They are generated. The whole image is generated. Hence the term "AI generated image." It's never just copy-pasted from a real photo.

Even when you use the name of a celebrity, that face is generated... to look similar to that celebrity. But it's still generated.

Thanks for the update! I get it. It's a lot of work, and a balancing act for timing things as well. It's hard to live up to (a minor group of) people's expectations. But knowing what the state of play is with an update like this really helps 👍

@perchance@lemmy.world

There will always be oddities like this that crop up. On other models if it's not this, it'll be other things.

Because AI text models don't have any kind of internal consistency going on. They just generate a next word they think a human might put there.

The reason that AI chat is so good isn't because the AI is really clever, and remembers things and imagines a real world as a human would. It's because every time it generates the next part of the story or the next response, the creator of the generator page you're using is reminding it of all that stuff over and over again.

So that helps it keep on track and make sense. But it doesn't really know what it's doing whatsoever.

When you see weird stuff like this, it's not a "bug" with perchance, or a "bug" with the AI text model... it simply comes from the fact that AIs are not actually thinking at all to begin with. And we need to cajole it to look sane with all sorts of tricks behind the scenes, even to get it to be this good. There will still be funny things they do even with that.

I notice the code has [x = base.evaluateItem] which will naturally output base.evaluateItem, which is what you are seeing. You can avoid this to also give it what to output after a comma: [x = base.evaluateItem, ""] Or you could even combine the block with the next block like this: [x = base.evaluateItem, imageLayerCombiner(data)] so that it will set the variable and output just the layer thingy.

1. I feel like this makes some sort of sense. You're saying "no underwear" I guess?
2. The engine isn't infallible, no engine is. It guesses based on various heuristics whether something is nsfw, and goes with that. Sometimes sfw stuff is marked as nsfw, something nsfw stuff is marked as sfw. Maybe this will improve if the engine changes, but for now it's about par for the course.

Maybe there's some change that can be made in these areas, but just my guess is there's not a lot they can do about it.

I have a feeling Brave fails to pass some checks regarding this. People have had trouble in the past with that. So maybe try a different browser, see if that works.

I've never seen it myself, and I've experimented a lot with locked-in seeds and such too.

So is this something you've generated, or is that in the gallery? Either way, seems like it's displaying incorrectly of course. I don't see why there would be something you should do about it, if there's some bug in the system causing this.

Other than maybe bring it to the attention of @perchance@lemmy.world--which you've now done ;P

You should use the dev tools in your browser to figure this out. In the sources tab of dev tools, turn on "break on caught error" and see if you can get to the line that looks like that one it printed out in that message. That'll help you figure out what code is breaking.

Presumably it's in the dynamic importer plugin, so you could just look in that plugin's code on the site, and see where that is, and find out what's gone wrong that way.

Cool.....? I guess you should add a link to this post so people can just go look at it, instead of just look at an image ;p Also, "screenshot-plugin" seems to not actually exist so the instructions in the image will not work.

Have you renamed it to "simple-screenshot-plugin" perhaps? https://perchance.org/simple-screenshot-plugin

Could I make some suggestions for changes to how this works? I could make my own, but since you've put the effort into making this in the first place I don't want to derail that if I can help it.

Allow passing a full query selector / element reference to the function.

Have functions to call to just render the canvas and get that canvas. Another to render and get a data url. Another to render and download.

And possibly one to download a data url image, and another to download a canvas image.

Doesn't need any more coding really, just moving some code around. And makes what you could use the plugin for waaaaay more versatile.

Cool... What you could do is make a little bookmarklet that helps people do this, like I did for resetting the nsfw block. Like, takes you to the right page. If you're on the right page, asks if you want to export or import. For export, gives you a prompt with the JSON. For import, gives you a prompt to paste the JSON into. Something like that.

All ai generators on the site use the same "model" and engine to generate them. All this generator is really doing is adding something to the prompt and negative prompt.

If you click (i) on a generated image you can see the prompts actually used by the generator. Then copy them out into another generator (eg. https://perchance.org/ai-text-to-image-generator) and use "no style" and it would give you similar results.

@perchance@lemmy.world

Use a different emotion in the prompt I guess?

Something to remember though is that all words affect the generated image. So unless you've actually locked in how the character looks with other words in the prompt, using a different emotion word may change some aspect of how the character looks as well.

You could also use the BREAK keyword, which cuts the string of meaning. So then you can have one word relating to one part of the prompt but not other parts of the prompt. The classic example is blue dress yellow hat can produce various elements in the image being blue or yellow... vs blue dress BREAK yellow hat which would make only the dress blue, and only the hat yellow. (In theory.)

This would be good.

Interesting though... I've generated pictures of dragons that don't look like the furry style. Have you actually got the word "anime" in there if you want it to have the anime style?

You can also just not use the update at all. (People can still call it on the other side, but you don't need to use that as the mechanism by which you change stuff.) You can just use regular JS and set it up to work however you want it to work.

I'm not even sure what an app would be that the site is not. What are your goals for making an "app" for perchance?

Woooooooooah, that's a lot of negative prompts! I'm no expert but generally, people advise less is more. Essentially you're pulling toward and away from all sorts of directions that aren't necessarily necessary for the images that are being generated. So it can get in a tangle.

The way I think about prompting is, I'm building up what I want for that image. So rather than using the same massive negative prompt for everything by default, I start from scratch, add negatives for things I want to avoid that seem to crop up for the prompt I'm working on specifically. Things like that.

That said, I still get generated images that look pretty reasonable based on the main prompt itself. Must've been a freak accident 🤣

Great--I'm so happy! :D

I'll bookmark this comment and come back to let you know if I spot more logs that should be .debug()

You could even do something like localStorage[this.id]=this.value. Not sure if that would work for getting the value though... 🤔

Quick question on this topic... is localStorage "safe"? Does it only store for that specific generator (like, localStorage's behavior is overridden for that), or can it pick things up and change things from other generators?