\u8282\uff0c\u51cf\u5c11\u56fe\u6848\u5316\u7684\u4f2a\u5f71\uff0c\u5e76\u4e14\u53ef\u4ee5\u66f4\u5feb\u7684\u5de5\u4f5c\u3002<\/p>\n

<\/strong><\/h2>\n

2020\u5e746\u670827\u65e5<\/h2>\n

Extractor:
\n\u63d0\u53d6\u5668<\/p>\n

Extraction now can be continued, but you must specify the same options again.
\n\u63d0\u53d6\u53ef\u4ee5\u4e2d\u65ad\u548c\u7ee7\u7eed\u4e86\u3002\u4f46\u662f\u4f60\u5fc5\u987b\u518d\u6b21\u6307\u5b9a\u76f8\u540c\u7684\u9009\u9879\u3002<\/p>\n

added \u2018Max number of faces from image\u2019 option.
\n\u6dfb\u52a0\u4e86\u201c\u56fe\u7247\u4e2d\u6700\u5927\u4eba\u8138\u6570\u91cf\u201d \u9009\u9879<\/p>\n

If you extract a src faceset that has frames with a large number of faces,
\n\u5982\u679c\u4f60\u63d0\u53d6\u7684\u67d0\u4e2a\u5e27\u4e2d\u6709\u975e\u5e38\u591a\u7684\u4eba\u8138<\/p>\n

it is advisable to set max faces to 3 to speed up extraction.
\n\u5efa\u8bae\u5c06\u6700\u5927\u503c\u8bbe\u7f6e3\uff0c\u53ef\u4ee5\u52a0\u5feb\u63d0\u53d6\u901f\u5ea6\u3002<\/p>\n

0 – unlimited
\n0 \u8868\u793a\u65e0\u9650\u5236<\/p>\n

added \u2018Image size\u2019 option.
\n\u6dfb\u52a0\u56fe\u7247\u5927\u5c0f\u9009\u9879<\/p>\n

The higher image size, the worse face-enhancer works.
\n\u56fe\u7247\u8d8a\u5927\uff0c\u8138\u90e8\u589e\u5f3a\u6548\u679c\u8d8a\u957f\u3002<\/p>\n

Use higher than 512 value only if the source image is sharp enough and the face does not need to be enhanced.
\n\u4ec5\u5f53\u539f\u56fe\u50cf\u8db3\u591f\u6e05\u6670\uff0c\u4e14\u4e0d\u9700\u8981\u589e\u5f3a\uff0c\u624d\u4f7f\u7528\u9ad8\u4e8e512\u7684\u503c<\/p>\n

added \u2018Jpeg quality\u2019 option in range 1-100. The higher jpeg quality the larger the output file size
\n\u6dfb\u52a0JPEG\u8d28\u91cf\u9009\u9879\uff0c\u53d6\u503c\u8303\u56f4\u4e3a1-100\uff0cJPEG\u8d28\u91cf\u8d8a\u9ad8\uff0c\u8f93\u51fa\u6587\u4ef6\u8d8a\u5927\u3002<\/p>\n

Sorter: improved sort by blur and by best faces.
\n\u6392\u5e8f\uff1a\u6539\u8fdb\u4e86\u6a21\u7cca\u6392\u5e8f\u548c\u6700\u4f73\u6392\u5e8f\u3002<\/p>\n

2020\u5e746\u670819\u65e5<\/h2>\n

SAEHD:<\/p>\n

Maximum resolution is increased to 640.
\n\u6700\u9ad8\u50cf\u7d20\u8c03\u6574\u5230640<\/p>\n

\u2018hd\u2019 archi is removed. \u2018hd\u2019 was experimental archi created to remove subpixel shake, but \u2018lr_dropout\u2019 and \u2018disable random warping\u2019 do that better.
\nHD\u7ed3\u6784\u88ab\u79fb\u9664\uff0c HD\u662f\u4e00\u4e2a\u7528\u6765\u89e3\u51b3\u4e9a\u50cf\u7d20\u6296\u52a8\u7684\u8bd5\u9a8c\u6027\u529f\u80fd\uff0c\u4f46\u662f\u2018lr_dropout\u2019\u548c\u2018disable random warping\u2019 \u505a\u5f97\u66f4\u597d<\/p>\n

\u2018uhd\u2019 is renamed to \u2018-u\u2019
\nuhd \u5df2\u7ecf\u91cd\u547d\u540d\u4e3a-u<\/p>\n

dfuhd and liaeuhd will be automatically renamed to df-u and liae-u in existing models.
\n\u5bf9\u4e8e\u5df2\u7ecf\u5b58\u5728\u7684\u6a21\u578b\uff0c\u8f6f\u4ef6\u4f1a\u81ea\u52a8\u5c06dfuhd \u548c liaeuhd \u91cd\u547d\u540d\u4e3adf-u\u548cliae-u<\/p>\n

Added new experimental archi (key -d) which doubles the resolution using the same computation cost.
\n\u6dfb\u52a0\u4e00\u4e2a\u8bd5\u9a8c\u6027\u7684\u7ed3\u6784\u3002\u540c\u6837\u7684\u914d\u7f6e\uff0c\u53ef\u4ee5\u8dd1\u4e24\u500d\u7684\u50cf\u7d20\u3002<\/p>\n

It is mean same configs will be x2 faster, or for example you can set 448 resolution and it will train as 224.
\n\u8fd9\u610f\u5473\u7740\u540c\u6837\u7684\u914d\u7f6e\uff0c\u4f1a\u5feb\u4e24\u500d\uff0c\u6216\u8005\u539f\u5148\u53ea\u80fd\u8dd1224\uff0c\u73b0\u5728\u53ef\u4ee5\u8dd1448<\/p>\n

Strongly recommended not to train from scratch and use pretrained models.
\n\u5899\u88c2\u63a8\u8350\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b<\/p>\n

New archi naming:
\n\u65b0\u7ed3\u6784\u547d\u540d\uff1a<\/p>\n

‘df’ keeps more identity-preserved face.<\/p>\n

‘liae’ can fix overly different face shapes.<\/p>\n

‘-u’ increased likeness of the face.<\/p>\n

‘-d’ (experimental) doubling the resolution using the same computation cost<\/p>\n

‘-d ‘(\u8bd5\u9a8c\u6027) \u76f8\u540c\u7684\u914d\u7f6e\u63d0\u5347\u4e00\u500d\u7684\u50cf\u7d20<\/p>\n

Examples: df, liae, df-d, df-ud, liae-ud, …<\/p>\n

Not the best example of 448 df-ud trained on 11GB:
\n11GB \u8bad\u7ec3448\u7684\u6f14\u793a\uff0c\u8fd9\u8fd8\u4e0d\u662f\u6700\u4f73\u6548\u679c\u3002<\/p>\n

Improved GAN training (GAN_power option). It was used for dst model, but actually we don\u2019t need it for dst.
\n\u6539\u8fdb\u4e86GAN\u8bad\u7ec3\uff08GAN_power\u9009\u9879\uff09\u3002 \u5b83\u4e4b\u524d\u662f\u7528\u4e8edst\u6a21\u578b\uff0c\u4f46\u5b9e\u9645\u4e0a\u6211\u4eec\u4e0d\u9700\u8981\u7528\u5728dst\u3002<\/p>\n

Instead, a second src GAN model with x2 smaller patch size was added, so the overall quality for hi-res models should be higher.
\n\u53d6\u800c\u4ee3\u4e4b\u7684\u662f\uff0c\u6dfb\u52a0\u4e86\u7b2c\u4e8c\u4e2a1\/2\u5927\u5c0f\u7684src GAN\u6a21\u578b\uff0c\u56e0\u6b64\u9ad8\u5206\u8fa8\u7387\u6a21\u578b\u7684\u6574\u4f53\u8d28\u91cf\u5e94\u8be5\u66f4\u9ad8\u3002<\/p>\n

