AI image generator

How to write a good image prompt

Image models don't read your mind: they interpret a description. The gap between a mediocre result and one that looks professionally made almost never comes down to the model — it comes down to how much concrete visual information you hand it. A good image prompt reads like the note you'd give a photographer before a shoot.

The recipe this generator applies has five layers: subject (what's in frame and what it's doing), style (technique and finish), lighting (probably the single biggest lever on how an image feels), framing (camera distance and angle) and exclusions (what you don't want to see). Change one layer at a time and you'll see exactly what each contributes.

One detail that surprises people: technical descriptors work better in English, because that's the language the training images were captioned in. That's why this tool composes the final prompt in English. If you want to go further, run your description through Claude first, or use our text prompt generator.

What to expect from a free generator

This tool uses open models served for free, so the trade is what it is: you pay nothing, and in exchange there's a shared queue that only holds one of your images at a time, a limit of roughly one image every fifteen seconds, and quality that's great for sketches, ideas and social posts but doesn't always match Midjourney or the paid models from OpenAI and Google on final work. There will be stretches where the service is congested or has burned through its daily quota and generates nothing at all: when that happens the tool says so and offers to take your prompt elsewhere, which is what you were really building. To compare the paid alternatives, see our model table and the cost calculator.

The classic weak spots are still there: text inside the image often comes out garbled, hands and fingers fail more than they should, and busy multi-character compositions tend to blend together. What works best is asking for fewer elements per image, describing where each one sits, and using the exclusions field for whatever keeps going wrong. If a result is halfway there, hit "Another variation": it keeps the prompt and changes the seed, which is the fastest way to explore without starting over.

Prompt examples that actually work

The fastest way to see the difference is to compare the same brief written two ways. A weak prompt names an object; a good one describes a scene. Since the dropdowns already handle style, light and framing, all your effort belongs in the first field.

Portrait. "A woman smiling" lets the model pick the age, the clothes, the background and even the camera lens, and by default it picks the most generic option it knows. "A baker in her sixties smiling behind the counter, flour on her apron and fresh loaves cooling behind her" gives it concrete anchors: age, trade, context and props. Pair that with realistic photography, a close-up and soft natural light and the result stops looking like stock imagery.

Product. "A bottle of oil" is a sticker. "A bottle of olive oil with a kraft paper label on a white marble table, an olive branch beside it" also describes the surface and the companion object, which is exactly what makes a product shot look expensive. Studio softbox lighting and a medium shot beat golden hour here.

Illustrated scene. When the output is going to be a drawing, say what is happening rather than only what is there: "an astronaut planting a small tree on the surface of Mars, Earth in the background" is a scene; "an astronaut and a tree" is two objects the model will arrange however it likes. Digital illustration, isometric low poly and watercolour respond particularly well to narrative descriptions.

The exclusions field. That last box is not decorative — it is how you fix the failures that keep recurring. "Text, letters, watermarks" avoids the unreadable signage image models insist on painting; "malformed hands, extra fingers" reduces the classic hand problem; "cluttered background" cleans up the composition. Only add what is actually going wrong, because a long exclusion list ends up competing with your own description. If you would rather start from something already written, our prompt library has templates ready to adapt.

Which aspect ratio to pick for where you publish

Format is not a cosmetic detail: it changes the composition the model produces, because a tall canvas pushes it towards tight framing and a wide one towards open shots. Choosing correctly up front saves you crops that ruin the image.

Square 1:1 is the safest and the one every model handles best: avatars, feed posts, thumbnails and anywhere you do not yet know where the image will end up. Landscape 16:9 is for article headers, video covers and slides, and it is also what wide establishing shots and landscapes exploit best. Vertical 9:16 is built for full-screen phone displays — stories, reels, short video — and works especially well with full-body framing and low angles. Portrait 4:5 takes up the most room in a vertical feed without becoming a story, and it is the best choice for portraits and product photography.

Two practical warnings. First: the further the canvas gets from square, the more anatomy gives the model trouble, so if you are generating people at 9:16 it pays to describe the shot precisely. Second: those numbers are what the tool requests from the service, but the returned image can land a few pixels off; if you need an exact size, adjust it afterwards in any editor. And if what you want is to compare which paid alternative gives better resolution, the AI tools directory lists the options sorted by use case, and the glossary is there if you trip over terms like seed, diffusion or negative prompt.