Text-to-Image Prompt Cheatsheet - Style Guide

Writing image prompts is hard because you must describe every visual attribute of a scene that does not exist, and the model only understands combinations of keywords. Most people over-invest in the subject and lose their yield in the back half of the prompt: shot size, lens focal length, light direction and palette decide whether the result is merely acceptable or immediately usable. This sheet is organised into subject and detail, style and artist, composition and lens, light and material, and negative terms with parameters, deliberately decoupled from any single tool; only the last group carries SD-style weight parentheses, step and sampler settings, and Midjourney separators and flags. Reach for it when the composition is right but the finish is not - the fix is a focal length and a light type, not more style words - or when a face keeps collapsing, which usually means the subject line lacks an explicit count and layout constraint. A good test of a prompt is whether a different model renders something equivalent from it; if it collapses on a swap, you have written one model's private shorthand rather than meaning.

AI Prompts·25 commands·Last updated 2026-08-23

Subject & Detail 5

Subject: {object/person}, details: {material/expression/action}

Subject first, then details

Shot: {close-up/full/wide}; {environment} bg

Framing + environment

Era/context: {cyberpunk 2099/Tang dynasty}

Context anchor

Count & layout: {three cats in triangular comp}

Explicit count/layout

Emphasize: (red hair:1.3)

SD emphasis weight

Style & Artists 5

Style: {watercolor/pixel art/3D render/film}

Explicit style word

Reference artists: by {Greg Rutkowski, Moebius}

Borrow style via artists

Genre: {steampunk/minimalism/surreal}

Genre sets tone

Render: {Unreal Engine 5/octane render}

Boost texture

Era feel: {1980s anime/baroque}

Time-based style

Composition & Lens 5

Lens: {85mm portrait/24mm wide/macro}

Focal length sets perspective

Angle: {top-down/low/bug's-eye/bird's-eye}

Angle sets mood

Comp: {rule of thirds/symmetry/leading lines/negative space}

Composition rules

Depth: {shallow DoF/all in focus}

Focus control

Aspect: {9:16 portrait/square/wide}

Ratio per platform

Light & Material 5

Light: {Rembrandt/neon/golden hour/volumetric}

Light type sets mood

Material: {brushed metal/glass/velvet/liquid}

Material adds realism

Palette: {cool blue/warm orange/low-sat/high-contrast}

Palette unifies emotion

Mood: {hazy/crisp/cinematic/dreamy}

Overall mood word

Weather/time: {rainy night/noon/fog}

Env light supplement

Negatives & Params 5

Negative: blurry, lowres, deformed, extra fingers, watermark

SD common negatives

MJ params: --ar 16:9 --v 6 --style raw

Ratio/version/style

MJ weight: separate with ::, ::2 boosts

MJ multi-prompt weight

SD steps/sampler: steps 30, sampler DPM++

Quality params

SD size: --w 1024 --h 1024 (or 768)

Size per model

Command Examples

Full SD-family five-block prompt with parameters

Subject: a young woman with (red hair:1.3), details: a linen shirt, looking at the camera
Style: watercolour, by Greg Rutkowski
Composition and lens: 85mm portrait, shallow depth of field, close-up, a forest at dusk in the background
Lighting and material: Rembrandt lighting, frosted glass texture, cool blue low saturation
Negative: blurry, lowres, deformed, extra fingers, watermark
steps 30, sampler DPM++
--w 768 --h 768

Each block does one job, 768 is the native SD training resolution where larger sizes invite repeated limbs, and omitting resolution or steps leaves both to guesswork defaults.

Midjourney: multi-prompt weighting with parameters

Subject: mountains above a sea of clouds::2, details: churning cloud layers
Style: surrealism, 3D render, octane render
Composition and lens: 24mm wide angle, bird view, leading lines, panoramic
Lighting and material: volumetric light, golden morning light, crisp
--ar 16:9 --v 6 --style raw

:: splits the prompt by intent and ::2 weights the subject; --ar sets the ratio for the target platform, and --style raw drops the default beautification.

