Hi everyone. I apologize for the post, and my poor English.
Before going to bed, I decided to generate some images and got curious about a particular token. I’ve not seen articles on this topic and wanted to share my experience with the community.
I used this prompt:
<girl, full height, far plan, side view, unlimited ocean of emotions, The Girl’s Incomprehensible Face, he pulls his hand to the left>
The result often came out as concept art. Intrigued, I tried removing some tokens to see which one or combination caused this effect. Initially, I suspected "far plan" but even with it off, the concept art style appeared randomly in about 60-70% of cases.
The token responsible turned out to be <The Girl's Incomprehensible Face>. This token consistently generated concept art. You can tweak it slightly, like using "boy, The Boy's Incomprehensible Face," and it effectively changes the character's gender during generation.
Character descriptions and Lora characters are well-represented. Occasionally, details differ between parts of the concept art, but overall, the characters remain consistent across different poses and spreads.
I initially thought this only worked with this checkpoint: https://civitai.com/models/6755/cetus-mix. However, I achieved similar results with other models too.
I think it's worth experimenting with. It would be great if more people could test this token with complex prompts and different text inversions to see how it performs. (I'm trying not to seem lazy, (but I really want to sleep)). Essentially as strong a token as - "abstract"
Overall, it's an interesting token, especially useful for training and pre-training, Lora character concept art. What do you think? I hope someone else finds this interesting too.
If interested in generational details, ask in the comments😉