How to Write Better AI Image Prompts — Stop Getting Weird Results

You type "a cat" into an AI image generator. You get a blurry, misshapen blob with too many legs. This happens because the AI has to guess everything you didn't specify.

A good prompt doesn't just describe what you want. It removes ambiguity. Think of it as giving directions: "Turn left at the gas station" beats "Go that way" every time.

The 5-Part Prompt Formula

Every effective image prompt has five parts:

1. Subject — What's in the image. Be absurdly specific. Not "a dog" but "a golden retriever puppy, 8 weeks old, sitting, head tilted, tongue slightly out."

2. Style — The artistic approach. Photorealistic, oil painting, 3D render, watercolor, line art, cinematic, vintage photograph. This sets the entire aesthetic direction.

3. Lighting — Where light comes from and what it looks like. "Soft morning sunlight from a window," "dramatic side lighting with deep shadows," "neon street lights at dusk," "studio lighting with a white backdrop."

4. Composition — How things are arranged. "Close-up portrait," "wide angle landscape," "overhead flat lay," "Dutch angle," "centered composition with negative space."

5. Details — The little things that sell realism. "Sharp focus on eyes," "shallow depth of field, bokeh background," "film grain, 35mm Kodak Portra 400."

Bad Prompt vs. Good Prompt: Real Examples

Bad: "a castle"
Good: "A medieval stone castle on a cliff at golden hour, photorealistic, dramatic backlighting silhouetting the towers, mist in the valley below, wide-angle shot, 4K sharp detail, cinematic color grading"

Bad: "a person reading"
Good: "Young woman reading a worn paperback in a sunlit cafe corner, watercolor illustration style, soft diffused window light, steam rising from a coffee cup on the table, shallow depth of field, warm earthy color palette"

Bad: "a car"
Good: "Vintage 1967 Mustang Fastback in forest green, parked on a wet cobblestone street at night, neon reflections on the paint, light rain, cinematic still frame, 35mm lens, moody teal and orange color grade"

The difference is specificity. Every extra word reduces what the AI has to guess.

Quality, Aspect Ratio, and Format

Our AI image generator (SDXL) gives you three quality levels. Low (1 credit) for quick drafts. Medium (2 credits) for most use cases. High (3 credits) when you need maximum detail — it runs more inference steps but takes longer.

Aspect ratios: 1:1 (square, Instagram), 3:2 (standard photo), 2:3 (portrait orientation). Pick based on where the image will be used.

Output formats: PNG (lossless, larger file), JPEG (compressed, universal), WebP (modern format, smaller than both). PNG for editing, JPEG for sharing, WebP for web.

Using Reference Images

Upload a reference image alongside your prompt. The AI uses it for style and composition guidance — not to copy, but to understand the vibe you're going for. This is the cheat code for consistent results across multiple generations.

Example: upload a photo with lighting you like, then prompt for a completely different subject in that same lighting style. The AI preserves the lighting and transfers it to your new subject.

Common Mistakes

Too vague. "Something cool" gives you AI roulette. Too contradictory. "Photorealistic cartoon watercolor" confuses the model. Pick one style lane. Too many subjects. "A dog and a cat and a bird and a fish and a-" — the AI loses track. One clear subject, maybe two. Negative prompts help. Add "no watermark, no text, no signature" if you want clean output.