neverUseThisName

Maybe you can apply PCA to 512-d vectors and reduce them to 128-d.

> Maybe you can download here and replace the two git-lfs pointer file in dex/dex/pth. > [gender_sd.pth](https://drive.google.com/file/d/1ancMcq5b_0nppdIAav1UgBxdiaMnnQPe/view) > [age_sd.pth](https://drive.google.com/file/d/1KtzJBcGPMwjYdxbJkfEeRFzxNMaTVykH/view) how to modify .pth file after downloading the models?

You need to have a dataset of face images and provide its path with `-sfp path/to/dataset` when running `faceBlending.py`. All arguments can be found in `get_args` function or by `python...

I don't think it has anything serious to do with dataset, but I'm not sure. Would you mind providing some samples of bad results? There are also some parameters you...

The problem might be that color_transfer estimates the color style of the whole image. If there are distinct components (e.g. foreground and background) as is your case, the resultant style...

It is in the pypy: https://pypi.org/project/color_transfer/ Just 'pip install color_transfer' will do.

No, I did not dive deep into this.

Just extract all faces and not swap faces from the same person.

Faces in my dataset are detected and cropped by MTCNN, but not aligned. It turned out if the dataset is large enough, it's easy to find at least one image...

No, I only used cosface.