DeepDream

[Yak]

Running a Space Giraffe screenshot: no surprise it's Space Oxen all the way down.



Another screenshot, more horns:



Oh dear. I think I preferred the horns.



more oxen





I guess I found the antelopes.

[Yak]

I downloaded a differently trained net to see what kind of stuff it'd come out with. First I cut it loose on the TxK logo:



It seemed to be thinking of bowling alleys and the floating heads of children (not Kevin Toms). It's supposed to be trained on "places", so let's give it the Space Giraffe screenshot and a glimpse of the Houses of Parliament and see what happens:



- certainly a lot of architectural structures emerge, some of it vaguely Parliamenty, and is that the Tardis on the left? And hints of a red telephone box in lower right? nope, just a trick of the algorithm and suggestion by the colours present in the SG screenie.

Dunno what the deal is with the giant elephant though.

[Yak]

Horndogs



View across my neighbour's field



Do not adjust your set

[Yak]

It's interesting to compare how changing one parameter changes otherwise identical runs. Here the input image is just a grid of tiny red dots on a black background, and I'm running to nearly the full depth of the network.

With the guide image of a picture of a highland cow I get



- looks like blenders or something. Changing the guide image to a picture of a radio yields



there are some common structures, e.g. the "horn" in the upper right, but the second image definitely inherits some of the "radioness" from its guide.

Sometimes it's good to leave enough of the source image that it remains obvious what it was. Mario takes a trip:



Let it go deep and it's all flesh and radios.

[Yak]

The "ascent step" mentioned in the source code affects how the dreamed images blend over the current frame. Reduce it a bit to soften the overlay of generated imagery and preserve the colouring of the source image for longer. These stilt-fish-poodles were generated at an ascent step of 0.9.

[(o)(o)]

Any chance of a Salvador Dali?

[gumboot]

Animated gif of ascent step ramped from x to y?


Viewing these pictures on my phone completely ruins them. Don't do that, phone-using people.

[Yak]

Here's fluffy Jerry transformed into a double-headed pug-slug.



Tried with Dali (no guide image; Dali should be his own guide). Here's the original; there has to be melty watches so:



4 generations in and Dali's toking up:



50 generations in and it's clean off, all trace of watches gone.

[(o)(o)]

Thx for Dali very nice

Her'es a few other art based ideas..
http://www.artpause.com/upload/uppic/24 ... poster.jpg
http://orig12.deviantart.net/a5d3/f/201 ... 53eljp.jpg
https://upload.wikimedia.org/wikipedia/ ... pheres.JPG
http://euler.slu.edu/escher/upload/9/90 ... it-III.jpg

Repeating patterns seems to be effective,, the Mario one and BBC test card are delicious ..

[Yak]

Here's the Mondrian at output 5a, no guide image, not trying any of my usual trickery of halting the run, changing shit and resuming it. I'll run them all on the same settings.



Definitely a chalicey thing going on with that.

Wain:



Right from the start this was insecty, getting a couple of fairly well defined bee-things within the first couple of iterations. Further iterations add fish, iguanas and again with the chalices. Maybe 5a just has a thing for chalices like some layers do for dog noses. We'll see as we move on.

Spheres:



Not quite as subtle as the original, instead of the outline of some features we have a melted load of eyechinflesh. The outer layer of spheres have retained their integrity and gained an aura of fishbudgietennisball.

At least the chalices have been given a rest.

MCE:



We lose the intricate detail of the original and are awash in a blur of laboratory glassware and chemistry students. And the bloody chalices are back, making me think that layer 5a is gagging for a drop of splosh. There is a faint whiff of bicycle about it too.

Personally I'd've run that last one using a higher rez original, fewer iterations and more of an elevation step to try to avoid losing all the detail in a wash of chemistry student. Bit I wanted to run all 4 to the same depth on the same settings to see what happened.

[gumboot]

It's a bit obvious, really, but:
https://www.google.com/search?tbm=isch& ... on+pollock

[(o)(o)]

Nice one..

That Dali in particular is nudging ever closer to 'actually a bit reminiscent of tripping fishbudgietennisballs'...

[Yak]

Run with the same settings and to the same number of iterations as the others, Pollock;s splatterings resolve into a bunch of bird, insect and lizard detail over a busy, hazy layer of organicness.



I thought I'd give MCE another go with one of his more linear tessellations and see how it goes:



a river of chins embellished with lizards and violins, according to 5b.

Let's shift the palette and try a run on 4e:



colour change and a different layer yield a much less literal interpretation of the picture.

[Yak]

Tried the Mondrian with the other differently trained net.



My first actual pagoda. And that *is* a red telephone box, next to some kind of Chinese structure. There's mountains and viaducts and bowling alleys and... is that people playing baseball?

It's like a postcard from Tripout City. Nice.

[gumboot]

Perhaps this technology can help us figure out alien messages...

big:

Spoiler: show