How to Write Better Prompts for an AI Image Generator (With Examples)
1) Start with the subject, then specify the scene
Strong AI image prompts begin with a clear subject and immediate context. Put the main subject first so the model anchors composition correctly, then add the environment, time, and action. This reduces “wandering” outputs where the generator invents extra objects or misplaces your focal point.
Template:
Subject + key attributes + action/pose + environment + time/weather
Example:
“Red fox with a torn ear, sitting alert on a mossy log, foggy pine forest, early morning, soft mist.”
2) Control composition with camera language
Most image generators respond well to photographic terms because training data includes captions and metadata. Use framing, lens, and angle to shape the result: close-up vs wide shot, eye-level vs low angle, shallow depth of field vs deep focus.
Useful terms: close-up, medium shot, wide shot, rule of thirds, centered, leading lines, symmetrical, negative space, low angle, overhead, macro, 35mm, 85mm, f/1.8, bokeh, deep focus.
Example:
“Portrait of an elderly woman with laugh lines, medium shot, eye-level, 85mm lens, shallow depth of field, soft bokeh lights behind her.”
3) Describe lighting like a cinematographer
Lighting is one of the fastest ways to improve prompt quality. Define the source, direction, color temperature, and mood. When lighting is unspecified, models often default to generic brightness.
Lighting options: golden hour, overcast softbox light, rim light, backlit, chiaroscuro, neon glow, candlelit, volumetric light, studio three-point lighting.
Example:
“Black cat on a windowsill, backlit by warm sunset, rim light outlining fur, dust particles in volumetric light beams.”
4) Choose an art style and medium precisely
Instead of vague terms (“cool art”), name the medium and production approach. “Oil painting on canvas,” “ink wash illustration,” “3D render,” or “paper cut collage” provides clearer constraints than broad labels. If you want realism, say so explicitly.
Example:
“Botanical illustration of a monarch butterfly, pen-and-ink with watercolor wash, off-white textured paper, labeled diagram style.”
5) Add material and texture cues
AI images improve when you specify surfaces: satin, weathered wood, brushed steel, cracked leather, wet asphalt, ceramic glaze. These cues help the model render believable detail and avoid plastic-looking outputs.
Example:
“Futuristic motorcycle helmet, matte carbon fiber shell, scratched visor, subtle dust, studio product photo on gray seamless backdrop.”
6) Use a constraint list to prevent unwanted elements
Many tools support “negative prompts” or exclusion terms. Even without a dedicated negative field, adding “no text, no watermark” often helps. Think of common failure modes: extra fingers, distorted faces, unreadable logos, random captions, clutter.
Common negatives: extra limbs, deformed hands, cross-eye, blurry, low-res, watermark, text, logo, overexposed, oversaturated, duplicate subjects.
Example (with negatives):
“Minimalist poster of a single white crane flying over a calm lake at dawn, pastel palette, lots of negative space — no text, no watermark, no border.”
7) Be specific with color palettes
Color direction prevents muddiness. Give a palette (two to five colors) and note saturation or contrast level. You can reference design language: monochrome, complementary colors, muted earth tones, high-contrast noir.
Example:
“Modern living room interior, Scandinavian design, muted earth tones (sand, olive, terracotta), soft natural window light, clean lines.”
8) Ground the prompt with references—without copying
Instead of naming a living artist, describe the traits you want: “bold brushstrokes,” “flat geometric shapes,” “1980s magazine airbrush,” “film-grain documentary.” This keeps the prompt ethically safer and more transferable across models.
Example:
“City street at night, cinematic neo-noir mood, high contrast, rain reflections, subtle film grain, anamorphic lens flare.”
9) Use ordered priority: what matters most goes first
Most generators weight earlier tokens more heavily. Lead with the subject and the biggest visual decisions (style, framing, lighting), then add secondary details (props, background items). If the model ignores a detail, move it earlier and simplify.
Example (prioritized):
“Macro photo of a honeybee covered in pollen, ultra-detailed, sharp focus on eyes, shallow depth of field, soft green garden blur.”
10) Avoid ambiguity and overloaded prompts
Too many adjectives can conflict: “minimalist, hyper-detailed, chaotic, clean.” Pick a direction and commit. If you need multiple variations, generate separate prompts rather than one “everything prompt.”
Bad: “A simple detailed ornate minimal busy logo…”
Better: “Minimal geometric logo of a mountain and river, two-color, flat vector, clean negative space.”
11) Add context for humans: age, ethnicity, wardrobe, expression
If you’re generating people, specify age range, facial features, hair, clothing era, and expression. Also define the activity to avoid stiff, mannequin-like poses.
Example:
“South Asian man in his 30s, curly hair, wearing a navy peacoat and scarf, candid smile, walking through a winter street market, snowflakes, warm stall lights.”
12) Prompt recipes (copy/paste examples)
A) Photorealistic product shot
“Studio product photo of a stainless-steel water bottle with condensation droplets, centered composition, softbox lighting from left, subtle shadow, white seamless background, 50mm lens, high detail — no text, no logo, no watermark.”
B) Fantasy landscape matte painting
“Epic fantasy valley with towering basalt cliffs and waterfalls, distant floating ruins, dramatic storm clouds, god rays breaking through, wide shot, atmospheric perspective, matte painting style, high detail, cool teal and slate palette.”
C) Children’s book illustration
“Whimsical children’s book illustration of a small raccoon wearing yellow rain boots, holding a red umbrella, puddles reflecting streetlights, simple shapes, soft watercolor textures, gentle outlines, warm and cozy mood.”
D) Logo / vector icon
“Flat vector icon of a sprouting seed inside a shield, simple geometry, consistent stroke width, two colors (forest green and cream), scalable, centered, no gradients, no text.”
E) Food photography
“Top-down food photo of shakshuka in a cast-iron skillet, vibrant tomato sauce, runny egg yolks, fresh cilantro, rustic wooden table, natural window light, slight shadows, high realism.”
13) Iteration workflow that reliably improves results
- Generate 4–8 variations with a clean base prompt.
- Pick the best image and identify one weakness (hands, background clutter, lighting).
- Modify only one or two prompt elements at a time.
- Use negatives to suppress recurring artifacts.
- Lock style and camera terms before adding fine details.
14) A practical “master prompt” structure
[Style/medium] + [Subject] + [Pose/action] + [Environment] + [Composition/camera] + [Lighting] + [Color palette] + [Quality cues] + [Negatives]
Example master prompt:
“Cinematic photo, astronaut sitting on a rooftop reading a paperback, futuristic skyline in the distance, wide shot, 35mm lens, dusk blue hour, neon rim light, muted blues with magenta accents, realistic textures, sharp subject, subtle film grain — no text, no watermark, no extra people.”