Clinamenic

Binary & Tweed
download-1.png


Yeah, so I have changed the text input to:

beige up | jade down | banknote engraving 'treasury'

will post results later
 

Clinamenic

Binary & Tweed
Yeah so it looks like this kind of steering could work, just very manually. I did the:

beige up | jade down | banknote engraving 'treasury'

run for 500 iterations, outputting every 50th, and ended up with

download-10.png

which i am using as the initial image for the text input (edit: that is, a new run with a new input):

swirl | fire | '$'

and this is the first result

download-11.png
 

Clinamenic

Binary & Tweed
So one could do this kind of reparameterization every so often, and spin the transformation into a different direction, with a different set of textual directions.
 

Clinamenic

Binary & Tweed
So now I;m gonna start with a cropping from one of my grandmother's pieces Black Girl's Window

Bildschirmfoto-2019-10-13-um-20-03-29.jpg

and I'm gonna try to have the GAN latch onto the existing imagery and extrapolate from it. I am using the text input:

hieroglyph | star | moon | eye
 

Clinamenic

Binary & Tweed
Although it seems like 50-image intervals here may be too much. I'll run this for a bit, but maybe next time I'll switch it to 10-image intervals.
 

Clinamenic

Binary & Tweed
Now I'm trying, with the same initial image:

star | moon | eye | sunburst | in the style of Betye Saar

and I've set the image interval to 10, instead of 50, which I think will make the alterations smoother, finer.
 

Clinamenic

Binary & Tweed
So this new one has four different runs manually edited together and configured. It started with the initial image, a cropping from a piece of her's called Phrenology Man:

download-00.png

And the first run was on text input

Phrenology Man | in the style of Betye Saar

with image interval 2.

And I ran that for maybe 20 iterations, then took that final output, and used it as the initial frame for a second run, with text input:

pink | halo | shine

with image interval 2

and I ran that forward and reverse, to get back to the final image of the first run, which I reused as the initial image for the third run, which had the text input:

afro | vibration | ripple

with image interval 2

and I ran that forward and reverse all the way back to the initial image of the first run, the actual cropping, which became the initial image for the fourth run, which had text input:

crystalline | smoke | violet

again with image interval 2.

So here is the final gif, with 239 images, mostly repeats.

 

Clinamenic

Binary & Tweed
This is how I imagine genetic mutation on the phenotypic end, the front end.

The different iterative sequences each with unique input texts, stemming from a common basis, mutative probes, etc.
 

william_kent

Well-known member
This is how I imagine genetic mutation on the phenotypic end, the front end.

The different iterative sequences each with unique input texts, stemming from a common basis, mutative probes, etc.

full disclosure: i was a beta tester for "cryptokitties" and although "breeding" was fun, as soon as real money was involved I was OUT

that's why GANs are interesting to me, I'm a open source kind of guy, no money involved at the moment, plenty of source code to play with
 
Top