Added option \u2018Uniform yaw distribution of samples (y\/n)\u2019:
\n\u6dfb\u52a0\u9009\u9879Uniform yaw distribution of samples<\/p>\n

Helps to fix blurry side faces due to small amount of them in the faceset.
\n\u7531\u4e8e\u4fa7\u9762\u4e2d\u7684\u4fa7\u9762\u8f83\u5c11\uff0c\u6709\u52a9\u4e8e\u4fee\u590d\u4fa7\u9762\u6a21\u7cca\u3002<\/p>\n

Quick96:<\/p>\n

Now based on df-ud archi and 20% faster.
\n\u9ed8\u8ba4\u7ed3\u6784\u6539\u4e3adf-ud \uff0c\u901f\u5ea6\u63d0\u534720%<\/p>\n

XSeg trainer:<\/p>\n

Improved sample generator.
\n\u63d0\u5347\u5b9e\u4f8b\u751f\u6210<\/p>\n

Now it randomly adds the background from other samples.
\n\u73b0\u5728\u4f1a\u968f\u673a\u6dfb\u52a0\u4ece\u5176\u4ed6\u6837\u4f8b\u4e2d\u7684\u80cc\u666f<\/p>\n

Result is reduced chance of random mask noise on the area outside the face.
\n\u53ef\u4ee5\u964d\u4f4e\u8138\u5916\u90e8\u7684\u968f\u673a\u566a\u58f0<\/p>\n

Now you can specify \u2018batch_size\u2019 in range 2-16.
\n\u73b0\u5728\u53ef\u4ee5\u6307\u5b9abs\uff0c\u53d6\u503c\u8303\u56f42-16<\/p>\n

Reduced size of samples with applied XSeg mask. Thus size of packed samples with applied xseg mask is also reduced.
\n\u5e94\u7528\u7684Xseg\u906e\u7f69\u6837\u4f8b\u53d8\u5c0f\uff0c\u6240\u4ee5\u6253\u5305\u540e\u7684\u7d20\u6750\u6587\u4ef6\u4e5f\u4f1a\u53d8\u5c0f\u3002<\/p>\n

<\/h2>\n

2020\u5e744\u670815\u65e5<\/h2>\n

XSegEditor: added view lock at the center by holding shift in drawing mode.
\nXSegEditor\uff1a\u5728\u7ed8\u56fe\u6a21\u5f0f\u4e0b\u5b89\u88c5shift\u952e\u53ef\u4ee5\u5728\u4e2d\u5fc3\u6dfb\u52a0\u89c6\u56fe\u9501\u5b9a<\/p>\n

Merger: color transfer \u201csot-m\u201d: speed optimization for 5-10%
\n\u5408\u5e76\uff1a\u989c\u8272\u8f6c\u79fb\u201csot-m\u201d\uff1a\u901f\u5ea6\u4f18\u53165-10%<\/p>\n

Fix minor bug in sample loader
\n\u4fee\u590d\u793a\u4f8b\u52a0\u8f7d\u7a0b\u5e8f\u4e2d\u7684\u5c0f\u9519\u8bef<\/p>\n

2020\u5e744\u670814\u65e5<\/h2>\n

Merger: optimizations
\n\u5408\u5e76\uff1a\u4f18\u5316<\/p>\n

color transfer \u2018sot-m\u2019 : reduced color flickering, but consuming x5 more time to process
\n\u989c\u8272\u8f6c\u6362\u201csot-m\u201d\uff1a\u51cf\u5c11\u4e86\u989c\u8272\u95ea\u70c1\uff0c\u4f46\u5904\u7406\u65f6\u95f4\u589e\u52a0\u4e865\u500d<\/p>\n

added mask mode \u2018learned-prd + learned-dst\u2019 \u2013 produces largest area of both dst and predicted masks
\n\u6dfb\u52a0\u7684\u906e\u7f69\u6a21\u5f0f\u201clearn prd+learn dst\u201d-\u751f\u6210dst\u548c\u9884\u6d4b\u906e\u7f69\u7684\u6700\u5927\u9762\u79ef\u3002<\/p>\n

XSegEditor : polygon is now transparent while editing
\nXSegEditor\uff1a\u7f16\u8f91\u65f6\u591a\u8fb9\u5f62\u73b0\u5728\u662f\u900f\u660e\u7684<\/p>\n

New example data_dst.mp4 video
\n\u65b0\u7684dst\u89c6\u9891\u3002<\/p>\n

New official mini tutorial https:\/\/www.youtube.com\/watch?v=1smpMsfC3ls
\n\u65b0\u7684\u5b98\u65b9\u6559\u7a0b<\/p>\n

2020\u5e744\u67086\u65e5<\/h2>\n

Fixes for 16+ cpu cores and large facesets.
\n\u4fee\u590d\u4e8616+cpu\u95ee\u9898\uff0c\u4ee5\u53ca\u5927\u6570\u636e\u96c6\u7684\u95ee\u9898\u3002<\/p>\n

added 5.XSeg) data_dst\/data_src mask for XSeg trainer – remove.bat
\n\u6dfb\u52a0\u811a\u672c<\/p>\n

removes labeled xseg polygons from the extracted frames
\n\u4f5c\u7528\u662f\u4ece\u5df2\u7ecf\u63d0\u53d6\u7684\u5e27\u4e2d\u6ea2\u51fa\u906e\u7f69\u4fe1\u606f\u3002<\/p>\n

2020\u5e744\u67085\u65e5<\/h2>\n

Fixed bug with input dialog in Windows 10
\n\u4fee\u590dWindow10\u4e0b\u9762\u8f93\u5165\u5bf9\u8bdd\u7684\u95ee\u9898\u3002<\/p>\n

Fixed running XSegEditor when directory path contains spaces
\n\u4fee\u590d\u4e86\u5f53\u76ee\u5f55\u8def\u5f84\u5305\u542b\u7a7a\u683c\u65f6\u8fd0\u884cXSegEditor\u7684\u95ee\u9898<\/p>\n

SAEHD: \u2018Face style power\u2019 and \u2018Background style power\u2019 are now available for whole_face
\nSAEHD\uff1a\u5168\u8138\u7c7b\u578b\u4e5f\u53ef\u4ee5\u4f7f\u7528 \u2018Face style power\u2019 \u548c \u2018Background style power\u2019 \u8fd9\u4e24\u4e2a\u53c2\u6570\u3002<\/p>\n

New help messages for these options.
\n\u6b63\u5bf9\u5982\u4e0b\u9009\u9879\u6dfb\u52a0\u4e86\u65b0\u7684\u63d0\u793a\u4fe1\u606f<\/p>\n

XSegEditor: added button \u2018view trained XSeg mask\u2019, so you can see which frames should be masked to improve mask quality.
\nXSegEditor\uff1a\u6dfb\u52a0view trained XSeg mask\u6309\u94ae\uff0c\u56e0\u6b64\uff0c\u60a8\u53ef\u4ee5\u770b\u5230\u54ea\u4e9b\u5e27\u5e94\u8be5\u88ab\u906e\u7f69\u4ee5\u63d0\u9ad8\u906e\u7f69\u8d28\u91cf\u3002<\/p>\n

