It has a very different interpretation than I expected of “Life, but not as we know it”. I was expecting one of the eldritch things from beyond the bounds of sanity, that it seems to be so good at, but instead got something genuinely poignant:
I have a feeling DALLE-2 will do a better job with some of the crazy prompts. From their feed we have stuff like
“something completely stupid, in the style of Magritte”
“a scene of perfect equanimity”
“triangle with four corners”
Yeah, I’d love to get my hands on that one. I wonder if someday the AIs will get good enough that they can pull off something like “Samura and ninjas drinking tea at Fuji with Bigfoot” flawlessly? (Spoiler alert, Nightcafe’s coherent AI can not.)
BTW, love that landscape painting.
That’s a pretty good face!
It’s not flawless, but the beta already does that, today. Did you check out the examples gallery (there are a couple of reddits aggregating them)?
cyborg apes building a base on Mars
Pulp Fiction played by the Muppets
an apple fighting a piece of broccoli
leopards playing dominoes in an alleyway
an elephant tea party on a grass lawn
Absolutely amazing. That needs to be a program you can install at home. Even if it does need one of those 15 kilowatt GPUs.
It looks to me like DALL-E is working at a higher level than Nightcafe: Like, it’s taking pieces that it knows are elephants and putting them with pieces that it knows are tea parties, or the like, as opposed to matching textures and more abstract patterns. Like, when Nightcafe draws an elephant, it somehow looks like an elephant, even though it’s so abstract that you can’t tell if it’s seen from the front or the rear. And of course it’s easy to look good, when all we see is the successes.
As it happens, “four-sided triangle” was one of my early experiments with Nightcafe. I think it did a better job of it, even if it’s not literally perfectly accurate:
That may be some sort of emergent property (the way neural network layers recognize straight lines at various angles, then at a higher level different textures and patterns, then pieces and patches, an eye, a tail, etc.), but it is not literally how it is programmed: DALL-E also uses CLIP and GLIDE and diffusion as components in its pipeline. The increased performance evidently comes from the way they put them together (and leveraged the aforementioned 15 kW GPUs for training the models)
Looks like the latest video card is supposed to provide 320 teraflops(!) for AI work. That’s more than 100,000 times faster than a Cray Y/MP from the late 80s.
This fall brings the Nvidia 4000 series which are to be power-eating monsters on the top end. Not that I expect anyone I know to own one but the new Titan is rumored to have 48GB of DDR6X and consume up to 900W. More consumer-grade cards will be lower but still hefty things.
I plugged “Tom Clancy’s The Division” into NightCafe and got an oddly accurate portrayal. It’s a video game about a government agency trying to hold New York together after a plague and the wintery image has piles of trash on a desolate urban street with an armed agent and even attempted the orange Phoenix logo. It’s certainly not an obscure title but it’s not like Pac-Man or Mario or something where it’s just part of the cultural flotsam either
I see they’ve categorized and added a lot more suggested modifiers.
Huh. That just happened in the last couple of hours. I see that they added ukiyo-e as an option, as I mentioned days I discovered worked. No Kei Toume Junji Ito, though.
I tried a number of the earier modifiers on tests with otherwise identical prompt, seed, and image. Most of them had very little effect. I also tried different comic/manga artists, such as Jack Sergio Aragones. Maybe that lattened e images a little, but nothing drastic. (You want to see drastic? Really—try Junji Ito.) i also hoped that lego-based prompts would make the image look legoized, but no luck.
I’m pleased with its interpretation of “The Man Who Wasn’t There”:
Thumbs up for Muppet Fiction! 
The sign in the middle almost says “the man”…
I only just noticed a link at the bottom of the page (buried in the “about us” info) for a pet portrait app. It looks to be just their style transfer app with a catchy name, but it put tge idea of a pet picture in my head like it was meant to. Searching through my recent pictures gave me a good one, and scrolling through their offered artworks provided the perfect choice. Here is The Yawn:
(I should have said purrfect, shouldn’t I?)
On a related note, does anyone know how to directly link an Imgur photo when it opens as a scaled webp and not as a jpeg link?
^^
Right-click on the image on the imgur page, choose “Open image in new tab”, then copy the url of the newly opened tab and paste it here on a separate line, like this:
I tried running some Diffusion on that old GPU. Not even close to enough RAM to max out the quality settings, and even low-medium settings lead to thermal overload 
So I will just have to cross my fingers for a DALL-E beta invite code…
That’s how it works some times for me. But others (like this time) it doesn’t give me an imbedable jpeg link, it gives me this:
