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).

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

• https://docs.docker.com/install/linux/docker-ce/ubuntu/
• https://docs.docker.com/compose/install/
• git clone https://github.com/EvgenyKashin/ganarts.git
• create .env with s3 access keys
• sudo docker-compose up -d