Why your AI images look generic, and seven prompt habits that fix it

Most disappointing AI images aren’t the model’s fault. They come from prompts that describe what is in the picture but say nothing about how it should look. Models fill that silence with their most average guess, and average is exactly what “generic” means.

The habits below come from comparing a large number of prompts with the images people actually wanted. None of them need special syntax, and all of them work across Midjourney, Stable Diffusion, Flux and Gemini.

1. Describe the light first

Light decides mood more than any other word in your prompt. “Overcast noon” and “low sun through fog” turn the same street into two different stories. Name the source of the light, its direction and its quality: hard or soft, warm or cool.

• Before: a woman at a tram stop, beautiful, detailed, 8k
• After: a woman at a tram stop at dusk, neon shop signs reflecting off wet pavement, soft rim light on her hood

The second prompt is barely longer, but every added word carries information the model can use.

2. Name a lens, not “high quality”

Quality words are noise; photographic vocabulary is signal. “35mm, shallow depth of field” implies intimacy and a blurred background. “16mm from a low angle” implies scale and drama. Models learned these associations from millions of captioned photos, so a lens and an angle do more than any number of “ultra-detailed” tags.

3. Commit to one style

Stacking “watercolor, cyberpunk, Studio Ghibli, photorealistic” forces the model to average four looks into mush. Pick one primary style and, at most, one modifier. If you want to explore, generate several versions that each commit to a single style, then compare.

4. Use colour words that carry mood

“Colorful” is vague. “Teal and amber” reads as cinematic; “oatmeal and clay” reads as calm and handmade. Two named colours are usually enough to steer a whole palette, and they are far more reliable than adjectives such as “vibrant”.

5. Say where the camera is

Eye level, overhead, low angle, over the shoulder, from across the street. Camera position changes composition more than any adjective, and it is the detail people most often leave out. If a result feels flat, the camera position is the first thing to add.

6. Write for the model you are using

The same idea, formatted wrong, loses a lot in translation. Midjourney likes compact phrases with parameters at the end, which our Midjourney prompt generator adds for you. Stable Diffusion responds to weighted tags and a separate negative prompt, as covered on the Stable Diffusion prompt generator page. Flux and Gemini prefer plain, complete sentences; see the Flux prompt generator for how that looks in practice.

If you use several tools, keep one description and reformat it per model rather than writing four prompts from scratch. The Midjourney parameters guide explains the parts that only Midjourney understands.

7. Reverse-engineer images you love

The fastest way to learn prompt vocabulary is to work backwards. Take an image with the look you’re after, turn it into a prompt, then change one element at a time (subject, light, palette) and watch what moves. We describe the full workflow in how to reverse-engineer any image into a prompt.

Do that a dozen times and you will start writing specific prompts without thinking about it.

A quick checklist

Before you press generate, check that your prompt answers these questions:

1. What is the subject, and what is it doing?
2. Where is it, and what surrounds it?
3. Where is the light coming from, and what is it like?
4. Where is the camera, and which lens?
5. Which single style or medium?
6. Which two or three colours?
7. Is it formatted for the model you’re using?

If a question has no answer in your prompt, the model will answer it for you, usually with the most ordinary option available.