?Filters like Pikazo?

Wow!! This app seems to be incredible, amazing and awesome for making art from photos

Thanks very much for the link and news

there are many beautiful and cool examples in the official website going down the page

www.pikazoapp.com

preview video trailer of Pikazo

Quote
Amy Hutton wrote: Wonder if there is any chance folks here might recreate some of these effects??

I think that maybe it could be possible perhaps in some way some of the effects, but others surely not, specially considering that this is done using a neural network software and is not like a filter effect

Extremely impressive!

Deepstyle unfortunately takes extremely long to render. Meaning sizes are probably going to be severely limited unless we either get over the inefficiency hurdle or we just make better processors or both. I think the implementation I've installed is limited to 512x512. But I'm sure there's a way to get over that by overriding some variable.

I guess the screencap-crop method works. Although just as another way, try using those detail-enlarge plugins too. It may help.

Quote
I think such a thing should be possible

Again, Deepstyle works on a totally different level. Heavily based on deep learning algorithms, these things actually understand images. The characteristics of deep neural networks are a lot more sophisticated than just a filter. Hence you can feed anything you want into deep style programs to output unique results. It's not a filter, it's a machine learning/understanding an image and producing an output based on it.

Machine learning is what enables search engines like Google to look at images and understand it's contents. Deep style (pikazo app) and deep dream are both products based on google's research into machine learning.

As for that, you can feed deep-style program pretty much anything, including stuff like rocks, tiles, paintings, drawings, scribbles, etc. Just for fun, I threw in a picture of a cat and styled it to some stone wall (wouldn't that make a nice filter idea? haha).

------
With that being said, I guess most people are looking for painterly effects. Not necessarily deep style. I'm just trying to make sure people understand the difference.

The "stoned" cat is pretty amazing, I must admit. Sorry, Rabid, Filter Forge is not made for this kind of processing. As Skybase has mentioned, deep-dream is too "data" intensive (large database in the Google version) and too expensive to process.

Using the free source code, I created one 1280 pixel wide image. It took 22 hours with the free deep-style software using a style and a source image. I think eventually someone will come up with a more efficient algorithm for the 2 image version (like Pikazo) and for the desktop. Google's version requires a very large database and is not a traditional "filter", more of an array of computers doing pattern recognition while learning on the fly.

I mean, honestly, I'd buy a print or framing of your cat picture there. (I'd buy a lot of prints of FF work, as well.)

I just look at that and feel this burning need to do work in this field - and I don't have an iPhone, either.

The deep style program actually works extremely well with patterned images. Specifically stuff with variants and noise.

There's a tendency for the program to produce artistically unsatisfying results (just a distorted mash of boring) when the seed image is too distinctive or too bare of detail (aka pictures with little specificity.) I think there's some kind of "just right for the program", for example the stones image came out pretty successfully.

My advice is - use the Color Control "Allow Images" feature in order to let a second image into your filter, Amy, and then play! I will be, all night perhaps.

Thank you from the bottom of my heart to one and all for this amazing thread!

Nothing I can find out there seems to match Pikaso, though.

Just one final note as I don't want to overpower the discussion: Don't give up on the "Deep Dream In FF" notion just yet.

I worked with Rabid's notion a bit to cannibalize the bomber from Buttons, and while I can't say it delivers Pikazo-like results, it definitely seems to be a step in the right direction. This is what I have telling the Ladybug filter image to draw itself with itself and then blend itself back in, for instance.

I'll see if I can push this "Dreaming, Moderately" submission in tonight too.

Getting better results than I thought, even on the non-preview stuff! I def recommend my Dreaming, Moderately filter once it's submitted.



I don't know that it's deep dreaming per se, but it might give you enough of a fix to get by - at least until Pikazo is available for more than just the iPhone!

Your images aren't displaying correctly.
Safari, Chrome. Both PC / Mobile don't display.

If you're hot linking, use imgur avoid direct links from other websites.

Is true that the images can not be seen, only the cat in he last post

Quote
LexArt I put it here with Image link and it works right as shown below

Well, it may work for you, and maybe you can see it, but I can not see them? Can you put a screenshot of how you are seeing them?

Here...result 1 as a direct image.

Obviously, my filter is not doing really serious neural-network analysis - the merge is a rather hacky blend, not a real "redraw" in the style of the original as Pikazo is doing. It can't match Pikazo's excellence in the least.

That being said, I've found that the filter is actually giving better results than any of the simpler deep-dream stuff out there. Rabid's simple suggestion of using the second image as the Bomber particle really gives a very nice effect. And I think that it may be easily possible to remove the "originalness" of the blend with more work on that part of the filtration step - maybe use the original picture as a map for Refraction or Noise Distortion, or else a custom map script.

And the stuff that the neural-network analysts are doing *is* possible in Filter Forge, according to my best analysis - it's just math. Very, very complex math, but math nonetheless.

Some fellow eyes on this whitepaper and its Methods section would be lovely.

http://arxiv.org/pdf/1508.06576v2.pdf

No, I do get that, Skybase. I get that we can't access VGG from within FF and that it's not designed for that.

I'm referring not to the specific methodology but the general process involved - the process of analyzing an image, then "redrawing" it in a style based on a second image.

Call it "shallow dream imaging", perhaps - the creation within a filter of a spontaneous, small set of filters approximating more complex tasks that NN's do during this process. A filter is not a neural network - but the multiple paths within filters operating on an image *are* a conceptual representation of what happens to images in human vision sometimes, breakdowns into different recognizable components that are recombined.

We clearly won't have access to large preexisting computer vision databases designed to help in shape recognition.

The process by which the whitepaper authors describe the breakdown of the first image into an outline/guide for placement, and the second into a set of texturing information/stylistic components to redraw an image in, though - that general process is tooootally possible in FF, especially with access to Lua in the scripting components. The amount that can be done on any data set with access to a Turing-complete processing language in my opinion can astonish, and an image is just a data set, at least as far as the language is concerned.

Pikazo is clearly leaps and bounds ahead of everyone else in this space.

But I'd already pick FF over all the rest I tested, and I tried, I think, three other sites/apps.

Or same the other way "Window" as target image, "Lifesaver" as source.

And another idea, using one image as the background, and using luminance masking on 2 other images and blending.

Nice and interesting examples you have put Rick

Garlic from Lifesaver.

This was the source image