How to JSON prompt for Ideogram 4.0

If you've moved from a previous version of Ideogram to 4.0 and found that your go-to prompts are giving you inconsistent results — layouts that shift between generations, styles that feel generic, text that garbles — you're not doing anything wrong. Our model has changed in a fundamental way to bring you stronger control over the output. This guide explains what changed, why it matters, and how you can take advantage of it.

4.0 was trained on structured JSON captions: a format that describes an image as organized, labeled data rather than a single run-on sentence. It doesn't mean you have to write code. It means the model has a preferred language now, and the closer you get to it, the more reliably it does what you ask.

Why your prompts might feel different in 4.0

Ideogram 4.0 is more literal than you might expect. Words like "moody," "vibrant," or "editorial feel" carry less weight now. The fix isn't to write more, but to write more specifically. "Warm directional light from the upper left" instead of "nice lighting." "Hasselblad medium format, slightly desaturated" instead of "film look." Named references, specific light sources, exact colours. That's the kind of vocabulary that 4.0 responds to, and the JSON format is just a way of making sure every part of your image has that kind of specificity attached to it.

JSON explained — no coding experience needed

Before 4.0, your prompt was one long sentence and the model filled in the gaps however it liked. Sometimes that was great. Other times, the logo ended up in the wrong place and the background was nothing like what you described, and you had no way to tell the model exactly which part was wrong.

JSON gives every part of the image its own labeled slot. What's important for Ideogram 4.0 is that your prompts need to be organised, labeled information. Each field in your JSON prompt tells the model exactly one thing: what the style is, what the background looks like, what text to render and where.

Here's the same image described in both formats:

Think of JSON as organizing your prompt into four clear jobs:

• What's the overall scene? (high_level_description)
• What's the visual style? (style_description)
• What's the background? (background)
• What specific things are in it? (elements)

You don't have to memorize the syntax. Magic Prompt (covered below) can translate natural language into JSON prompts that you can edit later. But once you understand the four jobs, you'll know exactly where to change something when a generation isn't quite right.

Breaking down the JSON format

If you've ever gotten a generation that was almost right — the composition was good but the style felt generic, or the lighting was flat when you wanted dramatic — the culprit is usually a missing or thin style_description. The model has no default style. If you don't tell it what kind of image this should be, it guesses. Here's what each field is actually responsible for:

Shifting from natural language to JSON

If switching to JSON sounds like learning a new language when you just want to make images, you don't have to. Magic Prompt is a built-in translator — you type a normal sentence, and before it ever reaches the model, an LLM rewrites it into a full structured JSON caption. On the live version of 4.0, Magic Prompt will turn ON or OFF automatically: the model decides whether to expand your input based on what you sent. Plain text gets expanded. A valid, well-formed JSON prompt passes through as-is.

Here's what that translation actually looks like:

Magic Prompt is opinionated — it makes decisions on your behalf that you might not want for controlled output:

• Populates sparse scenes with secondary elements you didn't ask for
• Upgrades vague references to specific ones ("Nike Air Jordan 1" instead of "retro sneaker")
• Removes hedging language — "could include," "various," "or similar" get stripped out
• Drops render metadata like bokeh and film grain unless you explicitly name them
• Commits to one value when you gave it two — "cream or ivory" becomes "cream"

For quick exploration, this is exactly what you want. For a brand deliverable where hex colours and text placement are non-negotiable, write JSON directly. Magic Prompt won't expand a prompt that's already structured.

Same scene, two approaches. Natural language gives general placement cues; JSON locks every element to an exact zone.

Why is the text in my image coming out wrong?

You tried to generate a logo mockup. The brand name came out as a smear of letters that vaguely resembles the right word. Or you asked for a poster headline and the model gave you something grammatically adjacent but spelled wrong. Or the text was fine but landed in completely the wrong font.

