DeepFaceLab 2021 版本记录

 

 

2021年11月20日

修复了合并中的 rct 颜色转移
固定模型导出。

 

2021年10月20日

SAEHD、AMP:随机缩放比例增加到 -0.15+0.15。 提升lr_dropout 能力以达到较低的损失值。

SAEHD:更改了 bg_style_power 的算法。可以更好的融合人脸

添加的选项
随机色调(Random hue)
饱和度(saturation)
光强度(light )
该参数仅在神经网络的输入处应用于 src 人脸集。

在面部交换期间稳定颜色扰动。 通过选择 src faceset 中最接近的颜色来降低颜色转移的质量。 因此 src faceset 必须足够多样化。 典型的值为 0.05。

Liae结构:当 random_warp 关闭时,inter_AB 网络不再训练以保持面部更像 src。

2021年10月09日

SAEHD:添加了 -t arhi 选项。 使脸部更像 src。
SAEHD和 AMP:

去除了周期性训练搞损失样本的隐藏函数。

修复了DirectX12版本导出.dfm的问题。

在样本生成器中,随机缩放从-0.05+0.05 增加到-0.125+0.125,提高了人脸的泛化能力。

2021年9月06日

修复保存错误
Fixed error in model saving.

AMP和SAEHD添加 blur out mask 参数
AMP, SAEHD: added option ‘blur out mask’

训练遮罩边缘的虚化
Blurs nearby area outside of applied face mask of training samples.

可使脸部附近背景变得平滑并且在交换的脸部上不太明显。
The result is the background near the face is smoothed and less noticeable on swapped face.

The exact xseg mask in src and dst faceset is required.

AMP和SAEHD:示例处理器数量不再限制为 8 个,因此如果你用的是16+核的AMD处理器,请增加虚拟内存。
AMP, SAEHD: Sample processors count are no more limited to 8, thus if you have AMD processor with 16+ cores, increase paging file size.

DirectX12 版本ml框架升级到了1.15.5。
DirectX12 build: update tensorflow-directml to 1.15.5 version.

 

2021年08月12日

遮罩模型:改进了预训练模型
XSeg model: improved pretrain option

通用遮罩:添加了更多人脸(该数据集没有公开)并使用预训练选项重新训练。现在质量更高了。
Generic XSeg: added more faces (the faceset is not publicly available) and retrained with pretrain option. The quality is now higher.

使用新的通用遮罩,更新了RTMWF数据集,添加了490张闭眼图。
Updated RTM WF Dataset with the new Generic XSeg mask applied, also added 490 faces with closed eyes.

2021年7月30日

导出AMP/SAEHD : 添加了“Export quantized” 选项(之前有用过)
Export AMP/SAEHD: added “Export quantized” option. (was enabled before)

可以使模型导出更快,如果你发现有有问题,关闭这个选项。
Makes the exported model faster. If you have problems, disable this option.

AMP模型
AMP model:

修改了CT模式的帮助信息
changed help of ct mode:

更改接近 dst 样本的 src 样本的颜色分布。 如果 src faceset 足够多样化,那么大多数情况下 lct 模式就可以了。
Change color distribution of src samples close to dst samples. If src faceset is deverse enough, then lct mode is fine in most cases.

默认的内部维度现在是1024
Default inter dims now 1024

返回 lr_dropout 选项
return lr_dropout option

最后的高损失样本行为 – 与 SAEHD 相同
last high loss samples behaviour – same as SAEHD

Xseg模型:添加预训练选项。
XSeg model: added pretrain option.

通用Xseg模型:使用预训练选项重新训练,质量更好了。
Generic XSeg: retrained with pretrain option. The quality is now higher.

使用新的通用遮罩更新了RTM WF数据集。
Updated RTM WF Dataset with the new Generic XSeg mask applied.

2021年7月17日

SAE/AMP: GAN模型恢复到了12月份的版本,至于哪一个好,可以在高清数据集上做一个测试。
SAE/AMP: GAN model is reverted to December version, which is better, tested on high-res fakes.

AMP 默认的变形因子现在是0.5
AMP: default morph factor is now 0.5

眼部和嘴部优先选项已经移除,改为默认启动
Removed eyes_mouth_prio option, enabled permanently.

遮罩训练选项已经移除,改为默认启动
Removed masked training, enabled permanently.

添加脚本
Added script

源素材和源素材的训练脚本
6) train AMP SRC-SRC.bat

AMP靠谱的训练方法
Stable approach to train AMP:

整理一个表情充分足够多样化的人脸集合
1) Get fairly diverse src faceset

设置变形因子为0.5
2) Set morph factor to 0.5

训练源源数据集到50万跌倒(越多越好)
3) train AMP SRC-SRC for 500k+ iters (more is better)

删除inter_dst的模型文件
4) delete inter_dst from model files

正常训练
5) train as usual

 

2021年7月1日

AMP模型: 修复预览历史

AMP model: fixed preview history

添加‘内部维度’选项,模型已经修改了。需要等于或者大于自动编码器维度。
added ‘Inter dimensions’ option. The model is not changed. Should be equal or more than AutoEncoder dimensions.

维度越多越好,但是需要更多显存,你可以根据你的配置来微调模型的大小
More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU.

移除预训练选项
Removed pretrain option.

默认变形因子设置为0.1
Default morph factor is now 0.1

如何去训练AMP
How to train AMP:

用常规的源素材和目标素材训练
1) Train as usual src-dst.

删除内部模型文件
2) Delete inters model files.

训练源素材和源素材,意思就是把源素材放到目标素材的文件夹
3) Train src-src. It’s mean place src aligned to data_dst

删除内部模型文件
4) Delete inters model files.

