Most image prompts collapse into the same hazy Midjourney-default look — backlit, slightly painterly, vaguely "epic." That's not the model failing. That's the model doing what every other prompt in its training data is doing.
The fix is specifying things the average prompt doesn't.
What to actually name
- Subject with one or two distinguishing details. Not "a woman" — "a 32-year-old woman in a tailored navy linen shirt."
- Camera + lens. "Hasselblad X2D, 85mm at f/1.8" produces a different image than "shallow depth of field" because it pulls from a different cluster of training images.
- Lighting source + quality. "Overcast soft-box key with a warm rim" beats "good lighting." Be specific about where the light comes from and how hard it is.
- Location + time. "Rooftop in Lagos just after sunset" gives the model an entire mood to work from. "Outdoors" gives it nothing.
- Colour grade. "Warm amber grade" or "cool teal-and-orange" is the difference between something that looks like a photo and something that looks like a render.
- Publication. "Shot for The New Yorker" or "Vogue editorial" pulls cleaner style than "professional" or "high quality." It tells the model what kind of image actually gets published in that style.
Drop the filler
Stop appending masterpiece, 8K, highly detailed, ultra-realistic, trending on artstation. These tokens added quality in older models; in current ones they push toward the same generic look you were trying to escape. Specify the actual look you want and trust the model to render it well.
If you can read your image prompt aloud and someone could sketch a thumbnail, you've specified enough. If they can't, you haven't.