search

NVlabs / SPADE

Semantic Image Synthesis with SPADE
https://nvlabs.github.io/SPADE/
Other
7.61k stars 980 forks source link

about the usecase of the project. #20

Open imnaren142 opened 5 years ago

imnaren142 commented 5 years ago

Hi, can we generate the images of fake humans using this code?

ArashHosseini commented 5 years ago

+1

taesungp commented 5 years ago

COCO-stuff does contain human label, but it is actually a very hard problem to generate realistic human on datasets of human in various poses. The generated human images on COCO-stuff look pretty bad. The current best GAN models can only generate realistic human when all the keypoints / part segmentations are given , or when the human images are aligned. I think our model can be used to generate realistic human images if it is trained with part segmentation, but there is no pretrained model that does that.

banyet1 commented 5 years ago

I've tried pix2pixHD(label and image) and other GAN to generate fake humans, the result is weird, most of them looks like just a mess of meats. Now I'm using stylegan (https://github.com/NVlabs/stylegan), maybe it's the light of the tunnel, but it's take a lot of time for training (2x1080ti, 10K images, 10 day's passed, and it still wade on and on, and on).

imnaren142 commented 5 years ago

@banyet1 can you post some of the example images which were generated using pix2pixHD?

mod-cpu commented 5 years ago

@naren142 here are a few I made w/ Pix2PixHD: http://www.mattkenney.me/fake-news-with-openpose-and-pix2pix/ http://www.mattkenney.me/openpose-pix2pix/

taki0112 commented 5 years ago

If you use the CelebA-HQ Mask dataset. SPADE can also generate fake humans.

Here is my tensorflow version result