Simply try the new amp model of deepfacelab (deepfake)

DeepFaceLab Related articles

One ：《 Brief introduction DeepFaceLab（DeepFake） Use of and things that are easy to be ignored 》
Two ：《 Keep talking DeepFaceLab（DeepFake） Evolving 2.0 edition 》
3、 ... and ：《 How to translate DeepFaceLab（DeepFake） Interactive synthesizer for 》
Four ：《 Want to improve DeepFaceLab（DeepFake） Quality precautions and skills （ One ）》
5、 ... and ：《 Want to improve DeepFaceLab（DeepFake） Quality precautions and skills （ Two ）》
6、 ... and ：《 Friendship tips DeepFaceLab（DeepFake） At present with RTX3080 and 3090 Compatibility problem 》
7、 ... and ：《 Use it efficiently DeepFaceLab（DeepFake） Some ways to improve speed and quality 》
8、 ... and ：《 Support DX12 Of DeepFaceLab（DeepFake） In addition to CUDA It can also be used. A Card 》

One , Not too new AMP Model

1.1 What is? AMP Model

stay 2021 year 5 month 30 In the Japanese version , The author tentatively added AMP Model .
The current version has 3 A model ：

• AMP
• Quick96
• SAEHD

AMP： Corresponding amplifier（ amplifier ）, Because the facial expression of the target is amplified to the source face .
This model can control morph factor（ Deformation factor ） Parameters , You can specify 0.0 To 1.0 Parameter values for .
Of course, there are many specific descriptions. Please see the author's original document .

1.2 Training AMP Steps for

1. Need a sufficiently diverse set of source face data .

2. Set up morph factor（ Deformation factor ） by 0.5

3. function 6) train AMP SRC-SRC.bat Script , Training source to source 50 More than 10000 iterations ( The more the better )

4. from model Delete... From the directory inter_dst file

5. Routine training —— Which is running 6) train AMP.bat

1.3 Final effect

The content of the test has been deleted , I found that I forgot the screenshot …… Only words can give an example .

AMP The biggest feature of the model is to specify morph factor Deformation factor .

• morph factor = 0.0： After synthesis, it is completely the target face , Equal to no face change at all .
• morph factor = 1.0： After synthesis, it is completely the source face .

So when morph factor Be situated between 0.0 To 1.0 Between time , The synthetic effect is More like source / More like a goal Between the floating .
Try it. It's true , But contrast SAEHD I don't seem to find AMP The advantages of , What the hell is it for .

I doubt , Is it a choice when it is difficult for the source to reproduce all the expressions of the target ？
If you know , Please grant me your advice .
.

Two , Other models and features

2.1 SAEHD

Normal model used at ordinary times , There are many parameter choices .
You can adjust the resolution and learn various options of the network according to your own graphics card .
abbreviation SAEHD：High Definition Styled AutoEncoder
This model has been used in previous articles （ Nothing else ）.
.

2.2 Quick96

The fastest model , There are no options , Fix 96 The resolution of the , The whole face .
Because the options are fixed , So it will automatically use the pre training data provided by the author （ So soon ）.
It can be used on low-end graphics cards , such as 2GB+ Of NVidia, perhaps 4GB+ Of AMD..

With this model, you can try it quickly , It works .
The disadvantage is that the resolution is too low , Quantitative change causes qualitative change , Sometimes it leads to poor final effect （ Not just resolution ）.
.

2.3 Removed model

• TrueFace： Can only N Card use , The author tried and removed .
• SAE： By SAEHD Replaced by .
• avatar ： Is it for animation , Forget the …… It was removed anyway , Its effect can be replaced .
• RecycleGAN： Uh , No impression at all ……