Are there plans to implement a higher resolution face swapper?

We would love to, but the author of insightface made the decision not to release their 256 and 512 models due to ethical considerations.

How to improve the quality of swapped faces?

We worked around the limitation of the model using a GFPGAN powered a face enhancer.

Is there a chance to speed up the face enhancer?

The underlying library, FaceRestoreHelper, which is used for face landscaping and dewarping, is not thread-safe and thus becomes the limiting factor.

What operating systems are supported?

We support Linux, MacOS and Windows.

Why is my GPU not fully utilized?

VRAM is the bottleneck due the memory hungry onnxruntime.

How much VRAM should my GPU have?

We consider 4GB as a minimum, everything above 8GB should work fine.

What is a good value for the execution threads?

One thread takes around 1GB of VRAM, therefore your threads count should less or equal to that.

How to adjust the temporary directory?

On Windows modify the TEMP environment variable, while on Unix based operating systems, alter the TMPDIR instead.

The program stops after 100% processing without any error.

We don't allow deepfake porn due to ethical considerations.
Last modified 19d ago