I think it's better. If the software was assuming too much, it would be more prone to draw something else than what it is asked. Clarifying assumption through text sounds a more efficient way to do things, until (or unless) you're working through a workflow where a full-fledged LLM interprets the scenes and adds the assumptions (but let the user see the exact query to the image generation model, so he can modify it if he wants to depart from the assumptions.
The same prompt with a mention on where she's looking at and rewriting the part where the hand is about to catch the child gives images like these ones:
Some cases of bad fingers, but the model has no trouble correcting it.
I think it's less prevalent in this model than what you say. The duchess was middle-aged, the Whisteria Lane resident wasn't particularly young-looking, the woman carrying the grocery store was on the elderly side, ... the necromancy teacher was explicitely described as in her fifties so it doesn't really count, and the young ones in many case had descriptions that veered toward a younger woman. Here we have an adventurer (not a profession one does for long before either retiring or dying), before we had a princess (and not a queen) and a schoolgirl. The only one that came out young without anything pointing to it was the elf mage. I'd say it's a slight tendancy, but not something that can't be overcome. It's a case where the model "assumes" something without being prompted (ie, the user prefers to see a young woman rather than a middle-aged woman).
A behaviour that is certainly exacerbated by RLHF (reinforcement learning by human feedback) post training of the model: images are pitted against each other and a real life human selects the one he likes better. I suspect young women wins all the time...
Last edited:
