Fancy homepage made by Samyak
Album Selection Screen
A style transfer. Each image represents a different weight in the strength of the style transfer when feeding forward through the Conv Net
Nick + Radiohead's OK Computer
Apoor + Pink Floyd's Dark Side of The Moon
Samyak + Radiohead's Hail to the Thief
Want your next Profile Photo to be cool like The Rolling Stones? Maybe you can impress your next Tinder match with a photo that pays tribute to Nirvana?
You can now do these things and more with Albumify.me! We use deep learning to perform style transfer of famous album covers to any photo of your choice!
We have created this app by incorporating the core idea of famous paper Gatys et al. that CNNs pre-trained for image classification already know how to encode perceptual and semantic information about images.
When we train a Convolutional Neural Network higher layers in the network capture the high-level content in terms of objects and their arrangement in the content image but do not constrain the exact pixel values of the reconstruction. To obtain a representation of the style of an input image, we employ correlations between the different filter responses over the spatial extent of the feature maps.
Basically style and content in a CNN are separable. We are taking content from one image and style from another and we are combining them.
We're a group of developers from across the world who grouped together to practice our Machine Learning skills. Machine Learning has a breadth of applications but we decided to learn how to do Image Style Transfer after reading articles about applications that transfer the styles of famous works of art to your photos. An excellent service that demonstrates this functionality is Deepart.io. We worked together to build our own deep style transfer program that transfers album styles to any image.
Computational Resources - Testing our program was very computationally expensive. It would take around 10 minutes to perform a single style transfer locally. Performing a style transfer when we deployed to google cloud platform would run through the $300 worth of credits we received from Google within a few iterations.
- We plan on finding a way to deploy this algorithm that will not exhaust our bank accounts :)
- Optimization of the algorithm to increase performance on higher resolution images and perform more accurate style transfers
Please download the script from github and give it a try! Thanks to the MHacks Nano team for (virtually) hosting the event! Hope to see some of you guys soon at MHacks 10
Our icon was borrowed from the great open source project Etcher