Multiple subjects: explicit count and layout

Subject: three cats in a triangular composition, details: one sitting sideways, two facing the camera
Composition and lens: 35mm, deep focus, medium shot
Environment: an old study with wooden bookshelves

Without a count the model usually draws one; naming three plus the triangular arrangement, then adding each pose, makes both number and layout controllable.

Context and era anchoring with a render engine for texture

Subject: a street repair robot, details: brass joints, exposed cabling
Era and context: cyberpunk 2099
Style: minimalism, Unreal Engine 5
Lighting and material: neon, a rainy night, reflections on wet pavement

One era line fixes the whole worldview, and the render engine supplies material and light realism — the cheapest jump from viewable to usable.

Common Pitfalls

Writing just a person as the subject with no count or layout: the model defaults to one subject and a multi-subject request silently fails.
Loading the prompt with style words but no focal length or light type: when the composition is right and the texture is not, the missing ingredient is lens and lighting, not more style.
Pushing SD-family resolution past 1024, where structural artifacts such as repeated limbs become noticeably more likely; 768 is the native training resolution.
Stacking a dozen negatives, which eats semantic space, pushes the subject down in weight, and flattens the image into something unfocused.
Over-weighting several terms above 1.3, which oversaturates the result and hardens edges — often worse than leaving them unweighted.

Tips

General text-to-image structure: subject → style/artist → composition/lens → light/material → negatives/params, stacked in order; earlier = higher weight.
SD uses (word:1.3) for emphasis and negative prompts to avoid flaws; Midjourney uses :: to separate multi-prompts with ::2 to boost, params via --ar/--v/--style.
Explicit style words (watercolor/pixel/3D render) and reference artists set the tone fast; focal length + angle decide perspective and mood — key to quality.
Negative prompts (blurry/lowres/deformed/extra fingers/watermark) are the SD safety net, sharply cutting wasted renders.

FAQ

In what order should a complete text-to-image prompt be stacked?

Five blocks in order: subject with details, then style or artist reference, then composition and lens, then lighting and material, then negatives and parameters. Earlier tokens carry more weight, so the subject has to come first and be concrete. People get stuck on the subject, yet most of the loss in usable-image rate comes from the back half: shot size, focal length, light type, and tone decide whether the result is merely viewable or actually usable.

Why specify a focal length instead of writing "beautiful"?

Focal length determines perspective, and adjectives change no geometry. 85mm compresses a face, 24mm opens up depth, macro gets close to texture. Pair that with a viewpoint (high, low, worm, bird), a depth of field choice (shallow or all in focus), and a composition rule (rule of thirds, symmetry, leading lines, negative space), and the spatial relations become controllable. The test of a finished prompt: another model should produce a comparable result from it.

Why does the same face keep coming out broken?

Almost always because the subject description omits an explicit count and layout. The model defaults to a single subject, so you must state the number and arrangement, for example three cats in a triangular composition. The other levers are weighting a key feature (the SD-style (red hair:1.3)), pairing shot size with environment (close-up, full body, or wide with a background), and anchoring era or cultural context (cyberpunk 2099, Tang dynasty).

How do the SD-family and Midjourney parameter syntaxes differ?

The SD family leans on weight parens, negatives, steps/sampler, and resolution: emphasize a feature with (word:1.3), exclude defects with the negative prompt, prefer steps 30 with sampler DPM++, and remember 768 is the native training resolution while 1024 and above tend to produce repeated limbs. Midjourney separates multiple prompts with :: and boosts with ::2, and takes --ar / --v / --style, such as --ar 16:9 --v 6 --style raw. The two syntaxes do not mix.

What do negative prompts actually fix?

They remove rejects, not bad compositions. blurry, lowres, deformed, extra fingers, and watermark are the generic SD negatives and they do cut the reject rate substantially, but no negative repairs a confused composition — that is a failure of the subject, composition, and lighting sections. Piling on negatives also consumes semantics and pushes the subject down in weight.

Official References

Each command links to its official documentation below, so you can verify the latest usage and read deeper.

Maintained by LaoHand

Publicly updated on Aug 23, 2026, continuously proofread against official docs.

Contact Us

Wrong command or description? Send us corrections, business inquiries or product feedback by email.