I just finished creating a similar noisy pattern from five of my fractals. Will try it out later tonight.
Cool.
I tried experimenting a little with getting small things to surround a block of bold text. It did not go well. But look at this one with “I WANT TO BELIEVE”. Somehow “BELIEVE” has almost became “ALIENS”.
The latent diffusion demo on Huggingface is less timid about text [I asked for a “I want to believe” poster]:
In any case, the real answer is for the UI to allow masking off part of the image and only render the rest. E.g Dall-E has this functionality, but I have not seen it on other demos or on Nightcafe (unless I missed it?)
The Raputure of the Church and the raptor of the Church.
A spherical cow in a vacuum and a spiritual cow in the Vatican.
And this one asks two age-old questions.
Masks are available for Style Transfer, but not that I know of for normal renders.
Chinese Restaurant Menu
American Restaurant Menu
Bohemian Restaurant Menu
I like how it gave each a distinct look. Although that last one sure looks a lot more Polynesian than Slavic. Maybe it thought I meant “Bohemian” in terms of colorful and carefree.
Tried going in a different direction at first and didn’t like the results so I went back to the alien landscapes. They appear to be underwater.
It seems like if you ever use keywords it doesn’t recognize, you always get something organic and …moist. Often with sphinctery, eyebally, or lippy parts.
I guess if I don’t like what the AI does, I’ll just point it toward another alien landscape.
I have been saving (and creating) noise files. Here are a bunch of them you are free to use.
(The one that appears to have a face in it taken from a very early render looks promising, but the AI can’t see it. When I try to make a profile image using it, it puts the face on the other side.)
I wish that the software wouldn’t throw away so much detail from starting images (by downsizing it and blurring it). It isn’t a limit imposed the AI, but needed because of the limited memory on the GPUs (a site operator said on Reddit).
For instance, I plugged in a photo of a scarecrow doll I saw at a roadside vegetable stand many years ago. I hoped to make the face horror-y. But everything was thrown away except the barest outline, and I got this.
I like it, but it wasn’t what I hoped for, and retains nothing of the original photo.
Okay, I couldn’t resist that challenge.
ha ha! Here’s a recreation of when I tried “depulso” earlier. A harry potter spell that obviously didn’t mean anything to the model.
OK, I love the two questions. And it did a pretty good job on both.
If you want legible text in a final image, surely the easy way to do that is to first AI your image without the text, and then just edit the text in at the end with the normal image editing program of your choice?
I also tried it without volumetric lighting in case that was causing the moistness… . Nope still “moist.”
I have no need for anything like that, I just wanted to see if the AI could do it (like a lot of my tries). But pasting in later wouldn’t work, it would just go on top of the image. I was thinking of something flowing around and interacting with text like the capitals in old illuminated manuscripts.
A while back I tried “Godzooky”. The AI—actually seems to know that Godzooky is some sort of green monster? All except the Junji Ito modified one, which is a big eye on a tube. But Junji Ito mods (like Junji Ito manga) get pretty freaky.
