-
Notifications
You must be signed in to change notification settings - Fork 2.7k
Techniques For Increasing Image Quality Without Buying a Better GPU
Single-Image Super-Resolution for Anime-Style Art using Deep Convolutional Neural Networks. And it supports photo.
Linux: https://github.com/nagadomi/waifu2x
Site (Has upload size limit): http://waifu2x.udp.jp/
Windows (Use Chrome's translate feature): http://inatsuka.com/extra/koroshell/
Waifu2x was designed to increase the quality of anime images and to re-size images. Thus depending on the style the re-sizing and/or noise reducing will work to varying degrees.
Using NIN to increase Places205-VGG image size and quality:
-
First, create an image with neural-style using the Places205-VGG model.
-
Depending on the GPU resources available to you, either convert the previously created image to a .jpg or leave it as a png.
-
Run the following command using the same style image that you originally used to create your content image. Make sure your content image is the one you created in step 1:
th neural_style.lua -style_image StyleImage.jpg -content_image ContentImage.jpg -output_image out.png -tv_weight 0.0001 -image_size 2500 -save_iter 50 -content_weight 10 -style_weight 1000 -num_iterations 1000 -model_file models/nin_imagenet_conv.caffemodel -proto_file models/train_val.prototxt -content_layers relu1,relu7,relu12 -style_layers relu1,relu3,relu5,relu7,relu9 -backend cudnn -cudnn_autotune -optimizer adam
Examples/Results:
Tubingen: https://imgur.com/a/ALzL7
Brad Pitt: https://imgur.com/a/Ws8x5
Escher Sphere: https://imgur.com/a/KS1mk
Notes:
This has only been tested with Starry Night and the example images. I used an Amazon g2.2xlarge spot instance which had 4GB of GPU memory.
I used the following command to generate the original images:
th neural_style.lua -style_image StyleImage.jpg -content_image ContentImage.jpg -output_image out.png -tv_weight 0.0001 -save_iter 50 -num_iterations 1000 -model_file models/snapshot_iter_765280.caffemodel -proto_file models/deploy_10.prototxt -backend cudnn -cudnn_autotune -optimizer adam
Download the Places205-VGG model here: http://places.csail.mit.edu/model/places205vgg.tar.gz