用常规的源素材和目标素材训练
5) Train as usual src-dst.

添加脚本
Added scripts

使用dfm.bat导出AMP模型
6) export AMP as dfm.bat

使用dfm.bat导出SAEHD模型
6) export SAEHD as dfm.bat

导出的dfm模型可用直播换脸软件(DeepFaceLive)
Export model as .dfm format to work in DeepFaceLive.

 

2021年5月30日

添加了实验性模型AMP(作为放大器,因为dst的面部表情放大到了src)
Added new experimental model ‘AMP’ (as amplifier, because dst facial expressions are amplified to src)

View post on imgur.com

这个模型具有变形因子,可以在合成之前设置0.0-0.1之间的值

It has controllable ‘morph factor’, you can specify the value (0.0 .. 1.0) in the console before merging process.

不同的脸部轮廓,会活的不同的下巴线
If the shapes of the faces are different, you will get different jaw line

View post on imgur.com

这就需要非常硬核的后期处理
which requires a hard post process.

但是你可以通过一个急于大佬人脸的预训练遮罩模型解决这个问题(包含在BT种子里面)。然后你可以在这个基础上继续训练dst
But 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.

在这种情况下你可以活的很好的融合效果。
In this case you will get more ‘sewed’ face.

View post on imgur.com

合成后的脸看起来不错
And merged face looks fine:

View post on imgur.com

基于大佬人脸训练的通用遮罩模型已经包含在种子文件里面。
Large dst WF faceset with applied Generic XSeg mask is now included in torrent file.

如果你的人脸集合足够多样化足够大,预训练的时候使用lct颜色转换模式比较有用。
If your src faceset is diverse and large enough, then ‘lct’ color transfer mode should be used during pretraining.

遮罩编辑器:现在删除后的人脸被移动到了_trash 目录,并且按钮移动到了窗口的右边。
XSegEditor: delete button now moves the face to _trash directory and it has been moved to the right border of the window

人脸打包工具现在会提示是否删除源文件。
Faceset packer now asks whether to delete the original files

模型自动保存时间从15分钟改成了25分钟。
Trainer now saves every 25 min instead of 15

 

2021年5月12日

数据集缩放工具已经支持改变图片的脸类型
FacesetResizer now supports changing face type

遮罩编辑工具:添加了删除按钮
XSegEditor: added delete button

遮罩训练提升
Improved training sample augmentation for XSeg trainer.

XSeg模型已经被修改,以便更好地处理大量不同的面,因此您应该重新训练现有的XSeg模型。
XSeg model has been changed to work better with large amount of various faces, thus you should retrain existing xseg model.

添加了一个基于大量人脸的预训练模型。它最适用于src faceset,因为它包含干净的面,也可以用于不是很复杂的dst镜头。
Added 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.

5.XSeg Generic) data_dst whole_face mask – apply.bat

5.XSeg Generic) data_src whole_face mask – apply.bat

2021年4月22日

添加了新的集成版DeepFaceLab_DirectX12,能在所有支持DirectX12的Win10系统上运行。
Added new build DeepFaceLab_DirectX12, works on all devices that support DirectX12 in Windows 10:

比如
AMD Radeon R5/R7/R9 2xx series or newer

比如
Intel HD Graphics 5xx or newer

比如
NVIDIA GeForce GTX 9xx series GPU or newer

在N卡上面,DirectX12版本要比NVIDIA版本慢20~80%。
DirectX12 is 20-80% slower on NVIDIA Cards comparing to ‘NVIDIA’ build.

提升了Xseg训练时样例生成效率。
Improved XSeg sample generator in the training process.

== 23.03.2021 ==

SAEHD: random_flip已经被替为新的参数
SAEHD: random_flip option is replaced with

random_src_flip (default OFF)
Random horizontal flip SRC faceset. Covers more angles, but the face maylook less naturally
random_dst_flip (default ON)
Random horizontal flip DST faceset. Makes generalization of src->dst better, if src random flip is not enabled.

添加了人脸数据集图片大小修改工具
Added faceset resize tool via

脚本名如下:
4.2) data_src util faceset resize.bat
5.2) data_dst util faceset resize.bat

把人脸数据集调整到和模型像素一样,可以减少CPU负载
Resize faceset to match model resolution to reduce CPU load during training.

别忘了备份原始数据集
Don’t forget to keep original faceset.

2021年1月4日

SAEHD: GAN升级,使得预览图减少了生硬感,变得更加干净!
SAEHD: GAN is improved. Now produces less artifacts and more cleaner preview.

GAN的选项
All GAN options:

GAN强度
GAN power

强迫神经网络学习脸部的小细节。
Forces the neural network to learn small details of the face.

当开始lr_dropout,关闭random_warp训练足够之后,启用这个参数,开了之后就别关了!
Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don’t disable.

数字越高,会越生硬。比较常用的值为0.1
The higher the value, the higher the chances of artifacts. Typical fine value is 0.1

GAN Patch大小
GAN patch size (3-640)

数字越到,质量越好,同时也需要越多的显存
The higher patch size, the higher the quality, the more VRAM is required.

即使在最低设置下,您也可以获得更锐利的边缘。
You can get sharper edges even at the lowest setting.

典型值为 8
Typical fine value is resolution / 8.

GAN 网络维度
GAN dimensions (4-64)

GAN网络的尺寸
The dimensions of the GAN network.

尺寸越高,对VRAM的要求就越高。
The higher dimensions, the more VRAM is required.

即使在最低设置下,您也可以获得更锐利的边缘。
You can get sharper edges even at the lowest setting.

典型值为 16
Typical fine value is 16.

不同设置的比较图
Comparison of different settings: https://i.imgur.com/6IgvsLN.png