wty-ustc / e2style

[TIP 2022] E2Style: Improve the Efficiency and Effectiveness of StyleGAN Inversion

150 stars 10 forks source link

still bad inversion but i dont know why #22

Open adeelnazar opened 1 year ago

adeelnazar commented 1 year ago

still bad inversion but i dont know why

celebrity-pics-3 (1) (1) 320937818_909805843710812_7225152858816556845_n (1) (1) 320925575_562852398514446_8652596532889929974_n (1) (1) 320505752_520230476748434_7768053947597212110_n (1) (1) 320345500_931916501076882_7892912347158999581_n (2) (1) 4245574-tom-cruise (1) (1)

wty-ustc commented 1 year ago

Use this script to align the images first.

adeelnazar commented 1 year ago

Use this script to align the images first.

Pic of Angelina jolie and Tom cruise both pic are correct align but not good as result

wty-ustc commented 1 year ago

Can you send me an original image? I'll test it when I have a spare moment.

adeelnazar commented 1 year ago

Can you send me an original image? I'll test it when I have a spare moment.

Give me Whatsapp no or email??

wty-ustc commented 1 year ago

You can comment the image directly here.

adeelnazar commented 1 year ago

Y

ou can comment the image directly here.

adeelnazar commented 1 year ago

E2 Style I think it's best for inversion compare to others but for simple images but image with high make up and charming face are not good

wty-ustc commented 1 year ago

I agree. Our target inversion space is the W+ space of StyleGAN. Compared with peer work, our method is more superior. But for these high frequency features, inversion to the W+ space is not sufficient. See #15 for the reason and the solution.

adeelnazar commented 1 year ago

I agree. Our target inversion space is the W+ space of StyleGAN. Compared with peer work, our method is more superior. But for these high frequency features, inversion to the W+ space is not sufficient. See #15 for the reason and the solution.

Is it possible to inversion image with HFGI inversion and style mixing with e2style with latentcode from HFGI