czczup / urst

[AAAI 2022] Towards Ultra-Resolution Neural Style Transfer via Thumbnail Instance Normalization

urst's Issues

stroke size controlling

In your paper you mention changing the decoder of adain to change the stroke size.
what is the difference between the decoders e.g. "decoder_stroke_perceptual_loss_1.pth" ?
In my case, I would like to convert an ultra high resolution (8192x8192), but applying the same "huge" stroke size, as I would if I would resize the picture to 1024x1024

Possible to use TIN in Aesthetic UST method?

Hi @czczup! Thanks for your awesome work!

Since you've shown that many style transfer methods are compatible with TIN, I wonder if it's possible to use TIN in Aesthetic UST?

Following figure overviews how AesUST work.

They introduced AesSA module which confuses me where to put the TIN module. Could you please help me out?

Best,
Rahul Bhalley

possible to change from adam optimizer to L-BFGS?

possible to change from adam optimizer to L-BFGS?

not particularly excellent at any of this (code etc.) - so even just a simple, yes or no, will suffice, even as to prevent me from wasting hours attempting to cut and paste stuff that will never work?

thanks again

Google Colab

Hello!

Any chance you can make a google colab version? I am having trouble setting this up.

_pickle.UnpicklingError: invalid load key, '\xff'.

no problems, fixed. - thank you so very much for this, this is truly excellent, what a package!

Wrong instructions in Readme

I want to use Wang2020 Colaborative for style transfer. The readme says to use test.py and config, style arguments. There is no test.py fail and main.py (which I think you maean) lacks these attributes. Could you please update the instructions in the readme so that one can use the code easily?

None

Multi model style transfer consumes huge memory

Hi @czczup,

Thanks for implementing Multi Model Style Transfer. But it consumes very huge memory and then the program gets killed (for default patch and thumbnail sizes). I'm using --URST flag for these experiments.

This screenshot is for:

• patch size: 500
• thumb size: 512

This process gets killed (the default configuration):

TIN application

Hi,

Sorry to bother again - lets say; roughly, forgive my expression;

using cysmith neural transfer (LGBFS optimizer), I process 4 x (500 x 500) slices of a lets say 2000 x 2000 pixel image-A, and I process one total picture thumbnail of 500 x 500 pixel of image-A (i suppose, downsampled from the original 2000 x 2000 image-A)

now, using your TIN method;

would I be able to 'stitch' these back together and then get rid of the 'stitch' area problems?

is it possible to string together or isolate some code as to achieve something like this?

thanks again for your time and consideration

• apologies if ive missed or misunderstood something from your method / paper article

Li2017 Universal - unhashable type list

(Wang) C:\Users\Bob\URST\Li2017Universal>python test.py --content c:\users\bob\desktop\ds.jpg --style c:\users\bob\desktop\dsstyle.jpg --URST
Traceback (most recent call last):
File "test.py", line 166, in
wct = WCT(args).to(device)
File "C:\Users\Bob\URST\Li2017Universal\util.py", line 16, in init
vgg1 = torchfile.load(args.vgg1)
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 424, in load
return reader.read_obj()
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 370, in read_obj
obj._obj = self.read_obj()
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 385, in read_obj
k = self.read_obj()
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 386, in read_obj
v = self.read_obj()
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 370, in read_obj
obj._obj = self.read_obj()
File "C:\Users\Bob\anaconda3\envs\Wang\lib\site-packages\torchfile.py", line 387, in read_obj
obj[k] = v
TypeError: unhashable type: 'list'