>>3030697Illustrious is trained on top of SDXL and retains some understanding of words not in Danbooru tags. And txt2img is a very fuzzy thing, it's not like diffusion will suddenly fail if your prompt contains an invalid tag. For example, colors usually work for objects even if there's no Danbooru tag for that specific combination. I got pic related even though "maroon bikini" and "moon necklace" are not real tags. Also you can mispell things and still get what you expect most of the time. You'll get better and more predictable results with tags that actually exist but it's worth experimenting a bit if you need something that's not explicitly tagged.
All that said, stuff like "visual harmony through warm tones" and "ambient light interplay" usually means the prompt was written by an LLM. I don't know if any model would actually do anything with that vague bullshit, apart from adding some randomness to the gen.
