Raw LLM Responses

​@AxelTracer Diffusion models do not directly turn the vector into an image, the idea is that a noise texture gets slowly unnoised with the representation vector being "context" that perturbs the denoising. That's an implementation detail though. What is important is what it actually means to represent data as points in a vector space. When we have a huge dataset of art and we create a model that represents that art in a vector space, what eventually happens with a powerful enough model and enough art is that, loosely speaking to account for the actual data living on a lower dimensional manifold in that vector space, the basis becomes an abstract conceptual space where each direction corresponds to some high level artistic concept. It's because of this, that you can interpolate in the latent space of the image model to create new pieces. With a pure image model, you are limited to quite uninspired images that mix and match the underlying concept in some mash up of images used as inspiration. When human beings do this, when there are inspirations are incestuous and they're just only inspired by other art to make art, generally we call them hacks. AI slop and human slop are quite aligned in that way. The innovation is the combination of a language model with an art model and a shared latent space for speech and art. Now - you express ideas that are not in any of the art pieces but are present in the language the model trained on and there is an option to extrapolate in the latent space of art to ideas that havent been made into images before. That's extremely powerful. That's when you as a human can express a novel idea, where the prompt itself is interesting, and that can get turned into an image. The actual model architecture around it isn't that interesting really, what's important is the intuition for what it actually means to have a high dimensional abstract vector space of concepts. There is a hand tuned version of this that predates AI for what it's worth and it's called picrew with sliders, to a large extent the AI creates the universal picrew generator with sliders that cover the space of all art scraped and all things in between. In modern art galleries, for conceptual art pieces, it is common practice for the artist to send not a finished piece but assembly instructions and companion text for the work. The value of the art is not in that it's I don't know, a banana stapled to a wall, what the piece lives or dies by is the context and ideas behind it. I don't think we've seen many people who aren't just using these models to create more "BIG BOOBY TRENDING ON ART STATION HOT ANIME MANGA" or whatever, it's hard to know how interesting the AI slop can get if you try to make it interesting and novel. Ultimately, if you buy conceptual art as a thing that can have value, then it would have just as much value if it was a prompt to a model instead of instructions for assembly executed by gallery staff. Even then, it has to be acknowledge if the people prooooompting for big booby anime girls had technical ability what they ended up making would still be slop, because making generically hot anime girls without any context isn't artistically interesting beyond expressing technical ability.