Neural-style: Best settings to look like a Deepart.io output

[joejacksonnite]

Hello, What is the best settings to achieve a similar result like Deepart.io? Also how much RAM I need to generate images of 1000px using only CPU?

Thanks

[kineticsystem]

I just made a test for you: it needs minimum 12GB of RAM to generate 1000px images with CPU. It depends on the style size too, the bigger the more RAM you need. With 1300px you can easily fill up to 32GB of RAM.

I personally doubt you can get the same results like in Deepart.io, but this doesn't mean you get bad results with Neural Style, just different. For some styles I'm much more satisfied with the result of Neural Style.

I also want to add a note about the image size. I'm generating some images up to 1300px and I noticed a degeneration of image quality on large areas covered by uniform colors. They tend to come out considerably desaturated, like clearly visible grey patches. It may be related to the way the neural network has been trained, but I'm no expert.

So I'm moving in a different direction like generating smaller images e.g. 1000px and scale them up with other software (Alien Skin, Topaz, etc) I'm still investigating the best option.

Another option that has been mentioned in the forum is slicing the picture in overlapping patches, process them and stitch them back together, but I'm not sure how well this work with a neural network that has been trained to recognize special features in the given image.

What I'm saying is: do not assume that DeepArt works better and do not assume that generating bigger images will produce better results.

Take care,

Giovanni

[kineticsystem]

Regarding big images I'm having good results with this software as mentioned by other in the forum:
https://github.com/nagadomi/waifu2x

[joejacksonnite]

Thank you very much Giovanni, incredible help!

[baaart]

@kineticsystem but waifu2x needs CUDA so whats the improvement for CPU only workflow?:)

[3DTOPO]

@sheerun Come on now - 3 identical posts for your advertisement?

[3DTOPO]

I think it is in poor taste and not really appreciated, personally speaking.

[htoyryla]

I thought it somewhat funny to see the same thing in three subsequent emails with slightly different words. I assumed it was a link to a blog post or something, as usual in open source communities. Now that I checked the actual link and saw that it really is a commercial advertisement, I fully agree with @3DTOPO.

[3DTOPO]

I wonder what percent you plan on donating to jcjohnson? I suspect he has more time invested in your "solution" than you do.

[htoyryla]

I was not offended by your offer to sell settings. I was first amused, then a bit irritated reading the same thing in three emails one after another, and then when clicking at the link, finding no information, only an advertisement. It would certainly have looked better to see a blog article on the general principles followed by the offer to buy the actual settings.

As to deepart.io, as far as I know it is run by the people who made the original invention in the first place.

As to myself, I wouldn't have been a prospective customer anyhow. I have no need to make images that look like someone else's, and I have the time and resouces to experiment for myself.

[pocketmoon]

"I 💙open source software"

[htoyryla]

Once more, I did not disapprove selling knowhow, even by someone who did not develop the software in the first place. My reaction was to spamming the issues pages with an advertisement.

[MarieMandrela]

I got pretty close with the following parameters:

-num_iterations 1000 -init image -content_weight 5 -style_weight 10000 -image_size [biggest_dimension]

• The image is not initialized randomly but from the content image (this is very important).

• I increased the style_weight from the default 100 to 10000 at which point the result was good enough for me.

deepart.io:

neural-style (-init image -style_weight 10000 -image_size [biggest_dimension])

neural-style (-init random -style_weight 10000 -image_size [biggest_dimension])

neural-style (-init image -style_weight 10000)

[sheerun]

Here are parameters I've used:

th neural_style.lua \
-style_image $(pwd)/style.jpg \
-content_image $(pwd)/content.jpg \
-output_image $(pwd)/output.jpg \
-gpu 0 \
-init image \
-content_weight 5 \
-style_weight 1000 \
-style_scale 1 \
-image_size 500 \
-num_iterations 1000 \
-cudnn_autotune \
-backend cudnn \
-tv_weight 0.001

Please like this comment if they work for you, thank you