Prompt Engineering for AI Images: The 7 Elements Every Visual Prompt Needs

Text prompt engineering has RTF: Role, Task, Format. Image prompt engineering has something similar — seven decisions you make whether you know it or not.

Skip a layer and the model fills the gap. Sometimes beautifully. Often wrong for the brief. "A warrior in a forest" leaves environment, light, lens, and medium unspecified — so the model picks, and you iterate until luck aligns.

This framework works across Midjourney v7, FLUX.1, DALL·E 3 / ChatGPT Image, and SDXL. Only element 7 (parameters) changes shape per platform. Master the seven once; translate dialect per model.

The 7 elements (overview)

1. Subject — who or what, action, key details
2. Environment — where, time, weather, era
3. Lighting — source, direction, quality, color
4. Composition — framing, angle, depth, negative space
5. Style / medium — photo vs illustration, art movement, render type
6. Camera / optics — lens, aperture feel, film stock (when photographic)
7. Parameters / constraints — aspect ratio, model flags, negatives, must-avoid

Not every image needs all seven at full depth. Icons and abstract pieces might skip camera. Flat vector might skip lighting physics. But scan the list before you hit generate.

Element 1: Subject

The anchor. Noun first on Midjourney; clear head noun in FLUX/DALL·E sentences.

Include:

• Identity (person, object, creature) + action or state
• Distinctive attributes (material, color, era of clothing, scale)
• Count (one subject vs crowd — models default to solo)

Weak: "dog"

Strong: "golden retriever puppy sitting on grass, red collar, tongue out"

Avoid dumping twenty adjectives on the subject — save capacity for light and medium.

Element 2: Environment

Grounds the subject. Without it, you get studio void or random bokeh.

Specify:

• Location type (kitchen, alpine ridge, cyberpunk alley)
• Time (dawn, blue hour, midnight neon)
• Weather / atmosphere (fog, rain, dust motes)
• Depth cues (foreground rocks, distant city skyline)

Product shots: environment can be minimal — "white seamless backdrop" is a valid environment choice, not an omission.

Element 3: Lighting

Where amateur prompts die. "Beautiful lighting" means nothing.

Useful vocabulary:

• Direction: side light, backlight, rim light, overhead, underlight
• Quality: soft diffused, hard single source, overcast ambient
• Color: warm tungsten, cool daylight, neon magenta fill
• Named setups: Rembrandt, butterfly, split lighting, golden hour

Photography briefs: lighting is half the prompt. Illustration: replace with "flat even shading" or "dramatic chiaroscuro."

Element 4: Composition

How the frame is organized — independent of subject matter.

Terms that move pixels:

• Shot size: extreme close-up, medium shot, wide establishing
• Angle: eye level, low angle hero, bird's eye, Dutch tilt
• Lens behavior: shallow DOF subject isolation, deep focus landscape
• Layout: centered symmetry, rule of thirds, leading lines, negative space left for text

Design use case: "negative space on right third for headline" beats generating then cropping.

Element 5: Style / medium

Tells the model which rendering universe to simulate.

Pick one primary register:

• Documentary photograph, editorial fashion, product catalog
• Oil painting, watercolor, ink illustration, pixel art
• 3D render, clay model, isometric diagram
• Anime, comic cel-shade, Studio Ghibli-inspired (describe look, don't rely on living artist names when policies block)

Conflicts kill quality: "photorealistic watercolor 3D cartoon." Choose one; blend styles only deliberately with lower weights (SD) or --sw references (MJ).

Element 6: Camera / optics

Optional for non-photographic media. Essential for realism.

Examples:

• 85mm f/1.4, shallow depth of field — portrait separation
• 35mm street photography, slight grain — candid documentary
• 4x5 large format, tilt-shift miniature effect — architectural
• macro 100mm, focus stacking look — insects, products
• anamorphic lens flare, 2.39:1 cinematic — film still

Midjourney: pair with --style raw for photos. DALL·E: same words, no flags. SDXL: camera tokens + medium in positive prompt.

Element 7: Parameters / constraints

Platform-specific control layer.

Midjourney: --ar, --style raw, --stylize, --chaos, --sref, --cref, --no text, --v 7

DALL·E / ChatGPT Image: aspect in words; "no text overlay"; revision in chat

FLUX: quoted text strings; explicit "no watermark"; API size/aspect enums

SDXL: negative prompt field; (weight:1.2); CFG 5–9

Constraints also include must-not items: no logos, no extra people, no gore — use --no or negative field where supported.

