Less than 4 weeks to go. #btconf Düsseldorf. About time to get your ticket! beyondtellerrand.com/events/dusse...
09.04.2025 06:36
👍 5
🔁 2
💬 0
📌 0
Less than 4 weeks to go. #btconf Düsseldorf. About time to get your ticket! beyondtellerrand.com/events/dusse...
“To understand how we’ve achieved AGI, why it's only happened recently, after decades of failed attempts, & what this tells us about our own minds, we must re-examine our most fundamental assumptions — not just about AI, but the nature of computing.”
— @blaiseaguera.bsky.social & James Manyika
#ai
On April 24th I have the great honor to share the stage with Casey Reas and Christiane Paul for a conversation about generative systems at the Great Hall at The Cooper Union in New York City.
www.eventbrite.com/e/conversati...
Oops, I should have checked that a bit earlier: turns out my dawing board bends down 5 mm in the center. Not great if a millimeter more or less already makes the stroke width vary considerably.
Starting to get some more trust in my robot not running into a wall or trying to pierce my canvas, so I dare to allow it some "broad strokes".
I hope you explain to them that training a LoRA is not the same as training a GAN or a DiffusionModel from scratch. Since in case of the former you just nudge a model that has already been trained on millions of "other people's" images into the desired areas of its latent space.
Maybe my installation "Circuit Training" does match the criteria which is creating its own dataset in the exhibition context by inviting the audience into its automated photo-booth to become the training data:
www.youtube.com/watch?v=lXan...
Ah - re-reading your post - when you say "from scratch", does that include taking the actual photos, too, and not just finding and categorizing them? In that case of course I am off-topic here.
There was no real alternative back in the early days, if you weren't into just creating dogs, flowers, hotel rooms or tourist attractions (which to my knowledge were the only pre-trained models available back then)
All my early AI work was using models I had trained on my own datasets that I had collected/curated myself.
These for example are from my 2017 "Imposture" series that was trained on images and poses extracted from pornographic images.
Of course the loss explodes if I measure similarity instead of distance, duh. I still don't know what it tries to optimize for here, but I kind of like the Moomin-graffiti vibes.
Maybe I should turn down the learning rate somewhat. But it does explode in a quite interesting way.
It's easier for the algorithm to work with shorter segments.
Different starting conditions and target.
I have started to experiment again with diffvg and gradient descent on expressive vector strokes.
Oh, I fear that it will probably hurt your soul even more if you hear that Botto is now even creating #generativeart sketches with P5.js. And whilst a lot of the results are rather beginner level quality, some of them are quite interesting.
I'll be at @unireps.bsky.social this Saturday presenting a new experimental pipeline to visually explore structured neural network representations. The core idea is to take thousands of prompts that activate a concept, and then cluster and draw them using MultiDiffusion. 🧵👇
And for a glimpse into the "state of the art" at the time this is a good summary for what happened in 2016 and who was around:
www.alt-ai.net#watch
Those years might not line up with when the papers were published which typically was earlier, but for AI art the important dates were when the models and code were actually shared.
From my personal recollection I would say the history was:
2014 Image Classifiers (AlexNet, GoogLeNet)
2015 DeepDream / RNN (text generation)
2016 PPGN / StyleTransfer
2017 VAE / pix2pix
2018 pix2pixHD / CycleGAN
2019 StyleGAN / BigGAN
2020 StyleGAN2 / CLIP / VQGAN
2021 StableDiffusion
One very important moment in the early days of GANs and their influence on what we know now as AI art was @phillipisola.bsky.social releasing pix2pix - the reason being that it allowed to create larger and more "artistic" outputs than @ian-goodfellow.bsky.social's original GAN from which it derived.
The term "AI art" did not exist yet when people like Anna, Memo, Helena, Mike, Kyle, Gene, me and about maybe 5-10 others word-wide started experimenting artistically with them. So yes, I guess we shaped it.
Oh and diffusion models are not "early days" - that's like industrial age vs middle ages.
New year's resolution: remember to share more stuff in here.
I mean an LLM writing code, like p5.js which resembles a system that generates visuals or animations using various parameters and pseudo-random number generators - which is what I guess most people would understand as typical "generative art". And yes maybe a human prompted the LLM do to that.
This discussion is rather fruitless. AI art is generative in its nature - even the one that is just prompted. I wonder - what is your verdict on AI-coded "generative" (aka code-based/algorithmic) art? Does the code have to be handcrafted in order to be permitted to wear that label?
Sounds like "outgroup by" :)
What does it do?
Ah, too bad - so I giess like it's either a lot of splats or muffled sound.
It sounds like it is somewhat like a Fourier transform, but different.
Not sure though if spherical harmonics are too smooth for that or if it just a question of the right coefficient count.
