AI Image Prompt Guide: Style, Lighting, Tips

Every image prompt needs a clear subject. Before you think about style or mood, nail down exactly what should be in the frame.

Bad: "A cat"

Better: "A tabby cat sitting on a windowsill, looking out at rain"

Be specific about quantity, position, and action. If you want two people, say "two people." If they should be facing each other, say so. AI image models interpret your words literally, and they fill in anything you leave unspecified with random choices. The OpenAI DALL-E guide emphasizes that descriptive, detailed prompts produce more predictable results.

---

Layer your subject description from general to specific:

• Broad category -- person, animal, object, landscape
• Defining details -- age, breed, material, season
• Action or state -- running, melting, half-open

This order helps the model build a coherent mental image before adding fine details.

Watch out for clutter: two or three elements in a scene work well. Five or more competing focal points usually result in a cluttered, incoherent image. If your scene is complex, consider generating elements separately and compositing them.

Style is where your prompt goes from "a picture of a thing" to something with visual identity. You can reference:

• Art movements -- Art Nouveau, Bauhaus, ukiyo-e
• Media types -- oil painting, watercolor, 35mm film photography, pixel art
• Visual qualities -- flat illustration, hyperrealistic, low-poly 3D

Combining two or three style references often produces more interesting results than a single one. "Watercolor illustration with ink outlines in the style of botanical field guides" gives the model a clear direction that's more distinctive than just "watercolor." The Midjourney documentation shows how stacking style descriptors creates layered, nuanced outputs.

---

Medium matters as much as style. Specifying "digital painting" gives you a different texture than "acrylic on canvas," even with the same subject. Think about what you'd tell a human artist: the tool they should use, the surface they should work on, and the era they should reference.

When you want photorealism, add technical camera details:

"shot on Canon EOS R5, 85mm f/1.4 lens, shallow depth of field"

These terms activate the model's training data from actual photography, and the results look noticeably more like real photos than a generic "realistic photo" prompt.

Lighting changes everything about an image. The same subject under golden hour sunlight feels warm and hopeful. Under harsh fluorescent light, it feels clinical or unsettling. You need to specify lighting because the model's default is usually flat, even illumination with no particular mood.

Useful lighting terms:

• Golden hour / Blue hour -- warm or cool natural light at the edges of the day
• Backlighting / Rim lighting -- silhouettes and glowing edges
• Rembrandt lighting -- classic portrait lighting with a signature triangle of light
• Diffused overcast -- soft, even, shadowless illumination
• Neon glow / Candlelight -- colored or warm point-source light
• Volumetric light (god rays) -- visible beams through atmosphere
• Studio three-point lighting -- controlled, professional setup

Pair them with a time of day or weather condition for even more control: "golden hour lighting with long shadows on a dusty road."

---

Mood and atmosphere go hand in hand with lighting. Words like "moody," "ethereal," "gritty," "serene," or "dramatic" push the overall feeling of the image. These terms are imprecise on their own, but they work well as modifiers alongside specific lighting setups.

Bad: "dramatic mood"

Better: "Dramatic Rembrandt lighting in a foggy alley"

The Stable Diffusion documentation notes that lighting and atmosphere keywords are among the most influential tokens in image generation. Placing them early in your prompt, right after the subject, gives them more weight.

Composition tells the model how to arrange elements within the frame. Without guidance, you'll get a centered, medium-shot default. That's fine for headshots, but it's limiting for anything else.

Use photography and cinematography terms to direct framing:

• Extreme close-up / Close-up -- detail and emotion
• Medium shot / Full shot -- interaction and body language
• Wide shot -- environment and context
• Bird's-eye view -- looking down, makes subjects feel small
• Worm's-eye view -- looking up, makes subjects feel imposing
• Over-the-shoulder -- conversational perspective
• Dutch angle -- tilted frame for tension or unease

Compositional rules from photography translate well to prompts. Mention "rule of thirds" to push the subject off-center. Use "leading lines" to draw the eye toward a focal point. "Negative space on the left side" gives you room for text overlay in design projects.

---

Aspect ratio affects composition too. A 16:9 landscape naturally lends itself to environmental scenes, while a 9:16 vertical works for portraits and mobile content. The Midjourney docs let you set aspect ratio directly with the --ar parameter, and other tools have similar controls. Always set your aspect ratio before generating, not after, because cropping a square image into a wide one wastes most of the composition.

Color palettes set the tone of an image before the viewer even processes the subject. You can guide color in several ways:

• Name specific colors -- "teal and burnt orange"
• Reference a palette type -- "muted earth tones," "pastel palette," "monochromatic blue"
• Point to a cultural or era reference -- "1970s Kodachrome colors," "cyberpunk neon"

Texture adds tactile quality that makes images feel real or deliberately stylized. "Rough brushstrokes" feels different from "smooth airbrushed gradients." "Grainy film texture" feels different from "clean digital render." Think about what you'd feel if you could touch the image, and put that into words.

---

Combining color and texture creates a visual signature:

"Desaturated teal and amber with visible film grain and light leaks"

This is a specific aesthetic that the model can reproduce consistently across multiple generations. Useful when you need a series of images that look like they belong together, for example in a brand campaign or social media feed.

Quick fixes for common color issues:

• Oversaturated results -- add "muted colors" or "low saturation"
• Flat-looking images -- try "vibrant," "high contrast," or "rich colors"

Small color adjustments often have a bigger impact on perceived quality than changes to the subject itself.

Negative prompting lets you tell the model what to exclude from the image. Not all tools support it:

• Stable Diffusion -- dedicated negative prompt field
• Midjourney -- uses the --no parameter
• DALL-E -- handles exclusions less directly, usually through careful wording of the main prompt

Common negative prompt entries include: "blurry, low quality, deformed hands, extra fingers, watermark, text, cropped, out of frame." These address the most frequent artifacts in AI-generated images. Hands and fingers remain a weak point for most models, and explicitly mentioning deformities in the negative prompt helps, though it doesn't eliminate the issue entirely.

Think of negative prompts as guardrails, not creative tools. They're best for suppressing known problems rather than steering the creative direction. Telling the model what you do want is always more effective than trying to sculpt the image by listing everything you don't want.

---

A practical workflow:

• Step 1 -- Generate your first image without any negative prompts.
• Step 2 -- Note the specific issues.
• Step 3 -- Add those issues to the negative prompt and regenerate.

This targeted approach works better than starting with a long generic negative prompt list, because you're solving real problems instead of guessing at hypothetical ones.