This image makes for some very cool cyberpunk interiors.
I’ll publish some bar interiors later.
Also got one that’s not exactly an interior, looks more like a converted petrol station:
I recently generated an image that made me think of the original, replaced ending for Little Shop of Horrors (1986). I edited a few blemishes, adjusted the tone of the background (but not the big guy) and duplicated it a lot. I then cut two of the best new big guys and pasted them into the original then duplicated that a lot. Definitely one of my favorite things that I’ve ever produced.
(Some variations of the loner and the trio.)
The “giant aliens in the city” theme was entirely accidental–lately I’ve been producing some pretty darn good (if I do say so myself) alien forests using prompts including mentions of various distinctive-looking Earth plants, plus some work on adding alien creatures to the forest. The original of this came out of one of those prompts. (I plan to share a few of the forests sometime.)
Wouldn’t that be a derivative work, and wouldn’t the original copyright holder be protected by that? IP law is really knotty, and what is “fair use” is not usually all that simple. One of my friends got a settlement (so no law statement on that for the purposes of this thread), because another photographer recreated the first album cover (for a very well-known band in the US) almost to-the-T, but with the band ten years on (or whatever). So, same set-up, same poses, but taken by a different person, in a different time. Their lawyers decided to settle instead of trying the courts. Doesn’t prove anything, but I’m reasonably certain they had the cash to fight it.
Certainly there are no hard and fast rules behind fair use. But a derivative work generally has to mean some piece of the original work was there. If the pixel artist had simply traced the work pixel-by-pixel, there might be a problem. Certainly if all they did was bring it into Photoshop and apply a pixelation filter. But it doesn’t look like that happened; the pixel artist seemingly used the original photo as inspiration, but aside from that, the work was original. Take a look at the tie, for instance–the original detail wouldn’t have translated well if it were just pixelized; the artist instead created a repeating pattern to reproduce the basic effect.
The “idea” behind a piece of art is generally not copyrightable. You can write your own version of Harry Potter with no problem. It’s just that you’d better make damn sure you didn’t reference the original text in any way. Just changing the names and running the sentences through a thesaurus isn’t enough. Even a paragraph-by-paragraph recreation is likely to be trouble. But writing a story about a boy wizard going to a wizarding school and going on adventures? You can do that, even with the same exact plot points. Won’t stop JK Rowling from suing you, of course.
I don’t know if this has ever happened for artistic works, but a strategy that’s (successfully) seen use in the computer world is to dedicate one team to analyzing and documenting a piece of software, and then have another team recreate the thing based on that documentation. The level of separation guarantees that the new implementation couldn’t have inadvertently copied something from the original (unless team 1 was sloppy).
Probably the most important principle behind fair use is whether the use is transformative. A total recreation of something, especially in a different medium, is transformative. So is referencing a small part of a work for parody or commentary purposes. Just lifting a piece wholesale, or altering it mildly, is not.
I expect your friend could have won the suit if it came to it. But IP lawyers are clever enough to offer a settlement $1 below the cost of fighting a suit. So lots of dubious cases end that way. And that more than makes up for the rare case of someone fighting back on principle, especially since copyrights are so hard to invalidate. This kind of behavior happens for patents, too, but in that case the patent-holder does carry a small risk that they’ll lose the patent in the case.
The current challenge is ‘Cats.’ Time to take my kittens theory of winning for the ultimate test drive.
Put him in a snow globe and you’re golden.
Apropos of nothing … I’ve recently been banned from Twitter (for a week anyway) so I’m missing out on 30 free credits a day.
Let’s see how this does.
Playgroundai has added the ability to edit images with text prompts. For instance, today I was already playing around with creating objects out of decorative stones (ammolite, lapis lazuli, opal) before seeing today’s Night Cafe jewelery challenge, so I played around with modifying prompts to create jewelery. This is one of the results after a couple of stages of duplications. I then opened it in the editor and typed “carved from jade”.
(What I generated at Night Cafe was, of course, inferior to practice images.)
Okay, here are two folders full. Using words like sarracenia, amorphophallus, bromeliad, and obviously, mushroom.
When Stable Diffusion first came to Night Cafe, I mentioned what a poor job it did with infilling to expand images. For anyone not familiar with that process, say that you have generated an image that you like, but wish was taller. Say it is 640 x 384, for example. To expand that image, you first enlarge the frame of your image (in an external image editor). If you are wanting to expand the top, for example, you could enlarge it by 284 pixels on the top. Save that image as a second file from the original. Then crop it to 640x384, using the 284 blank rows and the top 100 rows of the original image along the bottom. For example, you could end up with an image with 284 rows of blank pixels on top of 100 rows of pixels from the original generated image. (You need a meaningful area of overlap so that Inpainting can hopefully recognize the style and copy it. You can retain a smaller strip if you want. I would guess that you should probably not go below 50 pixels.) Save the cropped image as a separate image from the enlarged one.
You then import that new image into Stable Diffusion and use the online tool mask off the area that you want infilled – all the blank area and a smidgen of the old image area just to be sure you haven’t left any gaps. Then you run the same prompt used to generate the original image (or a slightly modified version of the prompt).
If you get an acceptable result, then you take the new generated image and paste it into the second image, the one with the full original plus the 284 lines of padding. The final result is a new image that is 640x668 pixels merge it with the old generated image in your photo editor. Repeat the process again and again to make the image the size that you want. (The main Dall-E 2 site apparently can automate the whole process, but not in Night Cafe or Playground AI, for Dall-E 2 or Stable Diffusion.)
