Ideogram 4.0 JSON Prompt Format: Field-by-Field Guide

Why JSON Prompts Produce Better Results Than Plain Text

Every other image model you've used accepts plain text: "a golden retriever on a skateboard." Ideogram 4.0 is different. It was trained exclusively on structured JSON captions — a format that describes images field by field: the subject, the style, the lighting, each element's position. When you send plain text, the model has to guess how to map your words to that schema. When you send JSON, you're speaking the exact language the model learned from.

The practical difference comes down to three things plain text can't reliably do:

For simple generations — a landscape, a portrait, a basic product shot — plain text is often good enough. Switch to JSON whenever you need precise layout, specific colors, or readable text inside the image. That's where the gap becomes visible.

The JSON Schema: Three Top-Level Fields

Every Ideogram 4.0 JSON prompt has the same three-part structure:

{
  "high_level_description": "...",
  "style_description": { ... },
  "compositional_deconstruction": { ... }
}

Field 1: high_level_description

This is one or two sentences that summarise the whole image. The model reads this first to understand the overall scene before it processes the detailed fields below it.

Write it like a caption for a photograph. Include: camera distance (close-up, medium shot, wide), what the subject is doing, and where the scene takes place. Mention the medium here — photograph, illustration, graphic design — because it primes the style block before the model reads it.

// Photography
"high_level_description": "A medium-shot photograph of a barista pouring latte art in a cosy café at golden hour."

// Graphic design
"high_level_description": "A bold typographic poster for a fictional jazz concert. Clean navy background, amber headline, minimal layout."

// Illustration
"high_level_description": "A flat-design illustrated map of a fictional city district with four colour-coded zones and labelled streets."

Field 2: style_description

This block controls the visual style, lighting, medium, and color palette. There are two versions — one for photographs and one for everything else — and the key ordering inside each version is strict.

For Photographs

Key order must be: aesthetics → lighting → photo → medium → color_palette

"style_description": {
  "aesthetics": "warm, cinematic, golden hour",
  "lighting": "directional late-afternoon sunlight, long soft shadows from the left",
  "photo": "35mm, f/1.8, shallow depth of field, eye-level",
  "medium": "photograph",
  "color_palette": ["#F5C542", "#87CEEB", "#4A4A4A"]
}

For Illustrations, Graphic Design, and 3D

Key order must be: aesthetics → lighting → medium → art_style → color_palette

"style_description": {
  "aesthetics": "minimal, professional, geometric",
  "lighting": "even, diffuse studio lighting — no cast shadows",
  "medium": "graphic_design",
  "art_style": "flat vector design, bold outlines, generous whitespace, sans-serif typography",
  "color_palette": ["#FFFFFF", "#333333", "#0066FF"]
}

Valid medium Values

Key Order Reference Table

Note: color_palette is the only field that may be omitted. If included, it must always be last.

Field 3: compositional_deconstruction

This is the required field. It's where you describe the actual content of the image — the background and every element in it. It has two required sub-fields: background must come before elements.

"compositional_deconstruction": {
  "background": "...",
  "elements": [ ... ]
}

The background Field

Describe the setting in detail. Don't write one sentence — write the full environment: time of day, atmosphere, surface textures, depth, what's visible in the distance. The model uses this to build the world the elements live in.

"background": "The interior of a small independent coffee shop at night. Exposed brick walls painted in deep charcoal. Pendant Edison bulbs hang from a wooden ceiling, casting warm amber pools of light on the wooden counter below. Rain streaks down the large plate-glass windows on the left. A blurred street is visible through the rain."

Elements: type: "obj"

Each object in the image is an obj element. Give it a type, an optional bbox for placement, and a desc with as much detail as you can give.

Key order within an obj element: type → bbox → desc → color_palette

Elements: type: "text"

When you want text rendered inside the image — on a poster, a sign, a product label, a business card — use type: "text" instead of type: "obj". This is how Ideogram 4.0 achieves its best-in-class text rendering.

Key order within a text element: type → bbox → text → desc → color_palette

{
  "type": "text",
  "bbox": [30, 50, 350, 950],
  "text": "SUMMER SALE",
  "desc": "Large bold uppercase sans-serif text in deep red, occupying the upper third of the image.",
  "color_palette": ["#DC2626"]
}

Element Key Order Summary

bbox and color_palette are optional. If included, they must appear in the positions shown above.

Bounding Boxes: How to Place Elements on the Canvas

A bounding box (bbox) tells the model where to place an element on the canvas. The format is:

[y_min, x_min, y_max, x_max]

Think of the canvas as a 1000×1000 grid. The top-left is (0,0) and the bottom-right is (1000,1000). To place something in the right half of the image, your x_min would be 500.

Color Palette Conditioning

Adding a color_palette array steers the dominant colors in the image. This is one of Ideogram 4.0's strongest features for design work. Instead of saying "warm amber tones," you give exact hex codes and the model steers toward them.

"color_palette": ["#FF6B35", "#F7C59F", "#004E89", "#1A659E", "#2B2D42"]

You can also add a color_palette to individual elements in compositional_deconstruction. This gives the model per-element color guidance — useful when you want different zones of the image to have different dominant colors.

5 Prompt Examples — Copy and Run

Below are five complete JSON prompts covering different use cases: typography, product photography, cinematic scenes, UI mockups, and illustrated maps. Click View Prompt on any card to see the full JSON, and hit Copy Prompt to grab it. Paste directly into ComfyUI's text node and click Queue Prompt.

These are placeholder images — swap in the real generated output once you've run the workflow.

🔍 Click to zoomFantasy Movie Poster

Whimsical adventure movie poster combining storybook fantasy, mixed-media collage design, and large-scale cinematic typography.

▶View PromptCopy Prompt

🔍 Click to zoomFashion Editorial Portrait

High-end fashion and travel editorial photography with natural beauty, cinematic lighting, and realistic environmental details.

▶View PromptCopy Prompt

🔍 Click to zoomRealistic

Realistic image

▶View PromptCopy Prompt

🔍 Click to zoomAction Photography

Cinematic high-speed motorcycle photography featuring a perfectly sharp rider against dynamic urban motion blur and light trails.

▶View PromptCopy Prompt

🔍 Click to zoomAction Movie Poster

High-intensity Hollywood action poster with explosions, military aircraft, drones, futuristic warfare, and dramatic cinematic lighting.

▶View PromptCopy Prompt

Use AI to Write Ideogram 4.0 Prompts For You

Writing JSON by hand gives you the most control — but there's a faster way to get started. You can paste a special skill file into ChatGPT or Claude, describe any image in plain English, and the AI will output a complete, ready-to-paste Ideogram 4.0 JSON prompt for you.

Click Copy Skill File below. This copies a set of instructions that tells the AI exactly how Ideogram 4.0 JSON prompts work — field order, bounding boxes, color palette rules, and how to expand vague descriptions into detailed prompts.

---
# IDEOGRAM 4.0 PROMPT GENERATOR — SKILL FILE
description: Works on Claude and ChatGPT. Paste this entire file as your first message, then describe your image....
---

# Ideogram 4.0 Prompt Generator

This skill converts natural-language image descriptions
into the structured JSON caption format that Ideogram 4.0
was trained on...

[Full skill file copied to clipboard — scroll to read or click Copy Skill File above]

Open a new chat in ChatGPT or Claude. Paste the skill file into the message box and send it. The AI will confirm it understands the Ideogram 4.0 format and is ready to generate prompts.

🔍 Click to zoom

After sending the skill file, type the instruction below into the chat (or in a new message), then replace the placeholder text with your image description. Click Copy Instruction to grab it ready to paste.

You are now an Ideogram 4.0 JSON prompt generator. Follow the skill file above exactly. When I describe an image, output a complete Ideogram 4.0 JSON prompt in two blocks: first pretty-printed so I can read it, then minified in a code block labelled "Paste into Ideogram". Always expand vague descriptions with rich detail.

Here is my image:
[describe your image here]

🔍 Click to zoom

The AI outputs two blocks. The first is pretty-printed so you can read and check it. The second — labelled "Paste into Ideogram" — is the minified version ready to copy directly into ComfyUI's CLIP Text Encode node. Copy the minified block, paste it into the node, and click Queue Prompt.

🔍 Click to zoom

Magic Prompt: Skip Writing JSON Entirely

Writing JSON by hand produces the most precise results, but there's a faster option: Magic Prompt. It's a server-side tool that automatically expands a plain-text prompt into a full structured JSON caption. You write one sentence; it outputs the complete schema.

How to Enter a JSON Prompt in ComfyUI

In your Ideogram 4.0 workflow, the text input node is a CLIP Text Encode node — the rectangular node labelled "CLIP Text Encode" connected by a wire to the sampler. It's usually on the left side of the canvas. Here's how to find it and paste your JSON:

1. Look for the CLIP Text Encode node on the canvas. It has a large text area in the centre and a wire running from its output to the sampler node. Click inside the text area to activate it — the border turns blue when it's active.
2. Select all existing text in the field (Ctrl+A on Windows / Cmd+A on Mac) and delete it.
3. Paste your JSON directly into the field. ComfyUI treats the content as a plain string and passes it to the model — no special formatting is applied.
4. Click the orange Queue Prompt button in the top-right corner of the ComfyUI interface. A progress bar appears below it and the active node highlights green as each generation step runs.
5. The finished image appears in the Save Image or Preview Image node on the right side of the canvas once generation completes.

Troubleshooting

Results look the same as plain text — JSON doesn't seem to help

The JSON is valid but the model isn't benefiting from it. Check two things. First, confirm your key order in style_description is exactly right — see the key order table above. Second, make sure compositional_deconstruction is present and contains both background and elements. If either is missing, the model falls back toward plain-text behavior.

"Image blocked by safety filter" — gray screen output

The model's safety filter triggered on your prompt. This happens more often with plain text than with JSON — the structured format has a lower false-positive rate. If you're using JSON and still hitting this, check your desc fields for phrasing that could be misread as unsafe. Rephrase descriptively and neutrally. Restarting the generation with the same prompt sometimes also resolves a false positive.

Colors in the image don't match my color_palette

Two common causes. First, check that all hex codes are uppercase — #FF6B35 not #ff6b35. Lowercase hex is not in the training distribution. Second, add more contrast to your palette — if all your colors are similar tones, the model has less signal to work with. Include both highlight and shadow hex values explicitly.

Text in the image is blurry or incorrect

Make sure you used type: "text" and not type: "obj" for every text element. Confirm the literal string is in the text field and the styling description is in desc. If the text is still blurry, add a bbox to anchor it — floating text without placement coordinates is harder for the model to render cleanly.

JSON causes an error in ComfyUI (node turns red)

ComfyUI doesn't parse the JSON itself — it passes the string directly to the model. A red node means a different problem: most likely a disconnected wire or a missing model file. The JSON text itself never causes a node to turn red. Check your ComfyUI troubleshooting guide for node errors.

Elements appear in the wrong position

Double-check your bbox coordinate order. The format is [y_min, x_min, y_max, x_max] — Y first, then X. Swapping them is the most common bounding box mistake. If an element that should be on the left appears on the right, the X and Y values are reversed.

Frequently Asked Questions

Published: 2026-06-11 · Last updated: 2026-06-11 · Tested on RTX 4090 (24 GB VRAM) · ComfyUI v0.3.x · Ideogram 4.0 fp8 checkpoint