AI Image Prompt Generator

An AI image prompt generator that speaks each model’s language

Image models do not read prompts the same way. This AI image prompt generator takes one idea — a subject, its action and setting, a style, light, colour, framing and anything to keep out — and writes it five ways: labelled segments for GPT Image, a parameter line for Midjourney, weighted tags plus a separate negative prompt for Stable Diffusion, natural sentences for Flux and Ideogram.

The biggest difference is what happens to “leave out”. Stable Diffusion takes a real negative prompt, Midjourney uses --no, GPT Image follows an explicit constraints line, but Flux has no negative prompt at all and Ideogram 4.x only takes positive rephrasing — so the tool rewrites “blurry” as “sharp focus throughout” and flags anything it cannot rewrite.

Stuck for an idea? Surprise me builds one from a seed, and the same seed always returns the same idea, so you can share it. For deeper Midjourney work — version checks, permutations, multi-prompt weights — open the Midjourney prompt generator.

How to write an image prompt

1. Pick the model under Write the prompt for, or start from an example.
2. Fill in Main subject, an action and a setting; add exact words under Text in the image if you need lettering.
3. Choose one of 64 Style presets, then lighting, palette and mood.
4. Set shot size, angle, lens and Aspect ratio; for GPT Image the matching pixel size appears below.
5. List anything to Leave out (the preset buttons add common sets), check the list, then copy the prompt — and the Negative tab for Stable Diffusion.

AI Image Prompt Generator features

How each model wants the same idea

Sources read on 14 October 2026: OpenAI’s GPT Image prompting guide (21 April 2026) and image-generation API guide; the AUTOMATIC1111 wiki and ComfyUI README for emphasis syntax; Black Forest Labs’ FLUX.2 prompting guide and “Working Without Negative Prompts”; Ideogram’s prompt-structure and text guides. Google’s Imagen API page now says the Imagen models are shut down, so they are not a target here.

Worked example: one poster, five prompts

The start example — a jazz poster with the headline “JAZZ ON MAIN” in a 2:3 frame — gives this GPT Image prompt (size 1024 × 1536):

Intended use: event poster.
Scene: a rain-slicked city street at night.
Subject: an elderly saxophonist, playing a solo under a single streetlight, eyes closed.
Style: vintage screen-printed poster with a limited palette.
Composition: medium shot, low angle.
Lighting and colour: warm sodium-lamp glow; navy, mustard and cream; soulful mood.
Text: render "JAZZ ON MAIN" exactly once, across the top in bold condensed capitals, spelled exactly as written.
Constraints: no watermark, no extra text, no logos.

For Flux the same idea becomes sentences, and the exclusions turn positive — “watermark” and “extra text” become “plain, unmarked surfaces”, while “logos” is flagged because no positive rewrite fits:

An elderly saxophonist, playing a solo under a single streetlight, eyes closed. Vintage screen-printed poster with a limited palette. The setting is a rain-slicked city street at night. Warm sodium-lamp glow, navy, mustard and cream, soulful mood. Medium shot, low angle. The text "JAZZ ON MAIN" appears across the top in bold condensed capitals. Plain, unmarked surfaces.

Getting text right in images

• Keep the words short and in Latin letters where you can; long sentences invite typos.
• Say where the text goes and what it looks like — “across the top in bold condensed capitals”.
• For GPT Image use medium or high quality for small or dense text, as OpenAI recommends.
• On Stable Diffusion, add lettering afterwards in an editor if the checkpoint garbles it.