Run on Mobile?

[InternetMaster1]

Amazing Library!

Can this run on a mobile device, say Android?

[divyanshu092]

Hi,
Yes, this project is based on Google Colab so it can be run on any device.
You should first mount your google drive, then clone this repo and go to the project's Code folder and run the Photorealistic.ipynb. You can follow this article for cloning the repo on your google drive: https://medium.com/@ashwindesilva/how-to-use-google-colaboratory-to-clone-a-github-repository-e07cf8d3d22b

[InternetMaster1]

Thanks for the quick reply.

1. I meant, can I run on Android on a Mobile GPU similar to this library : https://www.tensorflow.org/lite/models/style_transfer/overview
   I would have to convert the model files to tflite?

2. What is the execution time for the style transfer and the size of the quantized mode? If you check the above link, there is a table at the bottom which gives the performance benchmarks for Google's implementation.
   

How does photorealistic style transfer compare?

3. The Google Style Transfer library's output creates a lot of distortions. If I take a human face as a content image and apply style transfer, it shows a lot of wrinkles on the face. Your library would give better results?

[divyanshu092]

Hello,
I don't have much idea regarding how this would run on Mobile GPU, but with the advent of Pytorch Mobile, it might be possible to deploy any Pytorch model to both Android and IOS.
The comparison would make sense when we actually deploy on mobile GPU. As of now, I am using Google Colab and the algorithm is an iterative one which improves over time. However, good results take time in the order of minutes.

If I take a human face as content image and apply photorealistic style transfer, the wrinkles/distortions highly depend on the style to which we are wishing to transfer. I would attach a few examples, so you can compare.

[divyanshu092]

It also requires some changes in the hyper-parameters while testing on human faces giving more weight to content.
I am attaching a content image, some style images and the photorealistic style transfer result images.