Failed to face swap with celeb_hq ?

what caused this？

It seems that the third and fourth faces are coped from the second face. There are two possible reason I could tell:
(1) Faces are not processed by the model into latent codes, or
(2) Mask is not processed correctly so that the generated swapped face cannot be put on the second face.

For both cases, you could check the middle results in the inference code (comments are marked in #):

One-Shot-Face-Swapping-on-Megapixels/inference/inference.py

Line 119 in 1af90db

def run(self, src_idx, tgt_idx, refine=True):

    def run(self, src_idx, tgt_idx, refine=True):
        src_face_rgb, tgt_face_rgb, tgt_mask = self.read_pair(src_idx, tgt_idx)
        source, target = self.preprocess(src_face_rgb, tgt_face_rgb)
        swapped_face = self.swap(source, target)
        swapped_face = self.postprocess(swapped_face, tgt_face_rgb, tgt_mask) # comment out this line to see if swapped face is generated

        result = np.hstack((src_face_rgb[:,:,::-1], tgt_face_rgb[:,:,::-1], swapped_face))

        if refine:
            swapped_tensor, _ = self.preprocess(swapped_face[:,:,::-1], swapped_face)
            refined_face = self.refine(swapped_tensor)
            refined_face = self.postprocess(refined_face, tgt_face_rgb, tgt_mask)
            result = np.hstack((result, swapped_face))
        cv2.imwrite("{}.jpg".format(self.swap_type), result)

    def swap(self, source, target):
        with torch.no_grad():
            ts = torch.cat([target, source], dim=0).cuda()
            lats, struct = self.encoder(ts)

            idd_lats = lats[1:]
            att_lats = lats[0].unsqueeze_(0)
            att_struct = struct[0].unsqueeze_(0)

            swapped_lats = self.swapper(idd_lats, att_lats) # use "swapped_lats = att_lats " to see if the face can be reconsructed
            fake_swap, _ = self.generator(att_struct, [swapped_lats, None], randomize_noise=False)

            fake_swap_max = torch.max(fake_swap)
            fake_swap_min = torch.min(fake_swap)
            denormed_fake_swap = (fake_swap[0] - fake_swap_min) / (fake_swap_max - fake_swap_min) * 255.0
            fake_swap_numpy = denormed_fake_swap.permute((1, 2, 0)).cpu().numpy()
        return fake_swap_numpy

where the swapped_lats is the latent code for swapped face and tgt_mask is the mask for pasting the swapped face.

thanks for your answer

I read the inference.py code and generated latent code, but I found the code in FaceTransferModule.py, The scale value is large, is there a problem with the weight。scale = torch.sigmoid(self.weight).expand(N, -1, -1)

def forward(self, idd, att):
        if self.type == "ftm":
            att_low = att[:, :self.swap_indice]
            idd_high = idd[:, self.swap_indice:]
            att_high = att[:, self.swap_indice:]

            N = idd.size(0)
            idds = []
            atts = []
            for i in range(self.num_latents):
                new_idd, new_att = self.blocks[i](idd_high[:, i], att_high[:, i])
                idds.append(new_idd)
                atts.append(new_att)
            idds = torch.cat(idds, 1)
            atts = torch.cat(atts, 1)
            scale = torch.sigmoid(self.weight).expand(N, -1, -1)
            print(f'scale: {scale}')
            latents = scale * idds + (1-scale) * atts

            return torch.cat([att_low, latents], 1)

I detected the mask and it is correct (the mask is generated using: https://github.com/zllrunning/face-parsing.PyTorch)

Can you give me a contact?

Using issue to communicate is somewhat inefficient。

(1) It's fine that scale is large, as it is learnt to highlight idd instead of att (You can change the scale for generating more samlpes that keep more att if you want, but we set it as a constant in this repo).
(2) When you use the mask generated from https://github.com/zllrunning/face-parsing.PyTorch, please note that the mask used in our model is not colored. The mask should be [0,1,2,...,18] encoded, where we recommend you use the official masks provided by CelebA-HQ-Mask.
(3) If you have more questions, you may want to give me your e-mail so that I can send you my contact (I don't want my contact to be public ;-) )

(1) It's fine that scale is large, as it is learnt to highlight idd instead of att (You can change the scale for generating more samlpes that keep more att if you want, but we set it as a constant in this repo). (2) When you use the mask generated from https://github.com/zllrunning/face-parsing.PyTorch, please note that the mask used in our model is not colored. The mask should be [0,1,2,...,18] encoded, where we recommend you use the official masks provided by CelebA-HQ-Mask. (3) If you have more questions, you may want to give me your e-mail so that I can send you my contact (I don't want my contact to be public ;-) )

my email: 2056374813@qq.com

(1) It's fine that scale is large, as it is learnt to highlight idd instead of att (You can change the scale for generating more samlpes that keep more att if you want, but we set it as a constant in this repo). (2) When you use the mask generated from https://github.com/zllrunning/face-parsing.PyTorch, please note that the mask used in our model is not colored. The mask should be [0,1,2,...,18] encoded, where we recommend you use the official masks provided by CelebA-HQ-Mask. (3) If you have more questions, you may want to give me your e-mail so that I can send you my contact (I don't want my contact to be public ;-) )

The mask that i use are [0,1,2,...,18] .

sent