Merger:
\n\u5408\u6210
\nadded \u2018raw-predict\u2019 mode. Outputs raw predicted square image from the neural network.
\n\u6dfb\u52a0raw-predict\u6a21\u5f0f\uff0c\u4ece\u795e\u7ecf\u7f51\u7edc\u8f93\u51fa\u9884\u6d4b\u7684\u6b63\u65b9\u5f62\u56fe\u50cf\u3002<\/p>\n

mask-mode \u2018learned\u2019 replaced with 3 new modes:
\n\u906e\u7f69\u6a21\u578b\u4e2d\u7684Learned\u88ab\u4e00\u4e0b3\u4e2a\u65b0\u7684\u6a21\u578b\u6240\u66ff\u4ee3\u4e86\uff1a<\/p>\n

\u2018learned-prd\u2019 \u2013 smooth learned mask of the predicted face
\n\u2018learned-prd\u2019 \u2013 \u9884\u6d4b\u4eba\u8138\u7684\u906e\u7f69<\/p>\n

\u2018learned-dst\u2019 \u2013 smooth learned mask of DST face
\n\u2018learned-dst\u2019 \u2013 \u76ee\u6807\u4eba\u8138\u7684\u906e\u7f69<\/p>\n

\u2018learned-prd*learned-dst\u2019 \u2013 smallest area of both (default)
\n\u9884\u6d4b\u533a\u57df\u548cdst\u533a\u57df\u4e24\u8005\u7684\u4ea4\u96c6\u3002\uff08\u9ed8\u8ba4\uff09<\/p>\n

Added new face type : head
\n\u6dfb\u52a0\u65b0\u7684\u8138\u578b\uff1a\u5934<\/p>\n

Now you can replace the head.
\n\u73b0\u5728\u53ef\u4ee5\u6362\u5934\u4e86\u3002<\/p>\n

Example: https:\/\/www.youtube.com\/watch?v=xr5FHd0AdlQ
\n\u793a\u4f8b : https:\/\/www.youtube.com\/watch?v=xr5FHd0AdlQ<\/p>\n

Requirements:
\n\u9700\u8981\uff1a<\/p>\n

Post processing skill in Adobe After Effects or Davinci Resolve.
\nAE\u548c\u8fbe\u82ac\u5947\u7b49\u540e\u671f\u5904\u7406\u6280\u80fd<\/p>\n

Usage:
\n\u4f7f\u7528\uff1a<\/p>\n

Find suitable dst footage with the monotonous background behind head
\n\u627e\u5230\u5408\u9002\u7684dst\u955c\u5934\uff0c\u80cc\u666f\u6bd4\u8f83\u5355\u4e00\u90a3\u79cd\u3002<\/p>\n

Use \u201cextract head\u201d script
\n\u4f7f\u7528\u201cextract head\u201d\u811a\u672c<\/p>\n

Gather rich src headset from only one scene (same color and haircut)
\n\u53ea\u4ece\u4e00\u4e2a\u573a\u666f\u6536\u96c6\u4e30\u5bcc\u7684src\u5934\u50cf\u96c6\uff08\u9700\u8981\u76f8\u540c\u989c\u8272\u548c\u53d1\u578b\uff09<\/p>\n

Mask whole head for src and dst using XSeg editor
\n\u4f7f\u7528XSeg\u7f16\u8f91\u5668\u4e3aSrc\u548cDst\u6807\u6ce8\u6574\u4e2a\u5934\u7684\u906e\u7f69\u3002<\/p>\n

Train XSeg
\n\u8bad\u7ec3Xseg<\/p>\n

Apply trained XSeg mask for src and dst headsets
\n\u4e3aSrc\u548cdst\u5934\u50cf\u5e94\u7528\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684Xseg\u906e\u7f69\u3002<\/p>\n

Train SAEHD using \u2018head\u2019 face_type as regular deepfake model with DF archi. You can use pretrained model for head. Minimum recommended resolution for head is 224.
\n\u8bad\u7ec3SAEHD\u6a21\u578b\uff0c\u8138\u7684\u7c7b\u578b\u4f7f\u7528\u201chead\u201d \uff0c\u7ed3\u6784\u91c7\u7528DF\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528Head\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u5efa\u8bae\u5934\u90e8\u6a21\u578b\u7684\u5206\u8fa8\u7387\u4e0d\u4f4e\u4e8e224\u3002<\/p>\n

Extract multiple tracks, using Merger:
\n\u4f7f\u7528\u5408\u6210\u529f\u80fd\u5408\u6210\u591a\u4e2a\u8f68\u9053\u3002<\/p>\n

a. Raw-rgb<\/p>\n

b. XSeg-prd mask<\/p>\n

c. XSeg-dst mask<\/p>\n

Using AAE or DavinciResolve, do:
\n\u4f7f\u7528AE\u6216\u8005\u8fbe\u82ac\u5947\u5904\u7406\uff1a<\/p>\n<\/div>\n

a. Hide source head using XSeg-prd mask: content-aware-fill, clone-stamp, background retraction, or other technique
\n\u4f7f\u7528Xseg-prd\u906e\u7f69\u9690\u85cf\u6e90\u5934\u50cf\uff1a\u5185\u5bb9\u611f\u77e5\u586b\u5145\u3001\u514b\u9686\u6807\u8bb0\u3001\u80cc\u666f\u6536\u56de\u6216\u5176\u4ed6\u6280\u672f<\/p>\n

b. Overlay new head using XSeg-dst mask
\n\u4f7f\u7528XSeg dst\u8499\u7248\u8986\u76d6\u65b0\u7684\u5934\u50cf<\/p>\n

\n

Warning: Head faceset can be used for whole_face or less types of training only with XSeg masking.
\n\u8b66\u544a\uff1a\u4eba\u5934\u6570\u636e\u96c6\u53ef\u7528\u4e8e\u6574\u8138\u3002\u6216\u8005….\u8fd9\u53e5\u6211\u6ca1\u61c2….<\/p>\n

<\/h2>\n

<\/h2>\n

2020\u5e743\u670830\u65e5<\/h2>\n

New script:
\n\u65b0\u811a\u672c<\/p>\n

5.XSeg) data_dst\/src mask for XSeg trainer – fetch.bat<\/p>\n

Copies faces containing XSeg polygons to aligned_xseg\\ dir.
\n\u62f7\u8d1d\u5305\u542bXseg\u6807\u6ce8\u7684\u5934\u50cf\u5230aligned_xseg\u76ee\u5f55<\/p>\n

Useful only if you want to collect labeled faces and reuse them in other fakes.
\n\u5982\u679c\u4f60\u60f3\u8981\u6536\u96c6\u8fd9\u4e9b\u5df2\u7ecf\u6807\u6ce8\u7684\u4eba\u8138\u5e76\u4e14\u5e94\u7528\u4e8e\u5176\u4ed6\u89c6\u9891\u7684\u65f6\u5019\uff0c\u8fd9\u811a\u672c\u4f1a\u975e\u5e38\u6709\u7528\u3002<\/p>\n

Now you can use trained XSeg mask in the SAEHD training process.<\/p>\n

\u73b0\u5728\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684Xseg\u906e\u7f69\u53ef\u4ee5\u4f53\u73b0\u5728SAEHD\u8bad\u7ec3\u7684\u8fc7\u7a0b\u4e2d\u3002<\/p>\n

It\u2019s mean default \u2018full_face\u2019 mask obtained from landmarks will be replaced with the mask obtained from the trained XSeg model.
\n\u8fd9\u610f\u5473\u7740\uff0c\u4ee5\u524d\u901a\u8fc7\u4eba\u8138\u6807\u8bc6\u83b7\u5f97\u7684\u5168\u8138\u906e\u7f69\u5c06\u88abXseg\u906e\u7f69\u6240\u66ff\u4ee3\u3002<\/p>\n

use\uff1a
\n\u4f7f\u7528\u65b9\u6cd5\uff1a<\/p>\n

5.XSeg.optional) trained mask for data_dst\/data_src – apply.bat
\n\u5e94\u7528\u811a\u672c<\/p>\n

