?Filters like Pikazo?

Quote
And there was some gray fade outside of the cat, so here using this other cat photo the result is more unexpected

But really cool.

So far deep style works very well with variable patterns. So Pollock's work does a good job.

I haven't quite gotten my own deepstyle filter "there" yet! But thanks to ThreeDee and Sphinx sharing their wisdom, I'm getting better.

I can't say I've actually mastered the "getting the final image to look like the style image" step, but at least the results I'm getting are more deepstylish.



+



=

I do certainly seem to be getting closer though.

Van Gogh's Starry Night series seems to be the classic test style for this style of deepdream, and this is quite a pretty Starry Flower And Ladybug.




+



=

Def haven't mastered this by any means, and what I have is definitely dirty, but it's also definitely looking more and more like deepstyle.

Can't wait for more input on this fantastic thread.

I want to make deepstyle for the rest of my life. Drugs could never be this addicting. Cattybug says hello!

Some more food for our minds here. This is the best open-source deepdream algo I've found so far; his work produces some really nice stuff on the level of Pikazo for sure.

https://github.com/jcjohnson/neural-style
https://github.com/jcjohnson/neural-st..._style.lua

I'm still parsing the code and it may take months for me to learn neural networks well enough to really advance our work here, but others may have more experience.

I can tell however that there is a concept of "style loss" that represents the difference between the original and the style to be imitated, and that's certainly something that could be possible with multiple script passes in FF.

It helps that both this page and FF's native scripting are Lua - so there are certainly features of the learning algorithms that could be repeated in FF!

Indeed. We certainly can't call custom libraries or distributed-computing databases, or run exoflop-level computations on our home machines.

Thus again, my notion of calling the process we're trying to invent "shallow dream imaging" - on-the-fly analysis of an image and a style and the redraw necessary, instead of true deepstyling relying on distributed processing or existing neural-net databases.

Also, thank you for the kind words on my filter. That being said, ThreeDee's work is definitely the crown of the camp so far. I keep hoping he'll let some more crumbs of knowledge dribble about his process. And I think there are things to learn from the process from all these whitepapers.

ThreeDee, I will worship you for years to come if you release your deepdream filter. This thread still haunts me. I know you don't feel it's perfect, but goodness, it's good.

They began selling in-app purchases where you buy these tokens called Jetons. My stuff's displayed in JPY so I got no exact $ equivalents but its around $5 or $6 for 500 Jetons. You kinda get the gist of the price tag here.

So for a 800x800 quality it's 79 jetons
For 1600x1600 its 349 jetons
For 3200x3200 its 1749 jetons

I think that's reasonable I figure for what it does.

As far as the workflow, you basically run an image through their service for free, and if you want to scale the resolution up, you pay. So there's no guess work involved. You do get a free 400x400 image before moving to larger images.

Personally, I wouldn't recommend the 3200x3200 and instead settle for 1600x1600 and use neural-network based upres techniques to scale up. The reason for this is that the 3200x3200 image... looks a bit dirty. I'm not sure why but the details are squashed and it just looks kinda ugly close up. No point rendering something that large and getting something kinda not as clean. Please note that that's just my personal taste, maybe it might turn out better for you guys.

There's been a number of proposed neural-network based up-res apps for example: http://waifu2x.udp.jp/index.html which have proven to be extremely effective with up-ressing illustrative art. I generally find Waifu to be the best when it comes to the quality of output. It almost scares me how good it is. And the name's pretty hilarious too if you know the meaning of it lol.

I bought 5000 jetons to see what would happen since I've always enjoyed neural network type of things and noticed that it's just a pain in the ass trying to do this on my AmazonEC2 instance.

If anybody's interested I can show you how to run your own Neural Network on a AmazonEC2 instance with 4GPUs. It's a bit more complicated and while you have more freedom to do whatever you want, it's generally recommended for those who really want to go on an adventure. As for pricing, as far as I've been running pictures through it, it's costed me a fair much. I say buying tokens may be slightly cheaper and more effective since you're not the one running the operation and there's a guaranteed uptime for that.

Hi !

never saw this deepstyle/neural-art thread before !

I'm practicing neural-art (jcjohnson version) since several months and it's properly astonishing ! results are unbelievable !

however I think it would be really hard to "adapt it" to filter forge, as skybase told, it's running in its own way (with Cudnn, the VGG system and so on), and it's what make results quite unique, not like a filter processed image, but more like an image rebuilt fr om scratch.

anyway, What i think is interesting, and what i'm experimenting, is to work with neural-art AND filter forge. I've tried it on animation (as I run Neural Art on Ubuntu, it's simple as hell to batch process an image sequence, you just have to be really patient), and then batch process this output sequence into FF (using the EarthBound FF batch Wrapper), and testing different FF filters, and then mixingg those sequences (the original one, the Neural-art one, and the FF ones) into after effects (or any other composition program) to control the final output, and it gives pretty unique results !

the main thing you have to consider about neural-art is it's huge need of VRAM and a Cuda GPU (so mostly every middle to High end Nvidia GPU). I started on a GTX 660 (Cuda compatible, but with only 2 Gigs of VRAM) and I was able to output only 300x300 px to 450x500 px neural-art image. With a gtx 1070 (8 Gigs of VRAM) I can go up to 1200x1200 px (depending on the Style image, some settings, use of ADAM optimizer or not).

don't even think about running it on CPU, It works but You will wait for ages to output a small image. (i'm talking about hours... wh ere the same process ran on a Cuda GPU take several minutes)

Skybase -> as for waifu2x, I used it several times to enlarge my neural-art outputs, I agree it does an amazing work but on specific styles, most likely when you feed the neural algorithm with US comics styles images/ serigraphic illustration/ "Art-Nouveau" and so on, But when it come to grainy/messy images with paint textures on it for example, waifu enlarged results aren't that convincing (mostly because the waifu model have been trained on japanese illustration, so it does a very impressive jobs on thin lines/curves enlargement and simple color texturing).
Anyway, there is another solution which is coming (no stable yet if i'm not wrong, I didn't tested it myself yet) to get larger outputs from Neural-art, it's a tiling system some people are working on, and it seem awesome ! see some development discussion about it here ->
https://github.com/jcjohnson/neural-st...-230351054
and here
https://github.com/jcjohnson/neural-style/issues/337