Imagine a custom deepfake!

Preparation - Now that you have on your face, you have set up your model, it's time things start!

Go to the tab process in the graphical user interface.

Here we will tell you where it is stored Faceswap all, we want to use, and start training - after reading the tutorial on https://deepfake-porn.com/howto-custom-deepfake/.

Face - This is where we'll tell Faceswap where the face is stored, and the location of each file in their lineups (if necessary).

An entry - that is the location of the folder that contains the "A" side is extracted as part of the extraction process. These are the faces that are removed from the scene of origin is replaced by an exchange of face. There must be some 1000-10000 faces in this folder.

List A - If you are training with a mask, or by using the "Landmark deformation" option, then an alignment file is needed for your face. It was produced as part of the extraction process. If the file exists on the face and a folder named alignments.json it will automatically be included in the process. Each face in front of the file must have an entry in the file alignments if training fails. You may need to combine multiple alignment files. You can find more information on configuring file alignments for training in the free sample.

Input B - which is the location of the folder containing the face "B" is extracted as part of the extraction process. These are the faces that switch on stage. There must be some 1000-10000 faces in this folder.

Lists B - If you train with a mask, or by using the "Landmark deformation" option, then you will have an alignment file for your face. It was produced as part of the extraction process. If the file exists on the face and a folder named alignments.json it will automatically be included in the process. Each face in front of the folder B must have an entry in the file alignments if training fail. You may need to combine multiple alignment files. You can find more information on configuring file alignments for training in the free sample.

Models - Options related training model:

Dir model - this is the model file will be saved. You must select an empty folder if a new model starts, existing or folder that contains the template file if it continues with the formation of the model start.

Trainer - This is the standard form for exchange. An overview of the different models available before.

Allows growth - [NVIDIA SOLO]. `Allow_growth` TensorFlow enables GPU configuration options. This option prevents TensorFlow allocate all of the GPU VRAM at launch, but it can lead to fragmentation of VRAM and slow performance. This should only be allowed if you have a problem with training (in particular, a cuDNN error is obtained).

Training - Training special arrangements:

Batch size - As described above, the image number of the batch size provided by a model as well. Increasing this number using VRAM will increase. Increasing the size of the lot is the formation of acceleration at some point.

small lots in a usual way that helps regulate the diffusion of the model. While large batches form more rapidly, the size of the batch in the range of 8 to 16 is likely to produce better quality. Yet it is whether other forms of regulation can replace or remove this requirement.

Iteration - The number of iterations to perform before an automatic stop training. It's there for automation, or the formation of the license after a while. Typically, training is stopped manually when you are satisfied with the quality of the preview.

GPU - [NVIDIA SOLO] - The number of train GPU. If you have multiple graphics processors installed in the system, you can use up to 8 of them to accelerate the formation. Note that in this way until the speed is not linear, and add graphics processors, lower yields become operational.

Ultimately, it can be a much more the size of multiple graphics processors cutting. You will always be a bottleneck for the speed of the GPU and VRAM its weakest, so it works best in training when the GPU is identical. You can learn more about the multi-GPU hardware here.

No newspaper - loss models and logging are provided to analyze data in TensorBoard and GUI. This means you mortals do not have access to this data. Realistically, there is no reason to disable logging generally not be verified.

Landmark chain - As mentioned before data is curved so that the NN can learn to create a face. reference channel is a process different deformation, which attempts to address the similar face chain random on the other side (for example, a number, has found some of the same faces of the set B, and applies the chain with some randomization). The jury is whether this offers benefits/differences of standard deformation to chance.

No Flip - Images are inclined to chance to increase the amount of data that NN will. In most cases, this is fine, but the face is not symmetrical so that for multiple purposes, this may not be desirable (eg mol to one side of the face). Overall, this should be allowed and should be left when the training starts. Later in the session, you can disable this for a trade.

Color does not increase - To increase Faceswap color (detailed above). Helps color matching/lighting/contrast between A and B, but sometimes it may be desirable, it can be deactivated here. Color increases impact can be seen below.

VRAM savings - Setting to save VRAM optimization.

Faceswap offers many optimizations that can save VRAM allow users to train the model otherwise be not be able to train. Unfortunately, this option is available for users of Nvidia. This should be the last port of call. If you can train a batch size of at least 6-8 without enabling this option, then you should do first, because everything comes with a speed penalty. All of these options can be turned on each other, to save battery.

Gradiente memory storage - [NVIDIA only] - MSG is an optimization method that saves VRAM with the cost of computing. In the best-case situation can split your needs VRAM 20% increase in training time. He is the first choice you have to try. You can learn more about saving memory gradient here.

Savings Optimizer - [NVIDIA only] - This can save a significant amount of VRAM to perform calculations CPU rather than GPU optimization. It comes with a cost of increased use of the system RAM and the speed of the slowest station. It should be the second option you try.

Table tennis - [NVIDIA only] - Alias ​​"The last resort". This is by far the worst of the VRAM saving option, but it may be enough to get what you need. It essentially divides the two models and forms the first half of the model at a time. This saves up to 40% of VRAM but will be more than twice the amount of time to form the pattern. This should be your last option tried.

NB logging TensorBoard / graphics are not available with this option.

Note: No preview as the training cycle will show took place on both sides of the model.