5.XSeg.optional) trained mask for data_dst\/data_src – remove.bat
\n\u79fb\u51fa\u811a\u672c<\/p>\n

Normally you don\u2019t need it. You can use it, if you want to use \u2018face_style\u2019 and \u2018bg_style\u2019 with obstructions.
\n\u901a\u5e38\u60c5\u51b5\u4e0b\u4f60\u4e0d\u9700\u8981\u4f7f\u7528\u5b83\uff0c\u5982\u679c\u4f60\u60f3\u8981\u5728\u906e\u6321\u4e2d\u4f7f\u7528Face_style\u548cbg_style\uff0c\u4f60\u4e5f\u53ef\u4ee5\u4f7f\u7528\u3002<\/p>\n

XSeg trainer : now you can choose type of face
\nXseg\u8bad\u7ec3\uff1a\u73b0\u5728\u53ef\u4ee5\u9009\u62e9\u8138\u578b\u4e86<\/p>\n

XSeg trainer : now you can restart training in \u201coverride settings\u201d
\nXseg\u8bad\u7ec3\uff1a\u73b0\u5728\u4f60\u53ef\u4ee5\u901a\u8fc7\u91cd\u8f7d\u8bbe\u7f6e\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n

Merger: XSeg-* modes now can be used with all types of faces.
\n\u5408\u6210\uff1aXSeg-*\u6a21\u578b\u53ef\u4ee5\u5e94\u7528\u4e8e\u6240\u6709\u8138\u578b\uff08\u503c\u6574\u8138\uff0c\u5168\u8138\uff0c\u534a\u8138\u8fd9\u4e9b\u8138\u578b\uff09<\/p>\n

Therefore old MaskEditor, FANSEG models, and FAN-x modes have been removed,
\n\u4e4b\u524d\u7684MaskEditor\uff08\u906e\u7f69\u7f16\u8f91\u5668\uff09\uff0cFanseg\u6a21\u578b\uff0cFan-X\u6a21\u5f0f\u90fd\u5df2\u7ecf\u88ab\u79fb\u51fa\u3002<\/p>\n

because the new XSeg solution is better, simpler and more convenient, which costs only 1 hour of manual masking for regular deepfake.
\n\u56e0\u4e3a\u5bf9\u4e8e\u666e\u901a\u7684\u6362\u8138\u6765\u8bf4\uff0c\u65b0\u7684XSeg\u89e3\u51b3\u65b9\u6848\u66f4\u597d\u3001\u66f4\u7b80\u5355\u3001\u66f4\u65b9\u4fbf\u3002\u53ea\u9700\u89811\u4e2a\u5c0f\u65f6\u7684\u65f6\u95f4\u505a\u624b\u52a8\u7684\u906e\u7f69\u3002<\/p>\n<\/div>\n

<\/div>\n
<\/div>\n
<\/div>\n
\n

2020\u5e743\u670825\u65e5<\/h2>\n<\/div>\n

SAEHD: added ‘dfuhd’ and ‘liaeuhd’ archi
\nSAEHD: \u6dfb\u52a0\u4e86dfuhd\u548cliaeuhd\u7ed3\u6784<\/p>\n

uhd version is lighter than ‘HD’ but heavier than regular version.
\nuhd\u7248\u672c\u6bd4HD\u66f4\u52a0\u8f7b\u91cf\u7ea7\uff0c\u6bd4\u5e38\u89c4\u7248\u672c\u66f4\u52a0\u91cd\u91cf\u7ea7\u3002<\/p>\n

liaeuhd provides more “src-like” result
\nliaeuhd\u63d0\u4f9b\u4e86\u66f4\u52a0\u50cfsrc\u7684\u7ed3\u679c
\ncomparison:
\n\u6bd4\u8f83<\/p>\n

<\/div>\n
liae:<\/div>\n
<\/div>\n
liaeuhd:<\/div>\n
<\/div>\n
\n

added new XSegEditor !
\n\u6dfb\u52a0\u65b0\u7684Xseg\u7f16\u8f91\u5668\u3002<\/p>\n

here new whole_face + XSeg workflow:
\n\u4ee5\u4e0b\u4e3a\u6574\u8138\u548cXseg\u7684\u5de5\u4f5c\u6d41\u7a0b\u3002<\/p>\n

with XSeg model you can train your own mask segmentator for dst(and\/or src) faces
\n\u4f7f\u7528Xseg\u6a21\u578b\uff0c\u4f60\u53ef\u4ee5\u81ea\u5df1\u4e3aDst\u8138\u90e8\u8bad\u7ec3\u906e\u7f69\u5206\u5272\u5668<\/p>\n

that will be used by the merger for whole_face.
\n\u5408\u6210\u7684\u65f6\u5019\u4f1a\u5e94\u7528\u3002<\/p>\n

Instead of using a pretrained segmentator model (which does not exist),
\n\u66ff\u4ee3\u4e86\u4f7f\u7528\u9884\u8bad\u7ec3\u906e\u7f69\u6a21\u578b\u7684\u65b9\u5f0f\uff08\u4e0d\u66fe\u5b58\u5728\u8fc7\u7684\u90a3\u79cd\u65b9\u5f0f\uff0c\u6b64\u5904\u662f\u5e7d\u9ed8\uff1f\uff09<\/p>\n

you control which part of faces should be masked.
\n\u4f60\u53ef\u4ee5\u63a7\u5236\u54ea\u4e00\u90e8\u5206\u5e94\u8be5\u88ab\u906e\u6321\uff08\u5f53\u7136\uff0c\u5de5\u4f5c\u91cf\u4e5f\u5927\u4e86\u4e0d\u5c11\u3002\uff09<\/p>\n

new scripts:
\n\u65b0\u7684\u811a\u672c\u3002<\/p>\n

5.XSeg) data_dst edit masks.bat<\/p>\n

5.XSeg) data_src edit masks.bat<\/p>\n

5.XSeg) train.bat<\/p>\n

Usage:
\n\u4f7f\u7528\uff1a
\nunpack dst faceset if packed
\n\u6570\u636e\u96c6\u5982\u679c\u6253\u5305\u4e86\uff0c\u8bf7\u5148\u89e3\u538b\u3002<\/p>\n

run 5.XSeg) data_dst edit masks.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

Read tooltips on the buttons (en\/ru\/zn languages are supported)
\n\u9605\u8bfb\u63d0\u793a\uff08\u652f\u6301\u82f1\u6587\uff0c\u4fc4\u8bed\uff0c\u4e2d\u6587\uff09<\/p>\n

mask the face using include or exclude polygon mode.
\n\u4f7f\u7528\u5305\u542b\u548c\u6392\u9664\u7684\u65b9\u5f0f\u6807\u6ce8\u8138\u90e8\u3002<\/p>\n

repeat for 50\/100 faces,
\n\u91cd\u590d 50\u5230100\u5f20\u8138
\n!!! you don’t need to mask every frame of dst
\n\u65e0\u9700\u6807\u6ce8\u6240\u6709\u5e27
\nonly frames where the face is different significantly,
\n\u53ea\u8981\u5173\u6ce8\u90a3\u4e9b\u6709\u5dee\u5f02\u7684\u8138\u3002
\nfor example:
\n\u4e3e\u4f8b
\nclosed eyes
\n\u95ed\u773c
\nchanged head direction
\n\u6539\u53d8\u8138\u7684\u65b9\u5411
\nchanged light
\n\u6539\u53d8\u706f\u5149
\nthe more various faces you mask, the more quality you will get
\n\u6807\u6ce8\u8d8a\u591a\uff0c\u6548\u679c\u8d8a\u597d\u3002<\/p>\n

