Most image-generation trouble starts before the render. You type a sentence, the model returns something glossy and wrong, and you spend twenty generations hunting for a fix. The expensive part is not the GPU time — it is the missing brief: nobody told the model what the image is for, which axis (light, framing, palette, brand) matters most, and what a good result even looks like.
A text LLM is surprisingly good at exactly that job. It will not paint, but it will interrogate a vague idea until it becomes a structured prompt, translate an art-direction vocabulary into tokens a diffusion model responds to, and critique the output you paste back. Treat it as the art director you cannot afford to hire — one who never gets tired of "make it warmer, but not orange".
Below are twelve prompt formulas, each tied to a concrete task: style, composition, light, branding, retouching, series consistency. Every one follows the same shape — task → prompt → example result — so you can copy it, swap the variables, and see what changes. Run them in any capable chat model; the image side works in Midjourney, DALL-E, Stable Diffusion and Flux, though each responds differently to syntax, which the prompts account for.
Why a text model beats your own prompt-writing instinct
Diffusion models do not understand intent, only proximity in their training distribution. "Cozy cafe" lands on a stock-photo average because that is the centroid of everything tagged "cozy cafe". An art director's job is to pull you off the centroid: pick a lens, a time of day, a film stock, a negative space choice. An LLM has read enough photography and design writing to do this on demand — if you ask for it in the right structure.
The trade-off is real. You gain speed and vocabulary; you pay in control drift — the model will confidently add details you never asked for, and if you accept them uncritically your whole series starts to look like the LLM's taste rather than yours. The prompts below are written to keep the human in the decision seat: the LLM proposes, you approve, the image model renders.
A second, non-obvious point: LLMs are much better at rewriting an existing prompt than at writing one from scratch. Give them a bad prompt and a diagnosis request and you get a sharper result than asking for a masterpiece cold.
The twelve formulas
1. The brief extractor — turn a vague idea into a structured prompt
Task: convert a one-line concept into a full prompt with subject, composition, light, palette, medium and mood, explicitly separated so you can edit one axis at a time.
Prompt:
You are an art director. I will give you a one-line image idea.
Return a prompt in this exact structure, one line per field:
SUBJECT: ...
COMPOSITION: ... (framing, camera angle, negative space)
LIGHT: ... (direction, quality, time of day)
PALETTE: ... (3-4 named colors, not hex)
MEDIUM: ... (photo / illustration / 3D render / etc.)
MOOD: ... (2-3 adjectives)
Do not add fields. Do not explain. Keep each line under 15 words.
Idea: [your idea]
Example result for the idea "a barista at the end of a shift":
SUBJECT: exhausted barista wiping a steel counter, apron stained
COMPOSITION: medium shot, slightly low angle, empty right third
LIGHT: single window light from left, late afternoon, hard shadows
PALETTE: warm gray, oxidised copper, off-white, deep umber
MEDIUM: documentary photograph, 35mm, shallow depth of field
MOOD: quiet, worn, unsentimental
Why it works: the field structure is a checklist against the classic failure mode — a prompt that describes the subject five ways and the light not at all. You can regenerate a single field ("redo LIGHT only, colder") without losing the rest.
2. The style translator — art-direction words into model tokens
Task: you know the feeling ("editorial, like a Sunday magazine") but not the tokens. The LLM bridges design vocabulary and image-model vocabulary.
Prompt:
Translate this art-direction brief into image-model tokens.
Brief: "editorial, like a Sunday newspaper magazine cover, restrained, lots of air"
For each of these axes give 2-3 concrete tokens a diffusion model recognises:
- photographic genre / era
- lighting setup
- color grading
- composition convention
Then give me ONE combined prompt line under 40 words.
Flag any token that is a brand name or living artist — I do not want those.
Example result:
Genre/era: mid-century editorial photography, 1960s magazine
Lighting: large softbox, high-key, minimal falloff
Colour grading: desaturated warm neutrals, low contrast
Composition: generous negative space, subject off-centre, horizon high
Combined: "high-key editorial portrait, 1960s magazine style, desaturated warm neutrals, subject off-centre with generous negative space, low contrast, soft large-source light"
Flagged: none
Why it works: the explicit ban on living-artist names keeps you out of both legal grey zones and the "style soup" that makes prompts unmaintainable. This is the prompt that most directly replaces a mood board conversation.
3. The composition fixer — diagnose why the frame feels wrong
Task: the render is technically fine but compositionally flat. You describe or paste the prompt; the LLM prescribes composition changes only.
Prompt:
Here is my current prompt: [paste]
The image comes out flat and centred. Do not change subject, palette or medium.
Give me exactly 5 composition edits, each as a concrete token I can add:
1. camera angle change
2. framing / crop change
3. negative space change
4. foreground-background layering
5. one element to remove
Format: token — one-sentence reason.
Example result:
1. "slightly low angle" — puts the subject above the horizon, adds authority
2. "tight crop on hands" — removes dead space above the head
3. "subject in left third, empty wall on right" — creates a place for a headline
4. "blurred foreground chair edge" — adds depth cue the flat render lacks
5. remove "symmetrical" — symmetry is what is flattening it
Why it works: constraining the LLM to composition prevents the usual failure where a critique request turns into a full rewrite and you lose the version you liked. One axis, five edits, applied one at a time.
4. The lighting simulator — predict the render before you pay for it
Task: choose the light before generating, and understand what each choice costs in mood and in render reliability.
Prompt:
Subject: [describe]. Give me a table of 4 lighting setups.
Columns: setup name | token string | mood it creates | what it hides | what it reveals | reliability in diffusion models (high/medium/low)
Then recommend ONE for a product shot that must feel premium but not cold.
Example result (excerpt):
| setup name | token string | mood | reliability |
|---|---|---|---|
| Rembrandt | "single key at 45 degrees, triangle of light on cheek" | introspective, classical | high |
| High-key | "large soft source, near-shadowless, white background" | clean, commercial | high |
| Hard noon sun | "direct overhead sun, short hard shadows" | harsh, documentary | medium |
| Practical neon | "mixed magenta and cyan practicals, wet reflections" | nightlife, cinematic | low — colour drifts |
Why it works: the "what it hides / reveals" columns are the part a prompt library never gives you. Hard light reveals texture and hides flaws; soft light does the reverse. Knowing that before rendering saves whole sessions.
5. The brand-kit enforcer — keep a palette and tone across generations
Task: you have a brand palette and a tone. The LLM writes a reusable prompt prefix that keeps every image on-brand, plus a checklist to reject off-brand renders.
Prompt:
Brand: [name]. Palette: [3-4 colours]. Tone: [3 adjectives]. Never show: [list].
Write a reusable prompt prefix (max 25 words) that encodes palette and tone
without naming colours as hex.
Then write a 4-point rejection checklist I can apply to any render in 10 seconds.
Example result:
Prefix: "muted terracotta and slate palette, soft daylight, matte finish, calm uncluttered staging, no glossy highlights"
Checklist:
1. Are there more than 2 saturated colours? reject.
2. Is there any pure black or pure white? reject.
3. Does the surface look glossy or plastic? reject.
4. Would this fit next to the last three approved images? if no, reject.
Why it works: the prefix is short enough that the image model does not dilute it, and the checklist turns taste into a repeatable test. This is the formula a marketer without a designer in the team will use most.
6. The negative-prompt generator — say what you do not want, correctly
Task: build a negative prompt that removes artifacts without removing the style.
Prompt:
I am generating: [subject + style].
List negative-prompt tokens in 3 groups:
A) technical artifacts (hands, text, watermark, extra limbs)
B) style killers specific to this look
C) composition killers specific to this look
Warn me about any token that is likely to leak into the positive image.
Format: group — comma-separated tokens.
Example result:
A — extra fingers, deformed hands, garbled text, watermark, jpeg artifacts
B — oversaturated, HDR, plastic skin, digital sheen
C — cluttered background, centred subject, fisheye distortion
Warning: "blur" as a negative will kill the shallow depth of field you asked for.
Why it works: group B and C are the ones beginners never write, and they are the ones that wreck a style. The leak warning is the non-obvious bit: negative prompts are not free — over-stuffed negatives pull the whole render toward a flat, safe average.
7. The prompt rewriter — upgrade a weak prompt without changing its intent
Task: take a mediocre prompt and improve it while preserving what the author meant.
Prompt:
Rewrite this prompt for [Midjourney / DALL-E / Stable Diffusion / Flux].
Keep the subject and intent identical. Improve only: specificity, ordering
(most important tokens first), and removal of vague adjectives.
Show the before/after and list what you deleted and why.
Prompt: [paste]
Example result:
Before: "beautiful modern office, nice lighting, professional, high quality"
After: "open-plan office, floor-to-ceiling windows, morning side light, oak desks, two people in conversation, wide shot"
Deleted: "beautiful" (unmeasurable), "nice" (unmeasurable), "professional" (redundant with subject), "high quality" (no effect on modern models)
Why it works: the "what you deleted and why" log teaches you to stop writing the four adjectives that do nothing. After a dozen rewrites you stop needing the prompt.
8. The series bible — keep 10 illustrations consistent
Task: you need a set of illustrations for a landing page or deck that look like siblings.
Prompt:
I need 8 illustrations for [project]. Build a "series bible":
- one fixed style block (medium, line weight, palette, level of detail)
- one fixed lighting rule
- one fixed composition rule (e.g. subject always right, text space left)
- 8 subject lines, each under 10 words
Output the bible as a block I can paste before every subject line.
Example result:
STYLE BLOCK: flat vector illustration, 2px uniform line weight, limited palette of 4 muted tones, no gradients, subtle grain
LIGHT: single implied light from top-left, no cast shadows
COMPOSITION: subject occupies right 60%, left 40% left empty for copy
SUBJECTS: 1) person opening a laptop 2) hand sorting cards 3) two people at a whiteboard ...
