To age a face in a photograph, change a hairstyle or make it smile, you need to study the semantics contained in the separate layers of the trained GAN model. Recent studies of generative adversarial networks have shown that different layers contain different semantics of synthesized images: some are responsible for color, others for textures, etc.
. , . :
StyleCLIP
Adobe , StyleGAN, .
CLIP, StyleGAN. , . , ArcFace. , : , , .
ReStyle
GAN- . , . , . , .
EigenGAN
, . - , , . , , , - , .
LatentCLR
GAN, . StyleGAN2 BigGAN.
Geometry-Free View Synthesis
, . , , , . . , .
, . , «» . , . , 3D- , .
Articulated Animation
Snap , , — , , . , , . , . .
VideoGPT
. — VQ-VAE, self-attention. GPT- .
, , .
MiVOS
. . . , .
, : , , . , , Adobe Premier.
DINO
, FAIR .
. , , .
, . , . ImageNet, — , : , , . , , .
PAWS, , .
ML :
Compositional Perturbation Autoencoder (CPA)
, . , .
FAIR . , , , ..
Transferable Visual Words
, . . . « », . .
:
That's all, thank you for your attention and see you next month!