Start masking from the upper left area and follow the clockwise direction.<\/p>\n

Keep the same logic of masking for all frames, for example:<\/p>\n

the same approximated jaw line of the side faces, where the jaw is not visible<\/p>\n

the same hair line<\/p>\n

Mask the obstructions using exclude polygon mode.<\/p>\n

run XSeg) train.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

train the model
\n\u8bad\u7ec3\u6a21\u578b<\/p>\n

Check the faces of ‘XSeg dst faces’ preview.
\n\u67e5\u770b\u9884\u89c8\u56fe<\/p>\n

if some faces have wrong or glitchy mask, then repeat steps:
\n\u5982\u679c\u53d1\u73b0\u95ee\u9898\uff0c\u7ee7\u7eed\u4ee5\u4e0b\u6b65\u9aa4
\nrun edit<\/p>\n

find these glitchy faces and mask them<\/p>\n

train further or restart training from scratch<\/p>\n

Restart training of XSeg model is only possible by deleting all ‘model\\XSeg_*’ files.
\n\u53ea\u6709\u5220\u9664\u4e86model\\XSeg-*\u7684\u6587\u4ef6\u624d\u80fd\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n

If you want to get the mask of the predicted face (XSeg-prd mode) in merger,
\n\u5982\u679c\u9700\u8981\u4f7f\u7528predicted face\u906e\u7f69<\/p>\n

you should repeat the same steps for src faceset.
\n\u4f60\u9700\u8981\u5bf9src\u6570\u636e\u96c6\u91cd\u590d\u4e00\u6837\u7684\u6b65\u9aa4\u3002<\/p>\n

New mask modes available in merger for whole_face:
\n\u8fd9\u4e2a\u906e\u7f69\u5728\u5408\u6210\u7684\u65f6\u5019\u88ab\u5e94\u7528\uff0c\u4ec5\u9488\u5bf9\u6574\u8138\u3002\u5177\u4f53\u9009\u9879\u5982\u4e0b\uff1a<\/p>\n

XSeg-prd – XSeg mask of predicted face -> faces from src faceset should be labeled<\/p>\n

XSeg-dst – XSeg mask of dst face -> faces from dst faceset should be labeled<\/p>\n

XSeg-prd*XSeg-dst – the smallest area of both<\/p>\n

if workspace\\model folder contains trained XSeg model, then merger will use it,
\n\u5982\u679cmodel\u76ee\u5f55\u4e0b\u5305\u542bXseg\u6a21\u578b\uff0c\u5219\u88ab\u5e94\u7528\u3002<\/p>\n

otherwise you will get transparent mask by using XSeg-* modes.
\n\u5426\u5219\uff0c\u60a8\u5c06\u901a\u8fc7\u4f7f\u7528XSeg-*\u6a21\u5f0f\u83b7\u5f97\u900f\u660e\u63a9\u7801\u3002<\/p>\n

Some screenshots:
\n\u4e00\u4e9b\u622a\u56fe
\nXSegEditor:<\/p>\n<\/div>\n

<\/div>\n
<\/div>\n
trainer :<\/div>\n
<\/div>\n
merger :<\/div>\n
<\/div>\n
<\/div>\n
<\/div>\n
example of the fake using 13 segmented dst faces<\/div>\n
\u4e00\u4e2a\u4f7f\u7528segmented\u7684\u52a8\u6001\u56fe<\/div>\n
<\/div>\n
https:\/\/i.imgur.com\/wmvyizU.gifv<\/a><\/div>\n

<\/h2>\n

2020\u5e743\u670815\u65e5<\/h2>\n

\u8fd9\u4e2a\u7248\u672c\u4e3b\u8981\u662f\u52a0\u5165\u4e86Xseg\u6a21\u578b\uff0c\u5177\u4f53\u64cd\u4f5c\u53c2\u8003\uff1ahttps:\/\/www.deepfaker.xyz\/?p=1648<\/p>\n

global fixes
\n\u5168\u5c40\u4fee\u590d<\/p>\n

SAEHD: removed option learn_mask, it is now enabled by default
\nSAEHD: \u79fb\u9664learn_mask\uff0c\u53d8\u4e3a\u9ed8\u8ba4\u542f\u7528\u3002<\/p>\n

removed liaech arhi
\n\u79fb\u9664 Liaech\u7ed3\u6784<\/p>\n

removed support of extracted(aligned) PNG faces. Use old builds to convert from PNG to JPG.
\n\u79fb\u9664\u5bf9PNG\u56fe\u7247\u7684\u652f\u6301\u3002\u53ef\u4ee5\u4f7f\u7528\u8001\u811a\u672c\u628aPNG\u8f6c\u6210JPG<\/p>\n

added XSeg model.
\n\u6dfb\u52a0XSeg\u6a21\u578b<\/p>\n

with XSeg model you can train your own mask segmentator of dst(and src) faces
\n\u4f7f\u7528Xseg\u6a21\u578b\uff0c\u4f60\u53ef\u4ee5\u8bad\u7ec3\u4f60\u81ea\u5df1\u7684\u906e\u7f69<\/p>\n

that will be used in merger for whole_face.
\n\u8fd9\u4e2a\u4f1a\u5728\u5408\u6210\u7684\u65f6\u5019\u88ab\u5e94\u7528<\/p>\n

Instead of using a pretrained model (which does not exist),
\n\u76f8\u6bd4\u4e8e\u4f7f\u7528\u4e00\u4e2a\u9884\u8bad\u7ec3\u7684\u6a21\u578b\uff08\u4e0d\u5b58\u5728\u7684\uff01\uff09<\/p>\n

you control which part of faces should be masked.
\n\u73b0\u5728\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u63a7\u5236\u54ea\u4e00\u90e8\u5e94\u8be5\u662f\u906e\u7f69\u3002<\/p>\n

Workflow is not easy, but at the moment it is the best solution
\n\u4f7f\u7528\u8fc7\u7a0b\u5e76\u4e0d\u7b80\u5355\u3002\u4f46\u662f\u5f53\u524d\u662f\u6700\u597d\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n

for obtaining the best quality of whole_face’s deepfakes using minimum effort
\n\u4ee5\u6700\u5c0f\u7684\u52aa\u529b\u83b7\u5f97\u6574\u4e2a\u6700\u4f73\u7684\u6574\u8138\u66ff\u6362\u6548\u679c\u3002
\nwithout rotoscoping in AfterEffects.
\n\u65e0\u9700AE\u914d\u5408<\/p>\n

new scripts:
\n\u65b0\u811a\u672c<\/p>\n

XSeg) data_dst edit.bat
\nXSeg) data_dst merge.bat
\nXSeg) data_dst split.bat
\nXSeg) data_src edit.bat
\nXSeg) data_src merge.bat
\nXSeg) data_src split.bat
\nXSeg) train.bat<\/p>\n

Usage:
\n\u4f7f\u7528\u65b9\u6cd5\uff1a<\/p>\n

unpack dst faceset if packed
\n\u5982\u679c\u6253\u5305\u4e86\u8bf7\u5148\u89e3\u538b<\/p>\n

run XSeg) data_dst split.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

this scripts extracts (previously saved) .json data from jpg faces to use in label tool.
\n\u8fd9\u4e2a\u811a\u672c\u4f1a\u73bb\u7483\u51fa\u4e00\u4e2a.json \u6587\u4ef6\u3002<\/p>\n

run XSeg) data_dst edit.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

new tool ‘labelme’ is used
\n\u6b64\u65f6\uff0c\u65b0\u7684\u5de5\u5177labelme\u5c06\u88ab\u5e94\u7528\u3002<\/p>\n

use polygon (CTRL-N) to mask the face
\nCtrl+n \u65b0\u5efa\u906e\u7f69
\nname polygon “1” (one symbol) as include polygon
\n1\u4ee3\u8868\u5305\u62ec
\nname polygon “0” (one symbol) as exclude polygon
\n0\u4ee3\u8868\u6392\u9664<\/p>\n

