Ideogram’s prompting guide does not sell a formula. It publishes an order: eight parts, with a flat warning that the first sentence is the one worth getting right. What follows is that structure, the limits the guide sets for lettering, two of its worked prompts, and the questions it never answers.
What the official prompting guide documents
The guide is one page with ten sections beneath it, and its numbering is not worth trusting: the contents list calls negatives section 5, while the pitfalls page titles itself “6”. Cite the heading, not the number. The page carries no publication date either, only “Last updated 4 months ago”, so every detail here is dated to our reading of it on 2026-09-15.
The documented prompt structure, part by part
The eight parts the guide names
| Part | Purpose, in the vendor’s words |
|---|---|
| Image summary | One sentence for the whole picture |
| Main subject | Colour, shape, material, texture — and your text |
| Pose or action | What the subject is doing |
| Secondary elements | Props and ambient detail around it |
| Setting | Where it happens, indoors or out |
| Lighting | How the light looks and the image feels |
| Framing | Camera angle, shot type, subject placement |
| Technical enhancers | Lens, bokeh, brush texture, rendering style |
Two rows carry more weight. The guide says the image summary is the part to get right if you write only one sentence, and it puts required text in the main subject details, quoted and early. You need not use all eight.
Assembling them into one prompt
[Image summary]. [Main subject details], [Pose or action], [Secondary elements],
[Setting & Background], [Lighting & Atmosphere], [Framing & Composition], [Technical enhancers]Order is not decoration: the model gives slightly more importance to whatever sits earlier, which is why the subject and any lettering belong near the front. Length has a ceiling too, roughly 150 to 160 words or about 200 tokens. Past that, the guide warns, the model may ignore the rest of your prompt.
How Ideogram wants text written
Legible lettering is what this model is sold on, and the method is disarmingly plain: write complete sentences with punctuation, state the words you want in quotation marks, and describe the context around them. Its own first example is a poster on a wall reading “Everything you can imagine is real.”
The five practices the guide names
- Describe the text visually.
- Position text early in the prompt.
- Enclose the exact words in quotation marks.
- Break longer text into chunks, each with its own placement.
- Reduce visual complexity elsewhere.
Where text rendering breaks
Three limits are stated without hedging. Longer text raises the odds of spelling errors and half-finished words; the guide compares it to a group photo, where more people mean more chances somebody blinks. The model is not built for text-heavy documents. And rendering is most accurate in English, with non-Latin scripts called often unpredictable.
A typeface cannot be requested by name either, so you describe the lettering instead. Misspellings have published exits: regenerate, swap long words for shorter synonyms, overtype in the editor and remix at high image strength, or start from an image that already carries the text.
The iteration advice is just as plain. Change one thing at a time, whether that is the wording, a single detail, the style or the aspect ratio, and judge each result on its own before touching the next variable.
Worked examples: two official prompt assemblies
The guide assembles its parts into three complete prompts; two show the same skeleton doing different jobs.
Example one: the perfume bottle
A product photo of a men's perfume bottle named "Nightlife for men" in a sleek studio setup. The bottle is tall and rectangular with dark glass, a matte black cap, and silver lettering. The text "Nightlife for men" appears on the label in bold, modern font. The bottle stands upright with a slight reflection on the surface below. A wristwatch and a pair of sunglasses sit nearby, adding a masculine vibe. The scene is set on a smooth black surface with blurred city lights in the background. Lighting is moody and cool, with soft blue highlights and deep shadows. The bottle is centered in the frame, captured at eye level.Every slot is filled in order, and the product name appears twice, quoted, in the second sentence — so the model had no reason to guess the wording.
Example two: the Rhinos logo
A logo design for a local football team called "Rhinos" in green, blue, and white. The main graphic shows a strong, stylized rhino head viewed from a three-quarter angle, with sharp lines and a bold expression. The word "Rhinos" appears in large, blocky letters beneath the icon. Stars and shield shapes accent the logo without crowding it. The background is flat white, with no scene or setting. The color palette is vivid, with deep blue outlines, white highlights, and green fills.Same skeleton, different constraints. There is no set to describe, so the background slot carries flat white instead, and the enhancer slot holds vector sharpness rather than depth of field. The brand name is one short word in quotes, the case the typography page rates safest.
The rewrites the guide publishes as fixes
The pitfalls page prints the bad prompt beside the fixed one.
| What it names | Instead of | Try |
|---|---|---|
| Vague adjectives | A beautiful forest | A dense forest with tall pine trees and soft rays of sunlight filtering through the branches |
| Abstract concepts | A symbol of hope | A single flower blooming through a crack in the concrete |
Two habits fall out: adjectives the eye cannot check are replaced by things it can see, and contradictions get resolved rather than split. Aspect ratio silently rewrites framing too, so describe the region you want shown. The getting-started guide covers where these prompts are typed.
JSON prompting, plain text, and what the guide leaves open
The caption schema the open weights take
Ideogram 4.0 was trained on structured captions rather than sentences, and the open-weights repository publishes the schema. Three top-level fields: a short summary, an optional style block, and a compositional breakdown that is always required, holding a background first and then a list of elements. Each element is an object or a text entry, and either can carry a bounding box of four numbers on a normalised 0 to 1000 scale.
Two rules catch people: key order is strict and depends on the caption type, and colours must be uppercase six-digit hex, up to sixteen in the style block and five per element.
Two paths, opposite advice
The repository states that passing a plain-text prompt straight to the model will not work and may trigger a safety warning, because the weights expect a caption. That is not true of the hosted service: in the app and on the API, Magic Prompt expands your sentence into a full caption, which is why plain text works there. One model, two input paths, opposite rules — our Ideogram API guide covers which one a request gets.
What the guide never settles:
- No success rate. Nothing opened here publishes a text-rendering accuracy figure for any language.
- No parameter count or pixel mapping. The repository describes the architecture, not the size, and bounding boxes carry no stated link to a resolution.
The Ideogram topic page collects the rest of our coverage.
Frequently asked questions
Do I have to write JSON? In the app and through the API, no. Magic Prompt expands a natural sentence for you, and the structure section is written for plain text. On the open weights, a caption is what the model expects.
Why does my text come out misspelled? Length is the documented cause, and non-Latin scripts are the other. Shorten the line, keep it in English, move it earlier.
Can I ask for a specific font? Not by name. Describe the lettering instead.
How long should a prompt be? Under roughly 150 to 160 words, about 200 tokens. Past that, the vendor warns the rest may be ignored.
Does more detail always help? No. The guide says a short or abstract prompt is sometimes the better way to explore, and the eight parts are optional.