Awesome
SimSwap++: Towards Faster and High-Quality Identity Swapping
Accepted by TPAMI (IEEE Transactions on Pattern Analysis and Machine Intelligence)
Xuanhong Chen, Bingbing Ni $\dagger$, Yutian Liu, Naiyuan Liu, Linzhi Zeng, Hang Wang
$\dagger$ Corresponding author
Project page of SimSwap++
<img width=10% src="./docs/img/attention.gif"/> Please note that the current respository is a supplementary material for the SimSwap++ paper, NO issues will be replied at this stage. Once everything is ready we will release training and testing codes like our SimSwap.
Methodology of SimSwap++
Additional Results:
[Source ID: Angelina Jolie Target ID: Gamora]
<img src="./docs/samples/id_angelina-jolie_attr_BG_TOYART_GAMORA.png"/>[Source ID: Tom Cruise Target ID: Gamora]
<img src="./docs/samples/id_tom_cruise_genes--attr_BG_TOYART_GAMORA.png"/>[Source ID: Brad Pitt Target ID: Gamora]
<img src="./docs/samples/id_BradPitt--attr_BG_TOYART_GAMORA.png"/>[Source ID: Will Smith Target ID: Gamora]
<img src="./docs/samples/id_will_smith_attr_BG_TOYART_GAMORA.png"/>[Source ID: Geoffrey Hinton Target ID: Gamora]
<img src="./docs/samples/id_hinton--attr_BG_TOYART_GAMORA.png"/>[Source ID: Captain America Chris Evans Target ID: Gamora]
<img src="./docs/samples/id_ChrisEvans--attr_BG_TOYART_GAMORA.png"/>[Source ID: Jim Carrey Target ID: Gamora]
<img src="./docs/samples/id_JimCarrey--attr_BG_TOYART_GAMORA.png"/>[Source ID: Matthew McConaughey Target ID: Gamora]
<img src="./docs/samples/id_MatthewMcConaughey1--attr_BG_TOYART_GAMORA.png"/>[Source ID: Jackie Chan Target ID: Gamora]
<img src="./docs/samples/id_Jackie-Chan-1200by667--attr_BG_TOYART_GAMORA.png"/>[Source ID: Messi Target ID: Gamora]
<img src="./docs/samples/id_messi2--attr_BG_TOYART_GAMORA.png"/>[Source ID: Kylian Mbappé Target ID: Gamora]
<img src="./docs/samples/id_mbp--attr_BG_TOYART_GAMORA.png"/>[Source ID: Cristiano Ronaldo Target ID: Gamora]
<img src="./docs/samples/id_Cristiano_Ronaldo--attr_BG_TOYART_GAMORA.png"/>[Source ID: Quentin Tarantino Target ID: Gamora]
<img src="./docs/samples/id_Quentin-Tarantino-2020--attr_BG_TOYART_GAMORA.png"/>Video in-the-wild Results (under construction):
Videos are generated frame by frame without any temporal smoothing (e.g., Kalman Filter), which may cause some subtle flickering.
Group1 [SimSwap++ (S)]:
Source ID: Scarlett Johansson Target ID: Shakira (1080p on YouTube)
Source ID: Dilireba Target ID: Shakira (1080p on Google Drive)
Source ID: Tom Cruise Target ID: Shakira (1080p on Google Drive)
Source ID: Elon Musk Target ID: Shakira (1080p on Google Drive)
Musk's face is always so recognizable!
Source ID: Keira Knightley Target ID: Shakira (1080p on Google Drive)
Source ID: Brad Pitt Target ID: Shakira (1080p on Google Drive)
Source ID: Nicole Kidman Target ID: Shakira (1080p on Google Drive)
Group2 [SimSwap++ (S)]:
Source ID: Kelly Clarkson Target ID: Taylor Swift (1080p on YouTube)
<img src="./docs/video/1.webp"/>Source ID: Geoffrey Hinton Target ID: Taylor Swift (1080p on YouTube)
Source ID: Gal Gadot Target ID: Taylor Swift (1080p on YouTube)
Source ID: Leonardo DiCaprio Target ID: Taylor Swift (1080p on YouTube)
<img src="./docs/video/4.webp"/>Source ID: Elon Musk Target ID: Taylor Swift (1080p on YouTube)
Source ID: Robert Downey Target ID: Taylor Swift (1080p on YouTube)
Source ID: Aamir Khan Target ID: Taylor Swift (1080p on YouTube)
Group3 [SimSwap++ (S)] Complex Movie Scene:
I am Iron Man
Few previous face-swapping papers have published such complex movie scenes
The subtle flickering can be resolved by introducing temporal smoothing, which is not our concern
Source ID: Captain America Chris Evans Target ID: Iron Man (1080p on YouTube)
Source ID: Scarlett Johansson Target ID: Iron Man (1080p on YouTube)
Source ID: Elon Musk Target ID: Iron Man (1080p on Google Drive)
Source ID: Leonardo Target ID: Iron Man (1080p on YouTube)
Source ID: Gal Gadot Target ID: Iron Man (1080p on YouTube)
VGGFace2-HQ Dataset
VGGFace2-HQ contains more than $1.36M$ $512 \times 512$ aligned face images and up to $9, 630$ distinct identities. In addition, this dataset consists of two parts:
- (1) a natural image sub-collection, which collects up to $200, 000$ images covering $1, 000$ different identities;
- (2) a synthetic image sub-collection, containing $8, 630$ cleaned and re-annotated identities (i.e., clean up the images with mismatching identities and low-quality faces in the cropped VGGFace2).
Download the dataset:
<!-- ***Limited by the capacity of the cloud disk, we divided the dataset into two parts*** -->Via Google Drive:
<!-- [[Google Drive] VGGFace2-HQ Part2 (89GB)](https://drive.google.com/drive/folders/1ZHy7jrd6cGb2lUa4qYugXe41G_Ef9Ibw?usp=sharing) -->We are especially grateful to Kairui Feng PhD student from Princeton University.
Via Baidu Drive:
[Baidu Drive] VGGFace2-HQ Password: sjtu
Samples from VGGFace2-HQ
Natural Part:
Synthetic Part:
