What is a JSON prompt generator?
A JSON prompt is a description split into named fields. Subject, action, setting, light, and aspect ratio each sit on their own line, so you can change the light without rewriting the whole sentence. Reprompt (reprompt.org) builds that object from words you type. It also returns a plain-text paragraph that says the same thing.
That is a different job from the image to JSON prompt tool. There you upload a photo and the fields describe that photo. Here you start with an idea, and the fields describe the picture or clip you have not made yet. If you already have the image, use the upload tool. If you only have a sentence, stay on this page.
Which fields come back?
Every result has subject, action, setting, composition, camera, lighting, style, colors, text to render, aspect ratio, and a negative or avoid line. Colors are a short comma-separated list, not a hex code. Text to render is empty when the picture should not contain words. The negative field is for things to keep out, such as a watermark or extra fingers. Leave it blank in the paste if the target model has no negative box.
Video targets add four fields: motion, camera movement, duration, and audio. Image targets leave those four empty so you do not paste a camera move into a still-image model by mistake. The page checks that the object is valid JSON before it shows the copy button. If the check fails, you get an error instead of a broken block.
How should you paste it?
Nano Banana (Gemini) and GPT Image accept the JSON as ordinary prompt text. Neither product has a separate JSON mode for this schema. Midjourney, Flux, Veo, Sora, and Kling want the plain-text paragraph. Midjourney parameters such as --ar belong after the sentence, not inside the JSON. A ComfyUI workflow is also JSON, but it is a node graph. Do not paste this object into a workflow loader.
For Veo, the public guides describe a clip as cinematography, subject, action, context, and style, with dialogue in quotes and sound effects labeled. Duration of 4, 6, or 8 seconds and a 16:9 or 9:16 frame are product settings as well as words in the prompt. Sora and Kling prompts work better with one camera move and one subject action. This page writes those into fields so you can see them. It does not call the video model.
Image JSON or video JSON?
Pick the target before you generate. The field list changes with that choice. An image prompt that mentions a dolly move tends to confuse a still-image model. A video prompt with no motion tends to freeze the clip.
| Field | Still image | Veo, Sora, Kling |
|---|---|---|
| Subject, action, setting | Required | Required |
| Composition, camera, lighting, style, colors | Required | Required |
| Text to render, aspect ratio, negative | Use when they apply | Use when they apply |
| Motion and camera movement | Left empty | One action, one move |
| Duration and audio | Left empty | Suggestion plus a sound cue |
| What you paste | JSON for Gemini and GPT Image. Text for Midjourney and Flux. | The plain-text paragraph |
Style stays descriptive. "Soft window light, glazed ceramic, shallow depth of field" is something every model can read. "In the style of" plus a living illustrator's name is a shortcut this page does not write. If you need a look, name the medium and the light.
What does a real run look like?
This example was generated on 2026-10-06 with the same instruction the tool uses. The idea was "a ceramic mug of coffee on a sunlit wooden windowsill, steam rising, quiet morning, no text". The target was Veo 3. Nothing in the JSON below was typed in by hand after the run.
The mug stays a mug. Steam is the motion. Morning light stays morning light in the lighting field and in the paragraph. Text to render is empty because the idea said no text. The camera field is the lens and angle, "50mm macro lens at eye level". The camera movement is one move, "push-in", and the paragraph uses that same move, then the audio cue.
Captured with a 50mm macro lens at eye level, a ceramic mug of hot coffee sits on a rustic wooden windowsill with delicate wisps of steam rising into the air, set against a quiet morning backdrop. The camera performs a slow push-in movement. The scene is illuminated by warm, soft morning sunlight, rendered in a detailed and naturalistic style. Ambient noise: birds chirping softly and the faint sound of a distant breeze.
{
"subject": "ceramic mug of coffee",
"action": "steam rising from the mug",
"setting": "wooden windowsill on a quiet morning",
"composition": "close-up shot",
"camera": "50mm macro lens at eye level",
"lighting": "warm, soft morning sunlight",
"style": "detailed, naturalistic, high-resolution",
"colors": "brown, white, gold, warm tones",
"textToRender": "",
"aspectRatio": "16:9",
"negative": "text, people, clutter, shadows",
"motion": "gentle rising steam",
"cameraMovement": "push-in",
"duration": "6 seconds",
"audio": "Ambient noise: birds chirping softly and the faint sound of a distant breeze."
}JSON prompt questions
Is a JSON prompt the same as an image-to-JSON extraction?
No. This page starts from a sentence you type. The image to JSON prompt page starts from a photo you upload. Use this one when you already know the idea. Use the other one when you want fields that describe a picture you already have.
Which models can take this JSON?
Nano Banana and GPT Image can take the JSON pasted as the prompt text. They do not have a separate JSON mode for this schema. Midjourney and Flux want the plain-text version. Veo, Sora, and Kling want the plain text too. The JSON is there so you can edit one field, then paste the text.
What is a Veo 3 JSON prompt?
It is the same field list as an image prompt, plus motion, camera movement, duration, and an audio cue. Veo still wants a paragraph in the product. Duration and aspect ratio are also settings in the Veo interface. The JSON does not replace those controls.
Does Reprompt generate the image or the video?
No. Reprompt writes the prompt. You paste it into Gemini, AI Studio, Midjourney, Flux, Veo, Sora, Kling, or another tool you already use.
Why are some JSON fields empty?
Image targets leave motion, camera movement, duration, and audio empty. Those fields are for a clip. Text to render stays empty when the idea says not to put words in the picture.
What will the tool refuse?
Reprompt doesn't produce prompts for sexual or explicit images, minors, or non-consensual edits of real people.
Already have a photo? Use image to JSON prompt. For a paragraph aimed at one model, use the prompt rewriter. For a Gemini image prompt, use the Nano Banana prompt generator. For a still you want to describe as a clip, use image to video prompt.