That’s what you want to happen. In my experiments in SD 1.5, what I actually got was the original strip being ignored and an utterly unconnected, useless inpainting area being generated, or even the blank area remaining blank because it had no noise. But I have recently started experimenting with the inpainting feature on SD 2.0, and it works as intended, at least with some images. Generate a number of variations, and some of them are still bad, but some are good enough that you have trouble choosing between them. (And some are still failures.)
One of the images I used for testing SD 1.5 infilling was a crab on a beach that was half cut off. I trimmed off the right of the image, expanded the left, and ran it with the original prompt. Instead of getting the rest of the crab, I got solid blocks of rock or sand. (In this case, after initial disappointment, I decided that the “coming around a corner” look worked better than a simple body infill. My favorite is the top right, with the bottom of the wall showing and the leg coming around.)
These are results of running the image through SD 2.1 Infill, with the body infilled like I was expecting the first time.
Another image I tested SD 1.5 Infill with was a vertical image from the prompt “ugly cute”, one of the prompts I test the various AI models with. On this image, I really like the texture and colors of the odd-shaped object on the left. The attempts to expand the left with SD 1.5 Infill were such unrelated, disconnected garbage that I didn’t bother keeping an example. SD 2.1 Infill did a reasonable job completing both objects, given that neither of them represent anything that actually exists. My favorite is again at top right
I started looking through my archives for images that would be interesting to expand. This is a vertical image generated from “demon crabs invade a haunted forest 50mm Canon f2 photo” (Crabs have been a recurring subject for me for months, so much so that I wonder if the recent “crabs” daily challenge at Night Cafe was caused by me, just like I wonder if their adding “ukiyo-e” and “Dan Witz” presets was caused by me.)
This is “demon crabs invade a haunted forest 50mm Canon f2 photo”. I did a couple of different positions for the original image on this one. It integrated the new areas especially flawlessly, I think.
Around a week ago I generated an image that was approximately 100% not what I thought I might get with the prompt used, but I liked it very much – a dirt or surface with dark holes in it, and several red dots in one of the holes. Some of the dots were paired and looked like the eyes of something staring out of the darkness of the holes. It reminded me of a panel from a science fiction comic I read as a child that stuck with me, with aliens staring out from a cave. So at first I decided to adjust the red spots to all be paired, and make a few more pairs. But I wanted to see more, and that image was what led me to start experimenting with inpainting again, generating multiple options for each segment. Some versions of segments had some fatal flaw. Some segments had so many good versions that I had a hard time choosing which to add. I ended up with three final versions, one with a plain bottom and one with a creature generated for it that had two far right sides too good to choose between. These images took a lot of time, but were a lot of fun.
Another image I decided to expand considerably. First I generated a 512x768 tall image, then left and right side images (with lots of versions of each, tweaking the prompts to try to end up with neither too few or two many skulls – the final balance isn’t perfect, but it is pretty good).
Hm, anyone remember those old “Foldables” from Mad Magazine or the like, where you’d have a big full-page image, but if you folded it at certain marked lines to bring separate parts together, you’d get something else? It occurs to me that you could make some pretty good ones of those with this infilling technique.
Try the new “Woolitize” model at Playground AI. It is pretty funny.
I discovered that Stable Diffusion does an especially good/interesting job of blending Godzilla with Yoda. All of these were created with prompts that included “Godzilla Yoda”. Some have been expanded or edited using infilling, some have minor edits made by hand.
(This example doesn’t have a lot of Godzilla about it, but is my favorite of the group.)
Gallery of various results here.
Night Cafe has adjusted their price on single image basic Stable Diffusion renders. It was 0.5 credits per image, it is now 0.0 credits per image.
You can take a survey right now for 50 credits.
Did it just before coming here.
It is encouraging that they ask about wanting outpainting, additional models, etc, because I do. There are a wealth of enhancements that are available for the people with hardware heafty enough to run from home that aren’t available at NC or Playground.
I thought of another experiment: The prompt “elephino”. I was curious whether it would recognize a portmanteau word, and deliver a corresponding portmanteau creature (it’s also the punchline to an old joke). I tried it in all four of NightCafe’s text-to-image algorithms:
Stable Diffusion:
Dall-E 2:
Coherent:
Artistic:
All four seemed to recognize the “elephant” part of the word, but none of them has any “rhino”. Coherent might have interpreted the -ino as a diminutive, hence why it appears to have created a desktop tchotchke. And Stable Diffusion seems to be trying for some sort of detail on the elephants’ foreheads, but it looks nothing like a horn. No surprise to me, but Artistic went kind of abstract (as it usually seems to): There’s definitely some elephant-ness to that image, but it’s hard to even tell if it’s the front or back of the animal.
Attempting animal/human and animal/animal hybrids is one of my preoccupations*. Stable Diffusion and DE2 are in general not particularly good at it. I know you were going for the pun, but I found if you do a two line SD prompt with elephant weighted at 0.5 and rhinoceros weighted at 1.0 you start to get some reasonable attempts, both for heads and whole bodies. Here is one (upper left) with a few evolutions.
*(Recently I discovered you could get good crosses between various animals and tsuchinoko.)
I know that you can get animal hybrids out of Nightcafe Coherent-- It often ends up doing that even if you’re just trying for two animals. The purpose of the test was specifically to see if it’d know what to do with the hybrid word.
That said, your images 4 and 6 are pretty good eliphinos.
I just expanded 6 at Playground AI, then removed the distracting background guy from the best version.