Assembly order (works on every model)

Draft template:

[Subject + action], [environment], [lighting], [composition], [style/medium], [camera if photo] + [parameters]

Worked example — editorial portrait

Subject: woman in linen blazer, confident posture

Environment: minimalist concrete interior, large window

Lighting: soft overcast daylight from left, gentle shadows

Composition: medium shot, eye level, subject right of center, negative space left

Style: editorial fashion photography

Camera: 85mm, shallow depth of field, natural skin texture

Parameters (MJ): --ar 4:5 --style raw --stylize 100 --v 7

One-line MJ: Woman in linen blazer, confident posture, minimalist concrete interior, soft window light from left, medium shot with negative space left, editorial fashion photography, 85mm shallow DOF --ar 4:5 --style raw --stylize 100 --v 7

One-line FLUX: Editorial fashion photograph of a woman in a linen blazer standing in a minimalist concrete room, soft overcast light from a large window on the left, medium shot with empty space on the left side of the frame, 85mm lens shallow depth of field, natural skin texture.

7-element checklist (print this)

Before generate, tick mentally:

• Subject specific enough to recognize in a thumbnail?
• Environment defined or intentionally minimal?
• Lighting direction + quality named?
• Composition / framing specified?
• Single clear medium/style?
• Camera only if photographic?
• Parameters / negatives for target platform?

Missing ticks explain 80% of "AI slop" outputs.

Common mistakes by element

Subject: vague noun, no action, attribute overload

Environment: default gray void; accidental clutter you didn't want

Lighting: "cinematic" without direction; mixed warm and cool unintentionally

Composition: no aspect intent; subject dead center when design needed text space

Style: three mediums at once; "hyperrealistic illustration photo"

Camera: lens words on flat vector art (noise); missing lens on product macro

Parameters: MJ photo without --style raw; SDXL with 40-line negative wall

How many elements to specify?

| Goal | Minimum elements |

| Quick mood explore (MJ) | Subject + environment + style |

| Product hero shot | All 7 |

| Social illustration | Subject + style + composition |

| SDXL technical render | Subject + environment + style + parameters/negatives |

| ChatGPT Image draft | Subject + environment + lighting in prose — revise in chat |

More specification = less variance. Creative exploration wants less; client deliverables want more.

7 elements vs reverse-engineering

When analyzing a reference photo (our reverse-engineer guide), label each layer you see — then rewrite for target model.

PromptMake /image runs this decomposition automatically: upload reference, pick Midjourney v7 / FLUX / DALL·E / SD, choose Recreate or Variation mode. Output maps to model dialect. 3 free /image generations per day.

Use /text to turn a creative brief into a structured 7-element outline before pasting into your generator.

Cross-model translation reminder

Elements 1–6 are universal content. Element 7 morphs:

• Same six layers → MJ adds flags at end
• Same six layers → DALL·E becomes sentences + chat revision
• Same six layers → SDXL splits negative medium conflicts

See DALL·E vs Midjourney vs Flux — which needs better prompts — for difficulty ranking.

Frequently asked questions

Is this better than random keyword lists?

Yes for consistency. Keyword lists skip composition and lighting systematically.

Do I need camera terms for anime?

Usually no — use "cel shading, anime illustration" under style instead.

What's the most skipped element?

Lighting. Fixing light fixes more outputs than adding "8k detailed."

Can I reorder elements?

Keep subject first on MJ. FLUX/DALL·E tolerate sentence flow; subject still early.

Seven elements for video (Sora, Runway)?

Add motion as part of subject/action; camera movement as composition+. Video guides on the blog cover motion separately.

Related articles

Midjourney prompts that work 2026 — element 7 for MJ in depth.

SDXL weights & negatives — element 7 for SDXL.

Reverse-engineer prompt from photo — extract elements from references.

DALL·E vs MJ vs Flux — dialect after elements are defined.

Bottom line

Every visual prompt secretly contains seven decisions. Name them on purpose: subject, environment, lighting, composition, style, camera (when relevant), parameters.

You don't need a paragraph for each — you need no accidental gaps. Build the image in words the way a director blocks a shot: who, where, light, frame, medium, lens, then the technical suffix for your model.