This is Part Two of my on-going series about prompt adherence, with a little more discussion about why I think this really matters. Today, a lightly NSFW image, that I found while trying to understand how to get hairy chests in Flux.

The Image

And then, the prompt

The image is a high-resolution photograph of a middle-aged Caucasian man standing against a plain, light gray background. He has a robust build, with a muscular torso and a slightly rounded belly, and is fully nude except for a white jockstrap that he is pulling down with his right hand. His skin is a light tan with a few freckles, and he has a thick, white beard and a bald head, with a few wisps of hair on his head. His left arm is partially visible, showing a tattoo of a tribal design on his upper arm. The man’s expression is serious, with a slightly furrowed brow and a direct gaze into the camera. The lighting is soft but highlights the natural texture of his skin and the coarse hair on his chest and arms. The overall mood of the image is raw and unfiltered, capturing a moment of vulnerability and confidence simultaneously. The photograph focuses entirely on the man, with no other objects or people in the frame, emphasizing his presence and the strength of his physique.

Once more, I asked Llama-3.2 to generate a list of concepts in the prompt, and it came up with 15. I did the not present (“-“), present, but with issues (“=”), or present (“+”). Flux did a lot better here. Out of the 15, 9 (60%) Flux got right, up by 46 percentage points. Now, you migth ask, “why is this in “Flux ignores everything, if it got over 50% right, and most of the others are little things?”

It’s actually those little things that got it here. The generator got so much right that it’s weird that it missed these things. I mean, I can make up “just so” stories about each omission, but there is no connective thread that explains all of them together. Most of these are also concrete things that are either there or not there, so it’s not a mattter of interpretation.

This is what bothers me about AI in general. We don’t know why it works and we really don’t know why it doesn’t work. It is a situations like this, where there are simple, stupid things that it got wrong even in the context of so many other things that it got right that this challenge really becomes apparent. Needless to say, I don’t know enough about how models really work to contribute anything beyond the observation.

Concept List

    • middle-aged Caucasian man
    • robust build
    • muscular torso
    • slightly rounded belly
  • = white jockstrap
    • light tan skin
    • freckles
  • = thick, white beard
  • = bald head
    • tribal tattoo on left arm
    • serious expression
    • furrowed brow
    • direct gaze into the camera
    • raw and unfiltered mood
  • = vulnerability and confidence