Deep image analogy style transfer, Feedback required

I am nearly done implementing this paper :


Its on pytorch, and I am getting these results :

As you can see, this paper produces amazing results(even though mine arent like the paper, its still pretty good)
I really want to make this paper very popular, write a couple blog posts about it. I have some new ideas that may improve the speed of this.

All this requires my results to closely match the paper’s results, which are:
40 PM
01 PM

If anybody is interested to work with me, I will give them access to my GitHub repo. I have been at it for a few days now, and im not sure whats causing the choppy results. Perhaps a new perspective will fix the issues.

EDIT:

This is what the current state is:

you can find the project at : https://github.com/harveyslash/Deep-Image-Analogy-PyTorch

It seems like you are really close on this. I would be willing to look at what you currently have and see if there is anything I could do to help. This is a pretty interesting idea. Do you see any real-world use-cases for this?

Hey thanks so much for your time.
I plan to make an app out of this that can generate high quality styles (far better than prisma) from selfies , pics etc.

since today is halloween , heres a fun image i made just now:

If you want to collaborate, send me your GitHub username , ill add you as a collaborator. If you want I can also allow you access to my server ( so that you can directly see the code in action)

@harveyslash I’d love to join your effort on this, Thanks

Summary

This text will be hidden

I sent an invite on github

after some minor tweaking this is what I am getting :
Unknown

1 Like

That looks cool. I cloned the code yesterday, but did not get the chance to run it due to my aws gpu instance limits

it turns out this is very very sensitive to neuron activation clip value : getting good results after trying several options :

2 Likes

Nice work @harveyslash! Will wait for your blog post.

I think ill freeze my code. Getting comparable results now

I have published the code.

If anybody requires explanation, or faces bugs, or in general wants to help improve the project, feel free.
respond here or in GitHub itself.

Blog post with detailed explanation coming soon.

3 Likes

A few images generated :

Sounds very interesting…I would like to get a chance to collaborate.

project is online, you can take a look and open issues.

You can also open an issue if you don’t understand a specific thing (i.e. doesnt have to be a bug)