Why it works: consistency in image models comes from a fixed block, not from a fixed description. The block is your contract; the subject lines are the variables. Paste the block first every time and the set coheres.
9. The retouch brief — describe the edit you want in words
Task: you have a render that is 80% right; you need to describe the remaining 20% for an inpainting or a re-render.
Prompt:
The image is [describe]. What is wrong: [describe].
Write an inpainting instruction (what to select, what to replace it with)
and a fallback full-prompt revision if inpainting fails.
Keep everything outside the selection unchanged.
Example result:
Inpaint: select the background wall only. Replace with "plain warm off-white wall, no fixtures, soft even light". Prompt: "warm off-white plaster wall, soft even light, no objects".
Fallback: add "clean uncluttered background, no wall fixtures" to the main prompt and re-render at same seed.
Why it works: inpainting models respond to short, local descriptions — not to your full scene prompt. Splitting the instruction into "select / replace" mirrors how the tool actually works and stops the model from redrawing the whole scene.
10. The model translator — same idea, four syntaxes
Task: your prompt works in Midjourney but not in Stable Diffusion. The LLM rewrites it per model, respecting each one's real syntax.
Prompt:
Here is a prompt written for Midjourney: [paste]
Rewrite it for DALL-E, Stable Diffusion and Flux.
Rules: Midjourney uses --ar and --style parameters and responds to short
comma-separated phrases; DALL-E prefers natural-language sentences;
Stable Diffusion responds to weighted tokens in parentheses and a separate
negative prompt; Flux handles longer natural language well.
Keep subject and composition identical. Explain each change in one line.
Example result:
DALL-E: "A documentary-style photograph of an exhausted barista wiping a steel counter at the end of a shift, lit by a single window on the left in late afternoon, with an empty right third."
SD: "(documentary photograph:1.2), exhausted barista, steel counter, single window light from left, late afternoon, empty right third" + negative: "glossy, oversaturated, centred"
Flux: longer sentence version, no weights, no parameters
Change log: MJ parameters removed; SD weights added on medium; DALL-E converted to prose.
Why it works: each model has real syntax differences — parameters, weighting, prose vs tokens — and a naive copy-paste loses them. The change log is the part you keep for next time.
11. The critic — score a render against the brief
Task: you cannot see whether the render matches the brief. The LLM turns the brief into a rubric and scores what you describe.
Prompt:
Here is the brief: [paste]. Here is my description of the render: [describe].
Score the render against the brief on 5 criteria, 1-5 each:
subject accuracy, composition, light, palette, mood.
For any score below 4, give the single token change most likely to fix it.
Example result:
Subject 5, Composition 3 (subject too centred), Light 4, Palette 5, Mood 3 (too glossy).
Fix composition: add "subject in left third".
Fix mood: add "matte finish, no glossy highlights".
Why it works: a rubric makes "I don't like it" into a fixable list. You stop regenerating blindly and change one token at a time, which is the only way to learn what each token actually does.
12. The idea expander — 10 directions before you commit
Task: you have a theme but no direction. The LLM proposes ten distinct visual treatments so you choose before rendering.
Prompt:
Theme: [theme]. Give me 10 distinct visual directions, each with:
- one-line concept
- medium
- one surprising element
- risk (why it might fail)
No two directions may share a medium. Rank by how easy they are to render reliably.
Example result (excerpt):
1. Concept: a desk at dawn, only the lamp is on. Medium: photo. Surprise: steam from an untouched cup. Risk: steam reads as smoke. Reliability: high.
2. Concept: the same desk as a paper diorama. Medium: paper craft photo. Surprise: visible cut edges. Risk: lighting flattens the depth. Reliability: medium.
Why it works: the risk column is what makes this useful — it front-loads the failure mode so you pick the direction you can actually execute, not the one that sounds best on paper.
What these formulas have in common
Every one of them does the same move: it narrows the LLM's job to a single axis and asks for a structured output. That is also the discipline that transfers to the image model itself. Vague in, average out; structured in, controllable out.
The skills underneath — structured thinking, breaking a vague goal into checkable parts, critiquing output against a brief — are the same ones that make you effective with any AI tool. If you want to go deeper on that working style, the roundup on 12 prompts that turn an LLM into your Kubernetes sidekick uses the identical structure on a very different domain, and the Blender modeling guide covers the craft side of visual production that no prompt replaces.
Where people go wrong first
Three failure modes show up again and again:
- Over-stuffing. Beginners pile twenty tokens into a prompt. Modern models weight the first few heavily; the rest dilute. Use the rewriter (formula 7) to cut.
- Trusting the LLM's subject. It will invent details — a logo, a tattoo, a wrong number of fingers. You own the subject; the model owns the craft.
- Chasing consistency by description. A fixed description drifts across generations. A fixed style block (formula 8) holds.
Start with formulas 1, 3 and 11 — brief, composition fix, critic. Those three cover most of the gap between "I typed a sentence" and "I directed an image". Add the brand kit (5) and series bible (8) the moment you need more than one image for the same project. The rest are there when a specific problem appears, and by then you will know which one to reach for.