Good call on it being a robot. The prompt was:
“Robot android cyborg Yurei parade | diorama | found objects | hyperdetailed colored pencil drawing | 8k resolution | hyperrealism | volumetric lighting | photorealistic | dungeons and dragons guidebook | postcard” - weight: 1
“Dan witz | margaret keane | junji ito | Leonid Afremov | Wadim Kashin” - weight: 1
“Golden | silvery | metallic | enamel” - weight: 1
(I keep wanting to come up with something strikingly different than anything before.)
Just got a “Sociable” badge and 10 credits for reaching 50 followers.
Night Cafe says that they are adding a new algorithm soon. I wonder what it will be?
I got a coherent face for the firs time. Do not know what I did for that to happen. Another song lyric: Sunshine in a bag
Did you use just the song lyrics as the prompt?
Nope. With this one I used Margaret Keane also.
Well there you go. That’s definitely what made it happen. (Her name is influential craiyon.com, too).
Yes, I tried one of those “new” beta invite links, and while the results (which I will not post here, with all the traceable watermarks) are undoubtedly impressive once you figure out how to fiddle with the prompt to produce exactly what you want, your 15 free credits per month do not go very far when it takes you 20 or so tries. It is an expensive commercial product.
I think it was already mentioned that not only can you upload your own image (or generate one from a prompt) and edit the whole thing, you can mask specific regions to work with. Too bad the paltry amount of free credits are not sufficient for any real experimentation, even the gamma version is already obviously a fun and useful tool to make some awesome art.
So can it actually reasonably generate the “glamor” images shown in promotional materials? I’ve been playing around with the copycat version and while it can produce fun freaky images, it produces nothing actually very good. (I can’t post any more examples in that thread because Discourse won’t allow me to make more than four posts in a row.)
I don’t mind blowing a few free credits to try out whatever prompts you suggest (NB you can do what I did and put in an email address and wait an unknown length of time for an invite link; they are pimping this out hard now, especially since Google already has something better), I’m just worried about posting them since I do not like their privacy policy.
Maybe if I crop some of the more obvious crap out. E.g., here is “a snake on a surfboard in Kowloon bay, 50mm f/1.8”, but that is pretty easy, of course:
Buffy the Vampire Slayer and Mork, eating breakfast in a Parisian cafe:
Hmm. But you could spend lots of credits engineering the prompt to get exactly what you want, or crop out Mork and start replacing just him, etc.
Fairly well composed. This one seems to think Spike is Mork. It seems to be like “Craiyon”, going for photographic images on some subjects and drawings on others.
If you want a photo, you need modifiers that imply a photo, e.g.:
Buffy the Vampire Slayer holding a kitten, three quarter profile, 50mm f/1.8
Those are definitely vastly beyond the capabilities of Craiyon.
And of course Disco Diffusion isn’t remotely similar.
I’m still wondering about the new AI at Night Cafe. Their teaser definitely shows Dall-E 2 quality images. Of course, they could be basically lying with those, but if not, what could it be? Their two others are both openly available elsewhere but I don’t know of any other open one of extreme quality. I wonder if there is any chance that they’ve been secretly training their own?
Dall-E 2 is on Github
even including some suggested improvements/hacks like combining Dall-E 2 + latent diffusion.
So is Imagen:
Also, if you know what you are doing, you could experiment with your own models. So it is conceivable that Nightcafe have been secretly training their own. At the moment in 2022, though, (and at any moment, relatively speaking) extreme processing power for training is not cheap. Hence why you can give OpenAI your credit card number right now and go nuts (I’m seeing 115 credits for $15), but why nobody is offering the same-quality pre-trained models for download or on a free-to-use web site.
photo of a huge pile of lollipops, studio, smoke, sharp, sigma art lens
cyberpunk elven girl archer, digital art
“a dog, a cat, and a duck playing croquet”, by Vashti Harrison
489 forks!
@running_coach should recognize the starting image here.
And this is the final evolution:
