Skip to content

Latest commit

 

History

History
32 lines (26 loc) · 1.3 KB

README.md

File metadata and controls

32 lines (26 loc) · 1.3 KB

GanArts

This T-shirt does not exist.

Images are generated with StyleGAN neural network, trained on 40,000 images of modern art. The training process takes 3 days on 4x1080ti GPU. Every 30 seconds there are 9 new images, and the old ones are deleted. To print these images, use specialized sites(place for your ad). To increase the resolution, you can use a super resolution (increase the image size from 1024px to 2048px).

Demo video: https://youtu.be/PQZX4r4ib9Y

Repo consist of generator(based on StyleGAN) and server(flask + gunicorn + redis + s3).

Deploy options

  • cpu: layers converting is needed(NCHW -> NHWC)
  • gpu: high renting cost
  • s3(current): generating on gpu server, save to s3, serve on cpu server

Generator generates images, makes preprocessing and uploads to s3. Backend worker downloads images and saves to redis each n seconds in one transaction. Main server gets 9 images and urls of full-size images(on s3) from redis and serves these to clients.

EC2