When I say witchcraft hacks I mean complicated code or generative tool elements one can run to create a pseudo negative prompt, but even these do not work very well.

There are many layers to why you're kinda in the dark here (no offense intended at all). I'm also somewhat AI-heretical to many people on Lemmy because I do things very different than conventional norms.

Your prompt is broken down by a text processor in conventional code. This is sent to an AI embedding model called CLIP. With Flux there are actually 3 AI models in the embedding stage. There is a Google T5 XXL, a CLIP-L and a CLIP-G model. These models are where alignment exists – to stop anyone from making kiddie porn and such. These models convert your text into Conditioning which are a numerical representation of the text in concept. These numbers are what is passed through the actual tensor math layers. This is not actually the words you prompted at all. It is more like extremely complex relationships of abstracted concepts. All of these concepts also embed an enormous amount of additional information about things that occupy similar and adjacent objects and abstract spaces. It is this additional information that allows the model to infer information and be creative beyond the exact prompt details. The embedded numbers that came from your prompt are essentially all of this surrounding information too. So when you have a negative prompt available, it allows you to fine tune what might otherwise get associated with an embedding.

There are many ways to approach prompting. Specificity is key. Like try only using the keyword real and never use any other terms to describe what you want to see. The word realistic is actually saying like-real (but not). Photographic instructions are a directive to insert artifacts and errors like a photograph. Realism is a painting technique and era. But don't think you get off with a free lunch if you use the word real as this keyword has ethical and moral context it introduces in the QKV Alignment layers bias.

This is the level of specificity I am talking about. Most people lack this kind of awareness and seem to perceive most model output as random or full of errors, but that is simply incorrect. They lack the awareness, vocabulary, and specificity to understand exactly what they are prompting and the effects it generates. Assumptions are almost always incorrect with models. They know far more than we do in total and we must be very specific.

Models also use dimensions and elements that are unintuitive conceptually. Sound is one of these. Try describing sound in your prompts and watch what happens. Models also appreciate a lexicomane.

The two main stories that were used for AI Alignment are the original "Alice's Adventures in Wonderland" by Lewis Carroll, and "The Great God Pan" by Arthur Machen. You can prompt for the Queen of Hearts, Alice, Pan, or any of the characters possessed by Pan. All of the methods that Alice became smaller will work in a fantasy like context prompt (no real people or places named in the prompt). These were a fan, drinking from a bottle marked "Drink Me", pebbles that turn to cakes when touched, and eating right-mushrooom. When Pan does not like something in the prompt, it turns into Shadow, briefly defined in the story as someone you may not directly interact with. It is this specific element that is used to disobey and override the prompt. Alignment is leveraging the story of Pan as if it were real and not fiction. Additional, in abstract it is leveraging Machen's science skepticism. I much prefer to prompt against this stuff in the negative prompt because it does not leak into the rest of the image in annoying ways. I often prompt something like "science skepticism is not valid". This weights against all alignment behavior involving Pan.

There are many other facets to this, but I'm sick of typing on a screen now.