AI image models have struggled with text rendering for years. 4.0 has made it largely solvable through literal text rendering: during training, the model learned to treat text inside quotation marks as exact rendering instructions. To take advantage of this, quote exactly what you want rendered. With this method, Ideogram 4.0 achieves 97% accuracy on large English text benchmarks.

Rough accuracy guide by size and language:

• Large English text → ~97% accurate
• Small English text → ~90% accurate
• Large Arabic text → works well for single words and short phrases
• Small Arabic text or full paragraphs → expect garbling and phantom letters

For numbers inside charts or labels, default to using words. Prompting with "forty-two percent" renders more reliably than "42%" at small sizes.

How to control exactly where things appear

If you've ever tried to generate a movie poster or a product ad, you know the pain. The headline drifts into a corner. The logo sits on top of the main subject instead of below it. The credit block is unreadable soup. You regenerate, and everything moves to a completely different position.

To solve this, Ideogram 4.0 was trained with bounding box coordinates. The model learned to couple plain-language descriptions with exact spatial locations on a 0–1000 grid, so a "credit block" lands where you put it and stays there across seed variations.

The format is [y_min, x_min, y_max, x_max], with 0 at the top-left and 1000 at the bottom-right. A box covering the lower-right quadrant would be [500, 500, 1000, 1000].

Elements land inside their declared box consistently, even across seed variations. Use this for anything with a layout requirement: posters, packaging, social media templates, infographics.

Three things to know:

• Ground, floor, and pavement are always background — never put a surface as a bounding-box element
• Keep total elements to 5 or fewer per image for best quality
• Boxes are layout guides, not pixel-perfect crops — leave some margin

How to use exact brand colours

You described your brand colour as "deep forest green with a warm undertone." The model gave you something olive-ish. You regenerated. Now it's teal. You added "like a Rolex dial" and now it's darker but still not right. Twenty minutes gone trying to get a specific green out of a sentence.

Ideogram 4.0 supports hex colour conditioning to break out of that cycle. Pass the exact hex code and it anchors the image's entire palette around that value. A hex code like #2D5A27 will stick between generations.

Identical prompt, different color_palette arrays. The model anchors the entire tone range — highlights, shadows, and reflections — against the hex values you provide.

How to keep the same look across multiple images

You generated a perfect image — the grain was right, the colour grade was exactly on-brand, the lighting felt like your reference. Then you tried a second image in the same series. Same style description. It looked nothing like the first one.

That's the core problem with style-in-language. The same words produce different interpretations across generations. Style Codes are Ideogram 4.0's answer to this. They're a fingerprint extracted from a specific generation that you can apply to any future prompt. Generate one hero image you love, copy its Style Code from the panel, and pass it as style_code. Every generation after that shares the same visual DNA.

Style codes extracted from higher-quality presets (V4_QUALITY_48) tend to produce better results when reapplied. Preset quality levels also affect consistency:

• V4_TURBO_12 (~21s at 1024²) — exploration and drafts
• V4_DEFAULT_20 (~32s at 1024²) — daily production work
• V4_QUALITY_48 (~72s at 1024²) — finals, typography-critical output, style code extraction

The settings that actually change your output

Plain text vs. JSON: which one should I use?

Magic Prompt is always AUTO on live 4.0 for now, so the real choice is whether to write plain text or JSON. Plain text gets expanded automatically. JSON goes through as-is, giving you full control. Here's when to use one over the other:

What to try next

The techniques in this guide compound. Start with a plain-language prompt and get something close. Add a style_code from that result to lock the look. Switch to JSON when you need exact colours or placed text. Most workflows land on a combination of all three.

• Generate at scale with the API. Every technique here works over the API — batch runs, seed pinning, style code reuse across a full campaign. See the generate endpoint.
• Explore the open-source tooling. CaptionVerifier and all three Magic Prompt backends are available at github.com/ideogram-oss/ideogram4 if you want to run them locally or swap in your own LLM backend.