Unable to reproduce quality

[SpaceCowboy850]

Hello,

I was interested in checking this out as it proposes to preserve larger features, create larger textures, and generate faster than Gatys. But outside your model that you've uploaded, I cannot produce good results on textures that I give it myself.

This is the source texture I used, a 512x512

After over an hour and half of training (on a Titan X Pascal), this is the snapshot:

It doesn't seem like the default parameterization do very well. We've trained for over 20 hours and it pretty much stays at this level of quality.

[nikjetchev]

Hi

thanks a lot for the comment, I may take a look and see what settings work best for that texture.
It is indeed sometimes trial-and-error to choose the right values for patch size, image size of the texture, and network depth (which determines the receptive field)

As a sidenote, please take also a look at https://github.com/ubergmann/psgan , our latest algorithm can handle much more diverse texture images.

[SpaceCowboy850]

That's great to hear! We've implemented gatys on C++ windows caffe but are now looking to get the speed and texture size up, so this looks promising if we can get it to work.

[nikjetchev]

Some of the artifacts of the images may be a general issue with deconv. filter implementations, see
http://distill.pub/2016/deconv-checkerboard/

I changed a bit the parameters and tried with your texture

nz = 20 # num of dim for Z at each field position
zx = 7 # number of spatial dimensions in Z
batch_size = 20
epoch_iters = batch_size * 500

also used 4 mirrored versions of your texture for data augmentation - since it is a bit small for a good texture

This is what I got, looks a bit better than yours ?

[SpaceCowboy850]

Ah, okay, that's good to know. I'll play around with it some more. I'm aware of the checkerboard link, so I'll take a look at that after we've played around with parameterization a bit. Thank you!