It was never the case for “coherent”/disco diffusion at Night Cafe.
One curious thing I’ve found with the offline Stable Diffusion–for img2img, if I use the same seed but different prompts, I get varying but similar images:
They all produce the same little head tilt and squished features (except for the statue one), which is not there in the original image. The color selection is similar–partially derived from the background, but not completely.
Not all of the seeds were so consistent, but most had some degree of similarity across prompts.
I’ve seen something similar, playing with Annie Leibovitz prompts. Note how similar the first two in each set (an ugly family and a dead family) are, where each of those had the same seed.
My latest algorithm (at Night Cafe) consists of running whatever prompts I’ve settled on with 80% prompt accuracy [I’ve found that that slider basically affects how sharp the image is and not anything to do with composition]. The prompts I’m falling back on are pretty well vetted by now, which I tweak here and there, so that image is usually pretty cool. Just to be safe, I evolve it - that is, use it as a start image - with a new random seed and then give it the 80% prompt weight as well as 80% noise (using identical prompts, of course), and I’m getting really interesting results. The evolved image, nine times out of ten, is a real improvement and most of the time as well, both images are winners.
Another thing I wanted to mention - in case no one has figured it out (it took me a little while) - is that when you get the “too many words” warning, it’s not counting words; it’s counting punctuation. Actually, it’s at least counting commas for sure - I’m not really sure about other punctuation .
I read in a couple of places that OpenAI Dall-E now allows you to “upload and edit photos with faces”.
BTW, I see that Nightcafe offers large packs of credits for around 5c/credit. But IME if you are really going to use it, you will easily find yourself doing a hundred renders just to perfect a simple final image.
Stable Diffusion disappointing has less of a grasp of some pulp culture/media characters than Min-Dalle and Craiyon. In the past I have tried E.T. and Alf and the two older systems drew them badly but clearly “knows” who they are while SD has no clue. Today I tried creating a rancor in SD. Again, the two older systems knew well what a rancor was but didn’t draw it very well, but SD didn’t have a clue, drawing some four-legged hippo-bulldog thing.
Lately I’ve been playing with gremlins. I get very interesting results, but they don’t seem to have much influence from the movie. And I just noticed that Night Cafe has added some complex new custom presets (they may have been there for a while, I haven’t looked in a while). Here are some scary ugly gremlins with the “portrait” preset:
But for something creepier, try in SD:
Scary ugly gremlin haunted forest 50mm Canon f2 photo
Scary ugly gremlin haunted forest 1900s photo
This one
leverages these models to make short videos (e.g., from a text prompt).
Does Nightcafe have it already? There were some video options there but I have not tried them.
There is now a completely free AI image website with some cool features, including the ability to load starting picture and choose between Dall E and Stable Diffusiin and bunch of other stuff:
Wow! I thought all the intensive computing time cost a lot.
My challenge to y’all. I can’t for the life of me generate an image of Putin riding the bomb falling out of the bomber like Slim Pickens.
All the art generators keep putting him on a horse.
One trick I’ve seen is to first get a picture of a random person riding the bomb, then use inpainting to replace the generic face with Putin’s.
I tried just to get slim pickens riding the bomb and couldn’t do that either. And don’t even try T.J Kong riding the bomb. You just get a bunch of King Kongs (Ape version).
The training costs a lot, but the evaluation (“inference”) is really pretty cheap. My GPU cost about $700 and does a Stable Diffusion run in maybe 2 minutes. Even if the GPU only lasts 2 years, that’s about a tenth of a cent per run. The power cost is even less; maybe a hundredth of a cent.
At least for OpenAI, they’re paying the salaries of a bunch of researchers and have to keep a decent-sized business running (even though it is a non-profit). That’s not true for the open-source versions.
That said, I’m not sure how playgroundai is providing free access to DALL-E. I thought that would have to go through OpenAI, which as I said charges a decent amount.
Maybe this is an ‘enjoy it while you can’ situation…
Well, at least when I added “like in dr. strangelove,” I got a nice image of Putin straddling Dr. Strangelove. (Apologies to the Dr. Strangelove in this thread)
Can’t wait to get my hands on that video generator, or one like it…
Congressional Rep Eshoo (CA-18) sent a letter to the NSA and the Office of Science & Technology basically saying they needed to “do something” about Stable Diffusion and AI rendering tech. The heart of her complaint was that Stable Diffusion was released knowing that there was porn in the database model and even giving directions to disable NSFW filters for local use. On one hand, plenty of art contains “NSFW” style content . On the other, plenty of nasty stuff meant to injure people also includes NSFW content. Eshoo largely hones in on the potential to create child pornography but, less mentioned, it’s now trivially easy to train a new subject into a Stable Diffusion model with only 10-20 decent photos. Arguably, the NSFW stuff you could make is the same as the stuff a reasonably skilled person could make in Photoshop but, when this thread started and Nightcafe couldn’t draw a horse correctly, I wouldn’t have thought we’d be at “generate piles of images of nearly any woman with an Instagram account at the press of a button” within six months.
Not sure what can be done at this point: cat’s out of the bag and even if SD added a strict NSFW filter, there’s plenty of people dedicated to breaking/removing it and distributing their version. Likewise, you could train a new non-NSFW model but people have already taken the existing SD model and retrained it to be more porny or to have more anime or more furry content, etc. Plus, of course, not being able to generate legitimate artwork containing nudity. Feels like music piracy or 3D printing guns – once it’s out there, it’s out there.
Yeah, I tend to agree. Plus, porn isn’t the only thing malicious you could do with this software. There’s a Corridor Crew video that shows just how well you can incorporate images of people to put them in various different clothes, makeup, etc.
With impainting, it’s relatively easy to train a model with someone’s face and then just replace the face of another person. Or just put a person in a photo they were never in.
A while back there were a few haunted forest posts. I’ve played around with that background in SD for a bit with nice results. (Especially when I use old styles of photograph.)
Also, there are art-selling websites that show simulations of the artwork hanging in various settings. Apparently some of those were scraped in training because occasionally I get an image laid out just like that.
