Using vggface2 based embedding in inswapper is not prducing correct results

[prashant-saxena]

Hi,

I'm trying to use vggface2 based embedding using facenet-pytorch Here is the code to generate:

import torch
from facenet_pytorch import MTCNN, InceptionResnetV1

class FaceEmbed:
    '''https://github.com/timesler/facenet-pytorch'''
    def __init__(self):
        
        device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

        # If required, create a face detection pipeline using MTCNN:
        self.mtcnn = MTCNN(device=device)

        # Create an inception resnet (in eval mode):
        self.resnet = InceptionResnetV1(pretrained='vggface2').eval()

    def get(self, img):
        # Get cropped and prewhitened image tensor
        img_cropped = self.mtcnn(img)

        # Calculate embedding (unsqueeze to add batch dimension)
        img_embedding = self.resnet(img_cropped.unsqueeze(0))

        return img_embedding.detach().numpy().flatten()

Then using a custom inswapper, I have replaced this line https://github.com/deepinsight/insightface/blob/da738727d5fe26d305499c311bd05cb51c3936d5/python-package/insightface/model_zoo/inswapper.py#L50 to

latent = embed.get(source_image).reshape((1,-1))

The resulting swap (Incorrect in picture) is not at all good. Any pointers?