It does not need to be sentient, the data contains the patterns and the rules that are extracted.
Those arrangements of pixels are not random, concepts and variables are also not random.
There is a signal in all of that not just noise.
AI created anime girls on pixAI are brutally good
- Thread starter zilkin
- Start date
Those arrangements of pixels are not random, concepts and variables are also not random.
There is a signal in all of that not just noise.
- May 3, 2018
- 93
- 183
I didn't say it was random. I said what I said.
Diffusion models work off of a predictive process similar to large-language models, where the AI uses randomly generated noise to generate a unique image based off of predictive patterns it establishes using CLIP language from the prompt to determine what cluster of the latent space to pull the likely pixel patterns from. But it does not know any "rules". None at all.
It's pattern recognition writ large. It's amazing, but it's a fallacy to think it "understands" what it's doing.
I have hundreds of hours of experience using it at this point, and trust me, it would be a lot easier if it understood rules. But it doesn't even know hands have five fingers - because they DON'T always have five fingers in the training data, depending on the perspective of the picture, whether it is a cartoon drawing with only 4 fingers, etc. Think about a hand salute. From some angles it looks like one thumb and one finger - the AI only understand 2D planes. It does not have an understanding of 3D space for it's compositions - it only looks like it does due to good training data.
It's important for people to really understand how it works, because misinformation is what gets artists up in arms about the AI "copying" them.
- Sep 3, 2018
- 281
- 526
That already contains the "rules".
Pattern recognition and rules are the same thing. Patterns without rules are not patterns, it is chaos, it is randomness, it is noise.
It understands the rules within the data AND ONLY within the data.
Yes it does not have cognition, it does not have an external filter to judge things by, it cannot use a Simulation to work out what is wrong.
Which is why you get those kind of errors, but having errors does not mean having no rules.
- May 3, 2018
- 93
- 183
I honestly don't get what point you keep trying to make.
Your first post I was replying to was you saying the AI had "internalized the rules" of "cloth dynamics", "fluids", etc. like some kind of human artist.
You are saying "rules" as in rules a human artist would follow.
Diffusion models do not operate by any art, anatomy, lighting, or drawing rules. They operate by matching CLIP language contained in prompts to latent clusters in their model. The latent space in the model is derived from dataset training that creates mathematical associations.
And, by DEFINITION, diffusion models DO work on patterns WITHOUT rules. They use random noise as an ESSENTIAL part of the process. What they make out of that noise is determined by predictive probability of what pixels will appear next to others, as determined by the current part of the latent space cluster the prompt has directed the model to.
Again, this is completely false.
Please, please, please watch a video on how diffusion models work. You are HURTING AI acceptance among artists and giving fuel to AI skeptics yourself when you misrepresent HOW the AI is actually working.
It's not like a human artists but they do internalize them.
And those mathematical associations represent?
They are not the pixels by themselves.
Why do you think it can generate images that we get in terms of the lighting, materials and shadows?
If it were the pixels or shapes by themselves you would expect the light and shadow to be all over the place, yet what we get is surprisingly consistent.
After all every image in the training data has its own completely different light setup that you cannot copy part of the image and paste it in another image.
It's precisely because it can abstract and generalize the lighting of a scene.
The fact that you can change the lighting with a Prompt should tell you as much.
If it does it for lighting it does it for all kinds of things.
People are to obsessed about how the algorithm works when we are long past that in terms of what is important, it is not the Algorithm it is the Data, the GPT models should tell you as much.
There is no fundamental difference in the algorithm between GPT 3, GPT 4 and GPT5, what is different is the sheer amount of Training Data and Computation.
Last edited:
- May 3, 2018
- 93
- 183
Seriously. I've told you why.
Look up a video on diffusion models. I'll make it easy for you.
You must be registered to see the links
that does a very good job of breaking it down simply. And
You must be registered to see the links
explains very well how AI only processes things as 2D.The reason they can give you a consistent image is same reason the GPT LLMs can give you consistent text and answers. It's predictive completion.
Yes. Exactly. The diffusion models DO work like the algorithms used in the GPT LLMs, and good training data DOES make a big difference.
And the algorithm is and always will be important because it tells the models how to process that data.
I'm excited about AI too, I've already told you I'm an enthusiast. But you can't go around misrepresenting how they work. How the DO work is already impressive.
- Dec 9, 2020
- 135
- 106
- Dec 9, 2020
- 135
- 106
The face and talking animations look very good... now if it could also animate the bodies... that would make for some real hentai if you know what I mean.
