Raw LLM Responses

@XetXetable Sorry about the dump I mainly write for myself as a way to collect thoughts and improve arguments. Sometimes I hit send. TLDR I suspect we hold very different standards for what level of control of the result communicates intent. Controlnets are nowhere close. Future controls may come. But with diffusers being like they are I don't see it. Controlnets are very specific to one aspect of the image like openpose or they're very underspecified in what aspects you control like canny. They leave a LOT of room for the AI to make decisions for you which is what we don't want when we make the case for artists intent. Control is fundamentally opposed to diffusers. Other systems may appear that focus on augmenting workflows and those have better hope of preserving intent. It's very hard to tell what's gonna happen there. It's also very hard to show that the artist is intending what they've created with AI tools. We really should be sceptical of this because unlike conventional art where the proof is in the countless strokes the artist drew. With AI there's nothing to say this is the case. Testing intent is hard. But consider that any intent communicated through controlnets or prompting *must* minimize the impact of a seed. Otherwise you _found_ what you liked, the randomness decided what the result was. Controlnets take some of that away. That's obvious from results. But the diversity remains very obviously. Fixed seeds (like enforcing someone use pi, a single valid seed) also helps that to some extent but still the prompting is just very variable. I don't know you guys experience with these tools but picking an image in your mind and reproducing it exactly like envisioned is incredibly hard. And for someone that's good at both tasks I'm confident it's just easier to draw it. Maybe one day there's a framework for much more advanced control features that includes very specific things like what stroke width do you have for some elements of an animated characters, how are they shaded etc. Nobody uses language with the precision that AI tools could actually have learnt that through prompting. Prompts are a very inconvenient way of configuring these if you wish to have the high standard for intent/control we're talking about. CLIP, which sits in front of SD is in part a language model. It takes in the words in the order you've specified and that's very important. Even if you don't do weighting based on where in the token string like most frontends do (weighting later words less usually). So when you want to change the brightness or hue of the light coming from the left side of the scene you're making. If you touch the prompt to achieve that change you've entered into a very complicated space becuase any change you make changes every aspect of the image to some extent. And the diffuser only landed on the image you're looking at through rather subtle things it saw in the seed noise. Think about just how little actual final image there is in sample 0 -> 1. That step dictates your image the most and it's based on your prompt. Stability is lacking. Someone is going to critique this as being too constrained and narrow minded about what AI art is "allowed" to do. But when you argue that people who use AI to generate art are artists their intent is vital and how well they control the results is much more important than the actual quality of the results.