How to turn an image into a text prompt

First, separate visual prompting from OCR.

“Image to text” can describe two different tasks. Optical character recognition extracts letters already printed inside an image: a receipt total, a sign, a label, or a paragraph. Image-to-prompt writing does something else. It translates the scene itself into instructions for generating, recreating, or editing an image.

If a photograph contains a red umbrella in a concrete corridor, OCR may return no text at all. A visual prompt should capture the umbrella, the corridor, the low camera angle, the wet reflections, the blue-hour light, and the visual tension created by one warm color inside a cool scene.

Decide what the prompt must do.

The same reference can produce several correct prompts because the intended job changes what deserves emphasis. Before describing anything, choose one of four outcomes.

This decision prevents a common failure: copying every visible detail when only the composition matters, or writing a vague style label when the user actually needs a close recreation.

Use a six-pass visual reading.

Read the image from stable facts to interpretive details. The order below keeps the prompt coherent and makes later editing easier.

1. Subject and action. Name the primary subject first. Add count, age group, form, clothing, pose, expression, or movement only when visually important.
2. Setting and relationships. Describe where the subject is, what surrounds it, and which objects affect the story or balance of the frame.
3. Composition and camera. Record shot size, viewpoint, orientation, subject placement, depth, crop, focal plane, and negative space.
4. Light and color. Identify the light source, softness, direction, contrast, time of day, dominant hues, accent colors, and whether shadows are open or dense.
5. Material and style. Name surfaces, texture, medium, rendering treatment, grain, edge quality, and final-use context. The image prompt style guide explains how to replace vague labels with visible traits.
6. Constraints. State what must stay, what may change, and what should be absent. Use negative language sparingly and only for failures you can anticipate.

For a reusable order and optional fields, continue with what a good image prompt should include.

Turn observations into one working prompt.

Suppose the reference shows a single person under a red umbrella in a brutalist corridor at blue hour. A weak caption identifies the objects but discards the visual logic. A useful prompt connects those objects to framing, light, surface, and mood.

The longer version is valuable because each phrase controls a visible dimension. Remove “quiet suspense” and the geometry still survives. Remove “centered one-point perspective” and the entire image structure can change.

A template you can edit

Use the template as a diagnostic tool, not a form that must be filled completely. If a field does not change the picture, leave it out.

Adapt the prompt only after the visual reading is sound.

A model-neutral description gives you a stable source of truth. From there, adjust syntax and control language for the destination rather than rewriting the visual analysis from scratch.

• Midjourney: keep the visual phrase compact, place parameters at the end, and distinguish an image prompt from a style reference. See the Midjourney reference workflow.
• FLUX: favor direct positive descriptions and explicit reference roles. Avoid long lists of contradictory adjectives.
• GPT Image or conversational editors: write the desired result as a clear production brief, then state what must remain unchanged.
• JSON workflows: separate reusable fields such as subject, composition, lighting, and constraints without assuming the model requires JSON. See the image-to-JSON prompt guide.

Remove details that sound precise but do no work.

Check the prompt before you use it.

• Can a reader identify the main subject in the first clause?
• Does the prompt explain the frame rather than only naming objects?
• Are light and color described through observable behavior?
• Does each style word have a visible meaning?
• Are the details consistent with one another?
• For an edit, is the change separate from the preserve list?
• Could you remove any phrase without changing the likely result? If yes, remove it.

When you want more comparisons, study the annotated AI image prompt examples. When you are ready to work from your own reference, generate a prompt from your image and edit it using this checklist.

Common questions