The difference between a great image prompt and a bad one usually comes down to specificity in four areas: subject, scene, style, and camera. Get those four right and most models produce something usable on the first or second try.
Subject, scene, style, camera
| Element | Weak example | Strong example |
|---|---|---|
| Subject | "a cute dog" | "A golden retriever puppy sitting in a wicker basket" |
| Scene | "in a kitchen" | "On a sunlit kitchen counter, soft morning light, shallow depth of field" |
| Style | "nice photo" | "Editorial food-magazine photography" |
| Camera | (unspecified) | "35mm, eye-level, three-quarter view" |
Each of the weak examples leaves the model to guess, and the guess is usually generic. Each strong example removes a decision the model would otherwise make for you.
What to leave out
- Adjective stacks. "Beautiful stunning amazing gorgeous photo" adds nothing: models don't weight extra adjectives as extra quality, they just add noise.
- Quality-keyword spam. "8k, hyperrealistic, award-winning" stopped meaningfully helping years ago as models improved their defaults; a couple of specific quality terms are fine, a wall of them isn't.
- Conflicting styles. "Photorealistic anime in a cubist watercolor" produces mush, because you're asking for three incompatible rendering approaches at once.
A reliable template
"[Subject doing thing], [scene with lighting and mood], [style], [camera/framing]."
Try it once on something simple: "A ceramic mug of black coffee on a wooden desk, late afternoon window light, editorial photography, 50mm, top-down": and you'll see the difference immediately compared to just typing "coffee mug photo."
Fixing the three most common failure modes
Failure: the subject looks generic/stock-photo-like.
Add one specific, slightly unusual detail: a chip in the mug's glaze, a specific breed instead of "a dog," a named material ("reclaimed oak" instead of "wood"). Specificity breaks the model out of its most statistically average output.
Failure: the lighting looks flat or artificial.
Name a real-world lighting scenario instead of describing an effect. "Soft morning window light" works better than "well-lit." Real photographers' lighting vocabulary (golden hour, overcast diffuse light, rim lighting) tends to produce more convincing results than generic terms.
Failure: the composition feels random.
Specify framing explicitly: "centered," "rule of thirds, subject on the left third," "close-up crop." Without this, the model defaults to whatever framing is most common for that subject in its training data, which is often a plain centered shot.
Iterating instead of restarting
If your first result is close but not right, don't rewrite the whole prompt: change one clause at a time and note what shifts. If the lighting is right but the angle is wrong, only touch the camera clause. This is the same principle as prompt engineering for text: isolate the variable you're actually trying to fix instead of re-rolling the whole thing and hoping.
When to stop prompting and start using a reference image
If you're three or four iterations in and still not hitting the look you want, most tools that support image-to-image or a reference/character image will get you there faster than continuing to refine text alone: text prompts are great at generating a concept, but weaker at exactly reproducing a specific existing look. Switch tools rather than fighting the text prompt past that point.
Want a ready-made starting prompt instead of writing one from scratch? Try the Image Prompt Generator or the Midjourney Prompt Generator for tool-specific syntax.
Building a personal prompt library
Once you find a subject/scene/style/camera combination that consistently works for your use case (product shots, social graphics, blog headers), save the full prompt text, not just the result. Swapping only the subject clause while keeping scene, style, and camera fixed is the fastest way to get a consistent visual style across a set of images: the same principle brand photographers use with a fixed lighting and lens setup across a shoot.
A quick reference for common use cases
| Use case | Style clause that tends to work | Camera clause |
|---|---|---|
| Product photography | Clean studio product photography | 50mm, eye-level, soft shadows |
| Blog header images | Editorial photography or flat illustration | 16:9 wide framing |
| Social media graphics | Bold flat illustration, brand colors | Square, centered subject |
| Team/personal headshots | Realistic photographic portrait | 85mm, shallow depth of field |
Start from this table instead of a blank prompt box, and adjust the subject and scene clauses to your specific need.
Negative prompts, when your tool supports them
Some image tools let you specify what to avoid directly, separate from the main prompt. If yours does, use it for recurring problems rather than fighting them in the main prompt: "blurry, extra fingers, text artifacts, watermark" as a negative prompt clears up a surprising number of otherwise-good generations. Don't rely on this as a first fix, though: a better-specified main prompt solves more problems than a longer negative-prompt list.