Clinamenic

Binary & Tweed
The output is impressionistic to say the least.
Yeah I'm not sure if its the particular GAN itself or if its because higher amounts of text just confuse GANs more generally. But yeah I wasn't getting especially interesting imagery, just found it interesting that you can type in new prompts and it almost keeps up in real time.
 

catalog

Well-known member
not had much good results with starryAI, the phone app one, but noticed that's got a pretty active reddit, so must be doing something right

 

catalog

Well-known member
hypnogram seems to have died - not working for me from home or at work - I wouldn't be surprised if they have run out of money for the GPU processing, it's been hammered the last few weeks - back to google collab for me


@catalog here's an online AI that allows you to upload photos - i've not investigated further because it relies on a "credit" system ( means money changes hands.. ) - it appears to be a variation of STYLEGAN ( i.e., style transfer )
it's actually pretty decent, although no option to add a text prompt.

used the initial image (on top) and then selected "bitcoin goddess" (below) (think that's what it's called) style

Screenshot 2021-11-23 175524.png

and it makes this

nightcafe 1.jpg

used all my credits doing that but you can buy more so might do.
 

catalog

Well-known member
not that i know of, but i feel like there's gonna be loads more web front ends coming up. paying for that processing power is the issue definitely.
 

Clinamenic

Binary & Tweed
not that i know of, but i feel like there's gonna be loads more web front ends coming up. paying for that processing power is the issue definitely.
The free google collab seems like it gives enough for our purposes. Pretty sure its just a daily refresh.
 

Dusty

Tone deaf
Related to this thread, but with a little more human intervention; I use an old android phone that was kicking about the house doing nothing to generate GAN landscapes online, then feed them into a glitch app to fuck them up a bit. Delete bits... keep bits... build up layers of glitched artefacts, before finally processing with photo editing software, (again, limited to Android apps only).

My main problem with GAN generation is the low resolution output, suitable for viewing on screen only. With these I've been working to effectively obscure the low res GAN and upscale it so that the full size files are roughly an A4 600dpi resolution.

So its artwork produced via selection. Picking the best AI generated landscapes, picking the best automated glitch artefacts, all done with nothing more than a few setting sliders and an erase tool.

www.glitchedwork.com
 

william_kent

Well-known member
here's another online AI art generator

neural blender

it works on a "credit" system, but it's possible to get a couple of "free" turns - they're a bit funny about copyright ( not yours unless you've paid for "credits" )
 

catalog

Well-known member
Related to this thread, but with a little more human intervention; I use an old android phone that was kicking about the house doing nothing to generate GAN landscapes online, then feed them into a glitch app to fuck them up a bit. Delete bits... keep bits... build up layers of glitched artefacts, before finally processing with photo editing software, (again, limited to Android apps only).

My main problem with GAN generation is the low resolution output, suitable for viewing on screen only. With these I've been working to effectively obscure the low res GAN and upscale it so that the full size files are roughly an A4 600dpi resolution.

So its artwork produced via selection. Picking the best AI generated landscapes, picking the best automated glitch artefacts, all done with nothing more than a few setting sliders and an erase tool.

www.glitchedwork.com
this is great stuff - you can use the colab notebooks to create any resolution you like (with concurrent processing time required) but running those on a phone would be pretty difficult. i don't think impossible but wold be pretty annoying.
 
Last edited:

william_kent

Well-known member
I've found the problem with the "pro" account on google colab is when you try to create larger images as it will crap out with an "out of memory" error and a hint that maybe the "pro+" account may be better ( and cost more.. )
 

catalog

Well-known member
i've always limited mine to low anyway, like never more than 600 square, cos i'm more interested in the videos anyway and only really youtube quality. but i've had those errors and they're pretty frustrating.
 

Clinamenic

Binary & Tweed
Related to this thread, but with a little more human intervention; I use an old android phone that was kicking about the house doing nothing to generate GAN landscapes online, then feed them into a glitch app to fuck them up a bit. Delete bits... keep bits... build up layers of glitched artefacts, before finally processing with photo editing software, (again, limited to Android apps only).

My main problem with GAN generation is the low resolution output, suitable for viewing on screen only. With these I've been working to effectively obscure the low res GAN and upscale it so that the full size files are roughly an A4 600dpi resolution.

So its artwork produced via selection. Picking the best AI generated landscapes, picking the best automated glitch artefacts, all done with nothing more than a few setting sliders and an erase tool.

www.glitchedwork.com
You consider representing these as NFTs? You could do a limited series.
 

william_kent

Well-known member
hmm.. I wondering if hypnogram have messed with their code - images are quick to render, but seem a bit lower quality and feature ugly thick black lines that weren't in the images created before the break in service - also, the auto naming of the downloaded files has disappeared
 

william_kent

Well-known member
NVIDIA have announced "gauGAN", their text to image application, although it does also allow the uploading of "sketches", "images", and "segmentation maps"

press release

the demo or whatever it is supposed to be is here - first impressions are a horrible interface, confusing workflow, functions that don't work - any text I entered just ended up with a image of the milky way - looks like there may be a learning curve - in fact it looks like the UI / UX has been designed by programmers from the 20th century
 
Top