Raw LLM Responses

The elephant in the room: Controlnets. Prompting has been essentially obsolete for the past 2 years. We went from essentially asking a drunk genie to do all the work for you to having more precision and intent than a brush. You can now control every hair on their head. There are controlnets for lighting, composition, poses, 3D depth, textures, colour, styles, brush strokes, latent coupling, composable LoRas, etc. and you can either feed in existing images to extract one (with no way to really maintain copyright or identify the author of the image fed in, especially once you use multiple interacting ones with varying strengths and noising), or you can manually control it (such as moving a rigged model's bones around the way a 3D artist does), or both. I doubt 60 year old bureaucrats even know what a ControlNet is. Kinda important, because this is the dominant way of generating images and video now, and it's pretty hard to legally specify or even vaguely hint how much prompting and the latent space affects a ControlNet. We're no longer feeding text into a neural network, we're feeding neural networks into the neural network, and this massively overcomplicates this issue, as well as makes everything they've said so far completely redundant. Obviously, they would have to overwhelmingly side with this being human authorship, and this will certainly become a precedent when it hits the courts, but if you reflect for even a moment on why this is the case, you realize it undermines the notion that prompted outputs belong to the neural network by revealing some uncomfortable truths about how AI generation actually works that people would rather remain ignorant of.