Originally published at: Dall·E's new outpainting feature extends images | Boing Boing
…
1 Like
Now I want to see the AI extend classical portraiture in ways that are entirely inappropriate. Kinda like the stretching Haunted Mansion pictures.
3 Likes
Yeah, to be honest I sort of hate this new technology. I already read that article about the guy who won a fine arts competition using Midjourney. Sort of sucks the soul of out art, IMO. Yay technology.
5 Likes
One consideration to mull over is that these machine learning models have a specific training set. Dall-E’s training set is actually a bit trimmed down from the original OpenAI GPT3 set. So like any ‘curve fitting’ model it only functions within the confines of its training set. Having typed that, it’s a pretty damn large training set (~8e8 images); and who’s to say that humans don’t also have a training set? Let’s just hope ours includes a lot of diverse sources, oh… like the real world, and not just what can be scraped off the internet.
Still would like to see a flowchart of how Dall-E is processing the text input. I suspect, but don’t know, that if you gave it the same text input twice you’d get two different results, and that would mean they’re “rolling the dice” for creativity in there somewhere.
4 Likes
People generally forget these days that work for work’s sake is not desirable. Ideally, all drudgery should be automated so people can focus on leisure activities, self improvement, etc.
4 Likes
But if you automate the drudgery before you implement another method for people to obtain the things they need to live, you are not going to end up with that ideal.
4 Likes
As an AI engineer of 20-some-odd years, I may be able to answer your question.
The short answer is that there is no flow chart, and in some sense it’s all dice rolls. The thing about these modern machine learning algorithms is that we don’t exactly “know” how they work. Or rather, what they do is an emergent property of a much smaller system which we do understand. What they are doing is taking a huge data set and using it to train a set of weighted nodes in a mesh (I say mesh because people misunderstand the official term “neural network”. That term has a lot of layperson baggage associated with it and poisons the conversation).
At the end of the day, it’s a just matrix math computation to run your words through the weighted network of nodes to generate each pixel in the image. It’s a giant statistics computation. You can’t think of it as a series of choices being made, like a flowchart. It’s more like probabilities. Think of it like, “in all the paintings in the world that have something to do with this word, most of them have a brown pixel here, so that’s what I’ll do”. There is a probability (aka randomness) at every tiny decision point there because it is a weighted node in the matrix, but the probability is not “intentional” as a dodge for creativity or anything like that. It’s no more “intentional” than the randomness in quantum mechanics is. It’s just part of how neural networks work.
10 Likes
Is…painting…now considered “drudgery?”
1 Like
I’m sure that Michaelangelo didn’t enjoy every moment of painting the Sistine Chapel.
3 Likes
It is if you are doing someone else’s painting for them with no artistic control and not enough money…
Isn’t creativity doing whatever comes to mind?
Real graphic designers can’t just copy classics or they get in legal trouble, or, worse yet the work they make is judged as “derivative.” Or they go to jail for forgery. But an AI is just doing what it does including glorious magic like this or stomping gamers at their own game.
A graphic designer needs to hear the clients desires and make it visual. AI is far far far away from being like that, but, it can do this real quick.
So I suppose it’s much like did YouTube destroy Hollywood because anyone can just upload a “movie.” Maybe that made sense, if you are a bot tell me I am wrong and why 
2 Likes
This topic was automatically closed after 5 days. New replies are no longer allowed.