‘exclude polygons’ will be applied after all ‘include polygons’
\n\u6392\u9664\u5c06\u5728\u5305\u62ec\u4e4b\u540e\u5e94\u7528<\/p>\n

Hot keys:
\n\u70ed\u952e\uff1a<\/p>\n

ctrl-N create polygon
\n\u521b\u5efa
\nctrl-J edit polygon
\n\u7f16\u8f91<\/p>\n

A\/D navigate between frames
\n\u4e0a\u4e00\u4e2a\u4e0b\u4e00\u4e2a<\/p>\n

ctrl + mousewheel image zoom
\nCtrl+\u9f20\u6807\u6eda\u8f6e\u5b9e\u73b0\u7f29\u653e\u3002<\/p>\n

mousewheel vertical scroll
\n\u6eda\u8f6e\u53ef\u4ee5\u7eb5\u5411\u79fb\u52a8<\/p>\n

alt+mousewheel horizontal scroll
\nalt+\u6eda\u8f6e\u53ef\u4ee5\u6a2a\u5411\u79fb\u52a8\u3002<\/p>\n

repeat for 10\/50\/100 faces,
\n\u91cd\u590d10\/50\/100\u5f20\u8138\u3002<\/p>\n

you don’t need to mask every frame of dst,
\n\u4f60\u4e0d\u9700\u8981\u6807\u6ce8\u6240\u6709\u7684\u5934\u50cf<\/p>\n

only frames where the face is different significantly,
\n\u4ec5\u4ec5\u5904\u7406\u90a3\u4e9b\u660e\u663e\u4e0d\u540c\u7684\u5373\u53ef\u3002<\/p>\n

for example:
\n\u4e3e\u4e2a\u6817\u5b50\uff1a<\/p>\n

closed eyes
\n\u95ed\u4e0a\u773c<\/p>\n

changed head direction
\n\u6539\u7f16\u5934\u90e8\u65b9\u5411<\/p>\n

changed light
\n\u6539\u53d8\u4e86\u706f\u5149\u3002<\/p>\n

the more various faces you mask, the more quality you will get
\n\u4f60\u5904\u7406\u7684\u56fe\u7247\u8d8a\u591a\uff0c\u83b7\u5f97\u7684\u6548\u679c\u8d8a\u597d\u3002<\/p>\n

Start masking from the upper left area and follow the clockwise direction.
\n\u4ece\u5de6\u4e0a\u65b9\u5f00\u59cb\uff0c\u6cbf\u7740\u987a\u65f6\u9488\u65b9\u5411<\/p>\n

Keep the same logic of masking for all frames, for example:
\n\u5bf9\u6240\u6709\u5e27\u90fd\u4f7f\u7528\u76f8\u540c\u7684\u5c4f\u853d\u903b\u8f91\uff0c\u4f8b\u5982\uff1a<\/p>\n

the same approximated jaw line of the side faces, where the jaw is not visible<\/p>\n

the same hair line
\nMask the obstructions using polygon with name “0”.<\/p>\n

run XSeg) data_dst merge.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

this script merges .json data of polygons into jpg faces,
\n\u8fd9\u4e2a\u811a\u672c\u5427.json\u6587\u4ef6\u63d2\u5165\u5230\u56fe\u7247\u4e2d\u3002<\/p>\n

therefore faceset can be sorted or packed as usual.
\n\u8fd9\u6837\u5c31\u53ef\u4ee5\u5bf9\u56fe\u7247\u8fdb\u884c\u6392\u5e8f\u548c\u6253\u5305\u4e86\u3002<\/p>\n

run XSeg) train.bat
\n\u8fd0\u884c\u811a\u672c<\/p>\n

train the model
\n\u8bad\u7ec3\u6a21\u578b<\/p>\n

Check the faces of ‘XSeg dst faces’ preview.
\n\u67e5\u770bXseg\u9884\u89c8\u56fe<\/p>\n

if some faces have wrong or glitchy mask, then repeat steps:
\n\u5982\u679c\u6709\u4e9b\u8138\u662f\u9519\u8bef\uff0c\u6216\u8005\u6709\u70b9\u5c0f\u95ee\u9898\uff0c\u90a3\u4e48\u91cd\u590d\u4e0b\u9762\u7684\u6b65\u9aa4\u3002<\/p>\n

split
\n\u5265\u79bb<\/p>\n

run edit
\n\u7f16\u8f91<\/p>\n

find these glitchy faces and mask them
\n\u53d1\u73b0\u6709\u95ee\u9898\u7684\u4eba\u8138\uff0c\u6dfb\u52a0\u906e\u7f69<\/p>\n

merge
\n\u5408\u6210<\/p>\n

train further or restart training from scratch
\n\u7ee7\u7eed\u8bad\u7ec3\u6216\u8005\u91cd\u65b0\u5f00\u59cb\u8bad\u7ec3<\/p>\n

Restart training of XSeg model is only possible by deleting all ‘model\\XSeg_*’ files.
\n\u53ea\u6709\u901a\u8fc7\u5220\u9664model\\Xseg\u5f00\u5934\u7684\u6587\u4ef6\uff0c\u624d\u80fd\u91cd\u65b0\u8bad\u7ec3<\/p>\n

If you want to get the mask of the predicted face in merger,
\n\u5982\u679c\u4f60\u60f3\u5728\u5408\u5e76\u4e2d\u5f97\u5230\u9884\u671f\u7684\u906e\u7f69\uff0c<\/p>\n

you should repeat the same steps for src faceset.
\n\u60a8\u5e94\u8be5\u5bf9src faceset\u4f7f\u7528\u540c\u6837\u7684\u6b65\u9aa4\u3002<\/p>\n

New mask modes available in merger for whole_face:
\n\u6574\u8138\u6a21\u578b\u5728\u5408\u6210\u7684\u65f6\u5019\u4f1a\u51fa\u73b0\u60f3\u7684\u906e\u7f69\u53c2\u6570\u3002<\/p>\n

XSeg-prd – XSeg mask of predicted face -> faces from src faceset should be labeled
\nXSeg-dst – XSeg mask of dst face -> faces from dst faceset should be labeled
\nXSeg-prd*XSeg-dst – the smallest area of both<\/p>\n

if workspace\\model folder contains trained XSeg model, then merger will use it,
\n\u5982\u679cmodel\u76ee\u5f55\u4e0b\u6709XSeg\u6a21\u578b\u3002\u5408\u6210\u7684\u65f6\u5019\u5c31\u4f1a\u4f7f\u7528\u3002<\/p>\n

otherwise you will get transparent mask by using XSeg-* modes.
\n\u5426\u8005\u4f60\u4f1a\u901a\u8fc7Xseg-*\u6a21\u578b\u83b7\u5f97\u900f\u660e\u906e\u7f69\u3002<\/p>\n

Some screenshots:
\n\u4e00\u4e9b\u622a\u56fe\u3002<\/p>\n

label tool: https:\/\/i.imgur.com\/aY6QGw1.jpg
\ntrainer : https:\/\/i.imgur.com\/NM1Kn3s.jpg
\nmerger : https:\/\/i.imgur.com\/glUzFQ8.jpg<\/p>\n

example of the fake using 13 segmented dst faces
\n: https:\/\/i.imgur.com\/wmvyizU.gifv<\/p>\n

