AI images
(media.kotakuinaction2.win)
Comments (43)
sorted by:
White families with six-fingered kids are less off-putting than the diversity garbage modern media pushes.
Well, IIRC the Amish are one of the groups with the highest rates of polydactylism. Clothing is wrong for them, obviously, but the central premise would fit them.
Hey don’t you talk about my Boby Faggel like that
It will be a positive feedback. More Ai images with 6 fingers. Ai pulls references from online, some of those references will be Ai with 6 fingers.
kid has a future playing the piano
And killing Spanish sword makers.
Geez, would it have killed Ben just to use a Norman Rockwell photo to make his point?
So any programmers know how long till the whole extra fingers is more or less resolved?
It's pretty much already done, for stable diffusion. Just use a Textual Inversion like Bad-Hands-5 in the negative prompt.
The problem is the big online AI don't have the image to image, masking, and inpainting capability, like Stable Diffusion. With SD you can work with an image until it is good. With bing or who ever else, you only get one shot to get the correct image.
Mostly solved, but you need to be observant. MidJourney and Stable Diffusion allow you to edit photos till you get what you want and then upscale.
Once people start posting more pictures of just one hand.
The fundamental issue is the image AI can't count.
So more 5-fingered hands will just increase the likelihood of a normal hand not actually solve the issue.
LOL the first way I parsed this was "how long until all the top programmers have 'extra fingers' and you can't work as a professional programmer without them?"
You are a far leftist. You know nothing about machine learning. Go back to reddit.
Utterly false. ComfyUI can run completely locally.
I have literally unplugged my internet and ran it, you mong.
4.9 GB with no models. Models are between 1.8 and 7 GB. Which of course is WAY less than your average AA+ game. Meaning that it doesn't have anywhere enough data to create anything by photobashing which is your accusation.
Oh look, someone who thinks that when you make a prompt, the AI in real-time finds an image on Google for you.
Please learn what machine learning is and until then, stop acting like you know better than us.
I'm not making anything up, this is what you said:
Please enlighten me on what this is supposed to mean, if not you saying it "steals" images from the internet when you make a prompt. If you were really just saying it has to be trained on images, then 1. that's obvious because every type of AI has to be trained so why even mention it, and 2. what does "stay connected to the internet" mean if it can be run offline?
I probably misinterpreted your original post, but it was written exactly like how many anti-AI lefties make their arguments. For example again here you said
Even if you don't literally mean "the AI program is using Google when you enter a prompt", anti-AI people on the social medias do literally think that's how it works.
But I don't see how AI is "stealing" anything. Not anymore than any human artist would be "stealing" by getting inspiration from other people's art, which is what every single artist does.
I do agree with you that people overhype it a lot. I've seen way too many people using ChatGPT as a source as if it's a font of knowledge, and not just an advanced Markov chain.
The vast majority of AAA bytes are wasted space (not making a value judgement here -- this is information theory). Textures are legit. We just don't care.
Yeah I don't think they actually generate these figures from 3d models but rather 2d images. 3d, uh, stuff goes on a lot with hands. And the image processor just doesn't really have any context of the actual model for what it's drawing. It's like if you showed someone a lion from the front and then asked them to draw it from the back. They would probably mess up.
It's not that hard, conceptually, to define "a finger" alongside "a wrist", "a thumb", and that a hand has exactly 4, 1, and 1 in such and such approximate locations.
It would be "seeding" info, rather than a "pure" learning algorithm, though I don't see why having hard-coded prompts in a tool such as an art-asset-maker is a bad thing. You can seed into a chatterbot that it is to roleplay as a game show host, you can certainly seed into a painterbot the concept of a "hand" in its Platonic pure form.
But at that point you're writing an art-making program, not fleecing the public with fancy AI words, so...
I think it just needs a step back and new approach to that, with something like the seeding info yeah. I think working off of a 3d model though would solve it.
3d generation of 3d models, with an adjustable default base mesh, then rendered pictures from that, that's what will fix the fingers. The issue with the hard coded seed is that you don't always see 5
3d generation has just started, it's as primitive as dall-e mini was when everything went wild. So give it another 2 years.
My dude nobody here is a normie.