People love a formula. They want a fill-in-the-blanks template that turns words into a great image every time. The honest version of the AI art prompt formula is real, but it is not magic — it is a priority order that keeps you from leaving out the things that matter most. Get the order right and you stop fighting the model.

Here is the formula I actually use, then the reasoning, then variations for Midjourney, Stable Diffusion, and DALL·E 3.

The AI art prompt formula, in one line #

[Subject + action] + [composition / shot] + [environment] + [lighting] + [style / medium] + [color / mood] + [technical] + [parameters]

That sequence is not arbitrary. Earlier tokens carry more weight in most models, so the formula front-loads the things that define the image and trails off into the things that merely polish it. You can drop any slot, but you should drop from the right, not the left.

Slot by slot #

1. Subject + action

The non-negotiable. Name the thing and say what it is doing. “A fox” is weak. “A red fox mid-leap over a frozen stream” gives the model a pose, energy, and a setting hint in five words.

2. Composition / shot

Borrow from photography and film: close-up, wide establishing shot, low angle, dutch tilt, overhead flat lay, rule-of-thirds, bokeh background. This decides where the eye goes.

3. Environment

Place and time. “In a misty pine forest at first light” does triple duty — location, atmosphere, and an implied color temperature.

4. Lighting

The mood engine. Always name it and, ideally, its direction: backlit golden hour, overcast soft light, hard noon sun, neon rim light, single candle.

5. Style / medium

Photo, oil painting, watercolor, 3D render, screen print, claymation. Pick ONE primary medium. Stacking “oil painting, 3D render, photorealistic” gives the model contradictory instructions and it averages them into mush.

6. Color / mood

A palette steers the emotional read fast: muted earth tones, high-contrast teal and orange, pastel, monochrome.

7. Technical

Lens, film stock, render engine, resolution cues: 85mm f1.4, shot on Portra 400, Octane render, volumetric fog. These sharpen realism but cannot rescue a vague subject.

8. Parameters

Tool-specific flags. Aspect ratio first, always.

The formula in Midjourney #

Midjourney likes comma-separated phrases ending in parameters. Here is the formula filled out:

a red fox mid-leap over a frozen stream, low angle action shot,
misty pine forest at first light, cold backlight through the trees,
wildlife photography, muted blue and rust palette, 200mm telephoto,
shallow depth of field --ar 3:2 --style raw --stylize 200 --v 6.1

Read it left to right and you can hear the formula. Drop --stylize to let the subject dominate; raise it toward 600–1000 to let Midjourney’s aesthetic take over. For anime, the same skeleton runs under --niji 6 — see Niji vs Midjourney for when to switch engines.

The formula in Stable Diffusion #

SD splits the work between a positive prompt, a negative prompt, and numeric settings. The formula maps cleanly onto the positive prompt, and weighting lets you enforce priority instead of relying on word order alone:

Positive: (red fox mid-leap:1.3) over a frozen stream, low angle,
misty pine forest, first light, (cold backlight:1.2), wildlife
photography, muted blue and rust palette, telephoto compression,
sharp fur detail, 8k

Negative: blurry, deformed, extra legs, fused anatomy, watermark,
text, oversaturated, cartoon, low quality

Steps: 32 | CFG: 6 | Sampler: DPM++ 2M Karras

The (term:1.3) syntax is the SD superpower — it lets you boost the leap and the backlight without rewriting the sentence. The negative prompt enforces the “what to avoid” half. Our negative prompts guide goes deep on building reusable negatives.

The formula in DALL·E 3 #

DALL·E ignores flags and weighting. Translate the same slots into prose and it follows intent surprisingly well:

A dynamic wildlife photograph of a red fox caught mid-leap over a
frozen stream, shot from a low angle. The setting is a misty pine
forest at first light, with cold blue backlight filtering through
the trees. Muted blue and rust color palette, telephoto compression
with a softly blurred background, sharp detail on the fox's fur.

Same formula, just spoken in full sentences instead of comma fragments.

Watch the formula build an image #

The formula is easiest to trust when you watch it accumulate. Start with only the subject, then add one slot at a time and notice what each contributes. Here is the same Midjourney image growing slot by slot:

1. a samurai                                  (subject)
2. a lone samurai, low wide shot              (+ composition)
3. a lone samurai, low wide shot, bamboo
   forest in morning mist                     (+ environment)
4. ...soft diffused light through the bamboo  (+ lighting)
5. ...ukiyo-e woodblock print style           (+ medium)
6. ...muted indigo and moss palette           (+ color)
7. ...visible paper grain, fine line work     (+ technical)
   --ar 2:3 --stylize 200 --v 6.1             (+ parameters)

By line three you have a scene. By line five you have a point of view. The last two lines only refine what is already there. That is the whole argument for the order: every slot you add late is polish, and every slot you add early is structure. If you ran out of patience and stopped at line four, you would still have a usable image — which is exactly why you drop from the right.

This also shows why the formula beats a single mega-prompt. Each line above is a checkpoint you could generate from and judge. You are not betting everything on one paragraph; you are building a result you can steer.

When to break the formula #

The formula is a default, not a law. Three times I deliberately ignore it:

  • Abstract or texture work — no real subject, so I lead with medium and color.
  • Logo and icon work — flat, centered, minimal; lighting and lens are irrelevant. The logo design AI prompts guide uses a stripped-down version.
  • Style transfer — when a reference image carries the look, the prompt only needs subject and adjustments.

A quick template to steal #

Keep this taped to your monitor and fill the blanks:

[subject doing something], [shot type / angle], [place + time],
[named lighting], [single medium/style], [palette], [lens/film]
--ar [ratio] --v 6.1

Run it once, then change a single slot per regeneration. That is the entire game. The formula gets you to a strong first draft fast; disciplined iteration gets you the rest of the way. If you are still learning the fundamentals, start with how to write AI art prompts, then come back here when you want to systematize it. Ready for harder control, like regional prompting and weighting math? That lives in prompt engineering for AI art.