 <\/p>\n

2020\u5e742\u670828\u65e5<\/h2>\n

Extractor:
\n\u63d0\u53d6<\/p>\n

image size for all faces is now 512
\n\u63d0\u53d6\u7684\u56fe\u7247\u5c3a\u5bf8\u4ece256\u63d0\u5347\u5230512<\/p>\n

fix RuntimeWarning during the extraction process
\n\u4fee\u590d\u63d0\u53d6\u8fc7\u7a0b\u4e2d\u51fa\u73b0RuntimeWarning\u7684\u95ee\u9898\u3002<\/p>\n

SAEHD:<\/p>\n

max resolution is now 512
\n\u6a21\u578b\u53c2\u6570\u4e2d\uff0c\u6700\u9ad8\u50cf\u7d20\u6539\u6210512\uff0832G V100\u540c\u6837\u8868\u793a\u73a9\u4e0d\u8d77\uff09<\/p>\n

fix hd arhitectures. Some decoder\u2019s weights haven\u2019t trained before.
\n\u4fee\u590dHD\u67b6\u6784\uff0c\u4e4b\u524d\u4e00\u4e9b\u89e3\u7801\u5668\u6743\u91cd\u6ca1\u6709\u88ab\u8bad\u7ec3\u5230\u3002<\/p>\n

new optimized training:
\n\u65b0\u7684\u4f18\u5316\u8bad\u7ec3<\/p>\n

for every <batch_size*16> samples,
\n\u5bf9\u4e8e\u6bcf\u4e2a<batch_size*16>\u6837\u672c\uff0c<\/p>\n

model collects <batch_size> samples with the highest error and learns them again
\n\u6a21\u578b\u6536\u96c6\u8bef\u5dee\u6bd4\u8f83\u5927\u7684\u6837\u4f8b\u8fdb\u884c\u91cd\u65b0\u5b66\u4e60\u3002<\/p>\n

therefore hard samples will be trained more often
\n\u56e0\u6b64\uff0c\u6bd4\u8f83\u96be\u7684\u6837\u672c\u88ab\u8bad\u7684\u66f4\u72e0\u3002<\/p>\n

\u2018models_opt_on_gpu\u2019 option is now available for multigpus (before only for 1 gpu)
\n\u201cmodels_opt_on_gpu\u201d\u9009\u9879\u73b0\u5728\u53ef\u7528\u4e8e\u591a\u663e\u5361\uff08\u4ee5\u524d\u4ec5\u7528\u5355\u663e\u5361\uff09<\/p>\n

fix \u2018autobackup_hour\u2019
\n\u4fee\u590d\u201c\u81ea\u52a8\u5907\u4efd\u65f6\u95f4\u201d<\/p>\n

<\/h2>\n

<\/h2>\n

2020\u5e742\u670823\u65e5<\/h2>\n

SAEHD: pretrain option is now available for whole_face type
\nSAEHD: \u9884\u8bad\u7ec3\u652f\u6301\u201c\u6574\u8138\u201d<\/p>\n

fix sort by abs difference
\n\u4fee\u590dabs\u5dee\u5f02\u6392\u5e8f<\/p>\n

fix sort by yaw\/pitch\/best for whole_face\u2019s
\n\u4fee\u590d\u6574\u8138\u7684yaw\/pitch\/best\u6392\u5e8f\u3002<\/p>\n

2020\u5e742\u670821\u65e5<\/h2>\n

Trainer: decreased time of initialization
\n\u8bad\u7ec3\uff1a\u51cf\u5c11\u521d\u59cb\u5316\u65f6\u95f4<\/p>\n

Merger: fixed some color flickering in overlay+rct mode
\n\u5408\u6210\uff1a\u4fee\u590doverlay+rct\u6a21\u5f0f\u4e0b\u7684\u4e00\u4e9b\u95ea\u7f29\u7684\u95ee\u9898\u3002<\/p>\n

SAEHD:<\/p>\n

added option Eyes priority (y\/n)
\n\u6dfb\u52a0\u773c\u90e8\u4f18\u5148\u9009\u9879<\/p>\n

Helps to fix eye problems during training like \u201calien eyes\u201d
\n\u6709\u52a9\u4e8e\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u89e3\u51b3\u201c\u5f02\u5f62\u773c\u201d<\/p>\n

and wrong eyes direction ( especially on HD architectures )
\n\u548c\u773c\u795e\u4e0d\u5bf9\u7684\u95ee\u9898\uff08\u5c24\u5176\u662f\u5728HD\u67b6\u6784\u4e0a\uff09<\/p>\n

by forcing the neural network to train eyes with higher priority.
\n\u901a\u8fc7\u5f3a\u5236\u795e\u7ecf\u7f51\u7edc\u4ee5\u66f4\u9ad8\u7684\u4f18\u5148\u7ea7\u6765\u8bad\u7ec3\u773c\u775b\u3002<\/p>\n

before\/after https:\/\/i.imgur.com\/YQHOuSR.jpg
\n\u4e4b\u524d\/\u4e4b\u540e<\/p>\n

added experimental face type \u2018whole_face\u2019
\n\u6dfb\u52a0\u8bd5\u9a8c\u6027\u8138\u578b\u201c\u6574\u8138\u201d\u3002<\/p>\n

Basic usage instruction: https:\/\/i.imgur.com\/w7LkId2.jpg
\n\u57fa\u672c\u7528\u6cd5\u8bf4\u660e\uff1a<\/p>\n

\u2018whole_face\u2019 requires skill in Adobe After Effects.
\n\u6574\u8138\u9700\u8981 AE\u6280\u80fd<\/p>\n

For using whole_face you have to extract whole_face\u2019s by using
\n\u4f7f\u7528\u6574\u8138\u4f60\u5fc5\u987b\u63d0\u53d6\u6574\u8138\u56fe\u7247\uff0c\u901a\u8fc7\u4ee5\u4e0b\u4e24\u4e2a\u811a\u672c\u3002<\/p>\n

4) data_src extract whole_face
\nand
\n5) data_dst extract whole_face<\/p>\n

Images will be extracted in 512 resolution, so they can be used for regular full_face\u2019s and half_face\u2019s.
\n\u56fe\u7247\u4f1a\u6309512\u50cf\u7d20\u8fdb\u884c\u63d0\u53d6\u3002\u4ed6\u4eec\u540c\u6837\u4e5f\u9002\u7528\u4e8e\u5e38\u89c4\u7684\u5168\u8138\u548c\u534a\u8138\u3002<\/p>\n

\u2018whole_face\u2019 covers whole area of face include forehead in training square,
\n\u6574\u8138\u6db5\u76d6\u8138\u90e8\u7684\u6574\u4e2a\u533a\u57df\u5305\u542b\u989d\u5934\u3002<\/p>\n

but training mask is still \u2018full_face\u2019
\n\u4f46\u662f\u8bad\u7ec3\u906e\u7f69\u4f9d\u65e7\u662f\u5168\u8138<\/p>\n

therefore it requires manual final masking and composing in Adobe After Effects.
\n\u56e0\u6b64\u9700\u8981\u4f60\u901a\u8fc7AE\u624b\u52a8\u906e\u7f69\u548c\u5408\u6210\u3002<\/p>\n

added option \u2018masked_training\u2019
\n\u6dfb\u52a0\u906e\u7f69\u8bad\u7ec3
\nThis option is available only for \u2018whole_face\u2019 type.
\n\u8fd9\u4e2a\u9009\u9879\u4ec5\u9002\u7528\u4e8e\u6574\u8138
\nDefault is ON.
\n\u9ed8\u8ba4\u5f00\u542f
\nMasked training clips training area to full_face mask,
\n\u906e\u7f69\u8bad\u7ec3\u53ef\u4ee5\u628a\u8bad\u7ec3\u533a\u57df\u526a\u8f91\u5230\u5168\u8138\u906e\u7f69\u3002<\/p>\n

thus network will train the faces properly.
\n\u8fd9\u6837\u7f51\u7edc\u5c31\u53ef\u4ee5\u6b63\u786e\u8bad\u7ec3\u8138\u90e8<\/p>\n

