etherial harbor (fabric remix)

 

ochun's paired energies

 

Paired energy in this case being ochun's positive vs negative latent space attractor.

north shore hazy sunday

 

Trying some experiments with the Shi-Labs direct video synthesis generative model. Two runs of 16 frames each, time expanded and heavily processed in Studio Artist below.

waikiki beach synthesis

 

Synthesis (via stable diffusion) to contrast the resynthesis examples. Difference between above and below is associated with a lot of negative text prompting in the examples below.

north shore stormy weather

 

Transgressions

reality inversion

 

Loab

 

Resynthesis of an image of 'Loab', which this sciencealert article describes as 'a nightmare face haunting ai art'.  Loab even has its own wikipedia page.  Urban myth?  Ai art trolling?  Real phenomena?

I tried an 'image to text' synthesis approach and generated the following text off of the Loab image used in the article.  Same image i used as the one-shot seed for generating above via image to image resynthesis.

"the woman just she their the woman she completely not having just her girl all her real <unk> some women have <unk> and she

several women very shocked her life like she just had a lot of fun around a giant mushroom giant woman and her little daughter

she almost died child who made famous was shocked at her death so very angry she <unk> joke and smiling at dead lady with ear

a woman with a child who's almost to hurt her mother had her ugly face, <unk> who she was dressed in a hospital"

I then ran a few iterations of a resynthesis embedding drift on the original Loab image and got the image below.

Gallery Show Resynthesis

 

hanapepe art night resynthesis drift

 

reconstitution

 

Restructurement

 

Resynthesis embedding drift session starting from the gallery show mutation output image above.

isometric alien abducting alligator

 

There's an interesting editors letter in the latest issue of Digital Art Live Magazine talking about isometric art.  In the letter, editor David Haden mentions a potential generative ai thought experiment using the text prompt 'UFO abducting an alligator from a creepy swamp at night'.  It inspired me to try a few experiments.

I first tried some experiments using just the subject matter text prompt.  Pretty hit and miss, most of the images are bad renders of weird alligators.  This one below seemed kind of in the ballpark, so i was primarily working off of it as a base for the rest of the experiment.

If you just try to modify the style prompt to incorporate some form of 'isometric grid' extension, the results tend to be pretty bad.  This includes various 'prompt engineering' approaches to coerce better output.  One example from that batch of experiments shown below.

I'm a big proponent of using image embeddings and image embedding manipulation or modulation to control generative image synthesis, so the rest of the experiment was really focused on working with generative resynthesis.  Most of those experiments used that image above that i said would be the base image.

I recursively resynthesized previous output images, subjected to a second push of the attention embedding to try and push it to have more of an isometric grid influence.  The first 2 images up at the beginning of the post are representative of where i ended up.

I also tried some 'sci-fi' poster variations, where i started with resynthesis of old 50's era alien invasion poster art.  You can see the progression of the 'pushed' recursive resynthesis cycles below.

Recursive Succession

 

Gallery Show paint synthesizer auto-mutate grab image above used as input to one-shot generative image resynthesis below.  Again, this is all about recursively manipulating the image embedding used as the input to the generative resynthesis algorithm.  At the end of the recursive resynthesis processing session you end up with an image transformation (bottom) derived from the input image above.

movement

 

Using an auto-write paint synthesizer live paint stream output to drive the generative image synthesis U-Net latent input.  Just the successive paint strokes above, below also incorporates a secondary wet water drip, so i used Animate rather than Stream output to generate the U-Net latent input for those.

head shots

 

Envision

 

Resynthesis Embedding Drift 3

 

Resynthesis Embedding Drift 2

 

Resynthesis Embedding Drift

 

Recursive processing generative image resynthesis experiment.  I'm not preterbing the image embedding in this one, just feeding the previous output back into the input for the next resynthesis cycle.