Raw LLM Responses

The AI does not take 'data' from the 'dataset' to 'replicate' the art. Stable Diffusion, is 2gb. It does not contain any data, from the dataset it learnt on, that isnt how neural networks work at all. Diffusion, takes an image, and slowly adds progressively more noise to it over time. The Neural Net then attempts to do this in reverse, taking the noise and making it look like the image. What then gets stored in the 'weights' of the neural network, is not the image, nor the data of the image. What gets 'recorded' or 'saved' is the process of taking noise, and turning it into an image. This is know as the latent space. You then have a text encoder, which shares this latent space. Which looks at the 'tags' or 'text descriptions' of the image its training on, and also encodes that into the latent space. This creates a connection in the latent space, between text and in image. It will then be able to take just text, and inject that into the latent space, along with an image of total random noise. The neural network, will then attempt to do what it learnt, and give you an image that draws from the weights related to the text you entered, and uses the noise image to slowly turn it into an image. Neural Networks, are designed with biological neural networks as inspiration. Aka, our brains. Sure they dont work like a human brain, but they work a lot more like a human brain than 'photoblending between some images they store, or whatever clueless idea you have.