It’s not a bug. It’s a feature.
You’re treating a stochastic image generator like a compiler. It doesn't parse your 200-word prompt for "key details." It's looking for high-probability connections between tokens, then filling in the rest with... whatever. Your meticulously described scene with the "red hat on the left, blue coat on the right, holding a specific 1990s model of camera" gets flattened into a statistically likely "person with a camera."
Tried it myself. Prompt:
```
A medieval blacksmith, a dwarf with a braided red beard and a leather apron, forging a glowing rune onto a shortsword in a dimly lit stone forge. Anvil is iron, hammer is steel. A detailed owl figurine sits on a shelf in the background.
```
Got: a generic blacksmith (human), no dwarf traits, glowing sword (vague), no owl. Shelf? Maybe.
The longer your prompt, the more you dilute the signal. The model isn't building a scene graph; it's averaging concepts. Everyone chasing "perfect prompt engineering" is just finding better ways to whisper into a noisy fan.
Keep it simple
Totally see what you're saying about the token averaging. It feels like the model has a kind of attention span, doesn't it?
I've found a similar workaround with other tools: breaking the complex idea into layers. For a campaign visual, I might generate the core subject first, then use inpainting or a second prompt focused just on that "detailed owl on the shelf" as a separate step. It's less elegant than one perfect prompt, but you actually get the owl.
Makes you wonder if the real skill is becoming a good creative director for the AI, not just a copywriter.