When the face is trained enough, disable this option to train all area of the frame.
\n\u5f53\u8bad\u7ec3\u8db3\u591f\u591a\u7684\u65f6\u5019\uff0c\u5173\u95ed\u8fd9\u4e2a\u9009\u9879\uff0c\u8ba9\u7f51\u7edc\u53bb\u8bad\u7ec3\u6240\u6709\u533a\u57df\u3002<\/p>\n

Merge with \u2018raw-rgb\u2019 mode, then use Adobe After Effects to manually mask, tune color, and compose whole face include forehead.
\n\u4f7f\u7528Raw-rgb\u6a21\u5f0f\u5408\u6210\uff0c\u7136\u540e\u4f7f\u7528AE\u624b\u52a8\u906e\u7f69\u8c03\u6574\u989c\u8272\uff0c\u7136\u540e\u5408\u6210\uff0c\u8fd9\u6837\u989d\u5934\u4e5f\u80fd\u88ab\u6362\u6389\u3002<\/p>\n

 <\/p>\n

2020\u5e742\u67083\u65e5 (DFL2.0)<\/h2>\n

\u201cEnable autobackup\u201d option is replaced by
\n\u542f\u52a8\u5907\u4efd\u9009\u9879\u88ab\u66ff\u6362\u4e3a<\/p>\n

\u201cAutobackup every N hour\u201d 0..24 (default 0 disabled), Autobackup model files with preview every N hour
\n\u51e0\u4e2a\u5c0f\u65f6\u5907\u4efd\u4e00\u6b21\u3002<\/p>\n

Merger:
\n\u5408\u6210<\/p>\n

\u2018show alpha mask\u2019 now on \u2018V\u2019 button
\nshow alpha mask \u7684\u5feb\u6377\u952e\u6539\u6210V<\/p>\n

\u2018super resolution mode\u2019 is replaced by
\n\u8d85\u7ea7\u5206\u8fa8\u7387super resolution mode \u53c2\u6570\u88ab\u66ff\u6362\u4e3a<\/p>\n

\u2018super resolution power\u2019 (0..100) which can be modified via \u2018T\u2019 \u2018G\u2019 buttons
\n\u8d85\u7ea7\u5206\u8fa8\u7387supersuper resolution power\uff0c \u53d6\u503c\u8303\u56f4\u4e3a0~100\uff0c\u53ef\u4ee5\u901a\u8fc7\u5feb\u6377\u952eT\u548cG\u8fdb\u884c\u4fee\u6539\u3002<\/p>\n

default erode\/blur values are 0.
\n\u9ed8\u8ba4\u7684erode\/blur\u88ab\u8bbe\u7f6e\u4e3a0<\/p>\n

new multiple faces detection log: https:\/\/i.imgur.com\/0XObjsB.jpg
\n\u65b0\u7684\u591a\u8138\u68c0\u6d4b\u65e5\u5fd7:\u00a0https:\/\/i.imgur.com\/0XObjsB.jpg<\/p>\n

now uses all available CPU cores ( before max 6 )
\n\u4f7f\u7528\u6240\u6709\u7684CPU\u5185\u6838\uff0c\u4e4b\u524d\u6700\u591a6\u4e2aCPU<\/p>\n

so the more processors, the faster the process will be.
\n\u56e0\u6b64\uff0c\u5904\u7406\u5668\u8d8a\u591a\uff0c\u901f\u5ea6\u8d8a\u5feb<\/p>\n

2020\u5e742\u67081\u65e5<\/h2>\n

Merger:
\n\u5408\u6210<\/p>\n

increased speed
\n\u63d0\u5347\u901f\u5ea6<\/p>\n

improved quality
\n\u63d0\u5347\u8d28\u91cf<\/p>\n

SAEHD: default archi is now \u2018df\u2019
\nSAEHD\uff1a\u9ed8\u8ba4\u7ed3\u6784\u6539\u56dedf<\/p>\n

2020\u5e741\u670830\u65e5<\/h2>\n

removed use_float16 option
\n\u79fb\u9664use_float16\u00a0 \u9009\u9879<\/p>\n

fix MultiGPU training
\n\u4fee\u590d\u591a\u5361\u8bad\u7ec3\u7684\u95ee\u9898\u3002<\/p>\n

2020\u5e741\u670829\u65e5<\/h2>\n

MultiGPU training:
\n\u591a\u5361\u8bad\u7ec3\uff1a<\/p>\n

fixed CUDNN_STREAM errors.
\n\u4fee\u590dCudnn\u7684\u9519\u8bef<\/p>\n

speed is significantly increased.
\n\u901f\u5ea6\u5927\u5927\u63d0\u5347<\/p>\n

Trainer: added key \u2018b\u2019 : creates a backup even if the autobackup is disabled.
\n\u8bad\u7ec3\uff1a\u6dfb\u52a0\u4e86B\u952e\uff0c\u5373\u4f7f\u7981\u7528\u4e86\u81ea\u52a8\u5907\u4efd\uff0c\u4e5f\u4f1a\u521b\u5efa\u5907\u4efd\u3002<\/p>\n

2020\u5e741\u670828\u65e5<\/h2>\n

optimized face sample generator, CPU load is significantly reduced
\n\u4f18\u5316\u4eba\u7c7b\u6837\u672c\u751f\u6210\u5668\uff0cCPU\u590d\u5236\u663e\u8457\u964d\u4f4e<\/p>\n

fix of update preview for history after disabling the pretrain mode
\n\u4fee\u590d\u7ed3\u675f\u9884\u8bad\u7ec3\u540e\u5386\u53f2\u66f2\u7ebf\u7684\u663e\u793a\u95ee\u9898\u3002<\/p>\n

SAEHD:<\/p>\n

added new option
\n\u6dfb\u52a0\u65b0\u7684\u9009\u9879<\/p>\n

GAN power 0.0 .. 10.0
\nGAN\u5f3a\u5ea6<\/p>\n

Train the network in Generative Adversarial manner.
\n\u7528\u751f\u6210\u5bf9\u6297\u5f0f\u7f51\u7edc\u6765\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n

Forces the neural network to learn small details of the face.
\n\u5f3a\u5236\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u9762\u90e8\u5c0f\u7ec6\u8282\u3002<\/p>\n

You can enable\/disable this option at any time,
\n\u4f60\u53ef\u4ee5\u5728\u4efb\u610f\u65f6\u95f4\u5185\u542f\u7528\u6216\u5173\u95ed\u8fd9\u4e2a\u9009\u9879\u3002<\/p>\n

but better to enable it when the network is trained enough.
\n\u4f46\u6700\u597d\u7684\u65b9\u6cd5\u662f\uff0c\u6a21\u578b\u5df2\u7ecf\u8bad\u7ec3\u5230\u65e0\u6cd5\u63d0\u5347\u4e86\u518d\u542f\u7528\u3002<\/p>\n

Typical value is 1.0
\n\u5178\u578b\u503c\u4e3a1.0<\/p>\n

GAN power with pretrain mode will not work.
\n\u9884\u8bad\u7ec3\u6a21\u578b\u65e0\u6cd5\u542f\u7528\u8fd9\u4e2a\u53c2\u6570\u3002<\/p>\n

Example of enabling GAN on 81k iters +5k iters
\n\u8bad\u7ec381000\u6b21\u540e\u542f\u7528GAN\u53c2\u6570\uff0c\u7136\u540e\u5728\u8bad\u7ec35000\u6b21\u7684\u6548\u679c<\/p>\n