Writing Image Prompts That Actually Work (Stop Typing '8k Masterpiece')

If you've ever typed something into an AI image tool and got back a picture that was technically what you asked for but nothing like what you pictured, the problem usually isn't the model. It's that you described the subject and left everything else to chance.

I spent my first week stuffing prompts with "8k, masterpiece, highly detailed, award winning" because that's what every tutorial from three years ago says. It does almost nothing on current models. Here's what actually moves the result.

The five-part structure

Not a formula to follow rigidly, just the five things worth saying, in the order that tends to work.

1 · Subject. What the picture is of, in plain words. "A ceramic coffee cup." Not "a beverage container."

2 · Setting. Where it is and what's around it. "On a pale oak table beside a linen napkin."

3 · Light. Direction, softness, warmth. "Morning light from a window on the left, soft shadows."

4 · Framing. Close-up or wide, eye level or from above. "Close, slightly from above."

5 · Style. Photograph, illustration, 3D render, only if you have a view. "Shot on a 50mm lens, shallow depth of field."

Put together:

A ceramic coffee cup on a pale oak table beside a linen napkin, morning light from a window on the left with soft shadows, close and slightly from above, shot on a 50mm lens with shallow depth of field.

Every clause there is doing work. Compare it with "coffee cup, 8k, masterpiece" and you can see why one of them produces the picture in your head and the other produces a coffee cup.

Light is the clause that buys realism

If you only add one thing to your prompts, add light.

"A product on a table" gives you flat, ambient, catalogue-of-nothing lighting. "A product on a table, low afternoon sun from the right, long soft shadow" gives you a photograph.

Direction, quality, colour, three words each, and they do more than every quality adjective combined.

What to leave out

Quality words. "8k", "masterpiece", "highly detailed", "award-winning", "trending on ArtStation." These were prompt-hacks for older models trained on captions that contained them. Current models on Gendia, Seedream, Nano Banana, GPT Image, Flux, are trained to follow instructions. Quality words just add noise.

Contradictions. "Minimal, richly decorated." "Wide shot close-up." The model will pick one, and not necessarily yours.

Everything at once. A prompt naming six objects, four colours and three moods produces mush. If a picture needs six specific things in specific places, that's a Canvas job, or a generate-then-edit job: not one prompt.

Saying what you don't want

Most models here take negatives inside the prompt itself rather than in a separate field. Phrase it as a positive where you can:

• Instead of "no text" → "clean surface, unbranded"
• Instead of "not blurry" → "sharp focus throughout"
• Instead of "no people" → "empty street at dawn"

Models are better at adding what you describe than at subtracting what you forbid.

The cheap iteration loop

This is the habit that separates people who spend 200 credits an afternoon from people who spend 2,000.

Draft on Flux Schnell at 2.5 credits. Rewrite the prompt five times. Ten runs is 25 credits: the price of half a Seedream image, and by the end you know exactly what your prompt needs to say.

Then run the finished prompt once on whichever model suits the job.

The mistake is iterating on a 125-credit model. You end up defending a prompt you're not sure about because you've already spent so much on it.

Two things that surprise people

Word order matters. Things stated early carry more weight. If the pose is the point, put the pose near the front.

Longer isn't better past a point. Around three to four sentences, extra description starts diluting rather than adding. If your prompt has grown to a paragraph, the fix is usually to cut, not to add.

Try it

Open the image generator on Flux Schnell. Write one bad prompt, three words. Then rewrite it with all five parts. Five credits total, and the difference between the two pictures is the whole lesson.

Then: choosing a model · every control explained