Ideogram 4 Does Not Want Your Sentence. It Wants a Layout Spec.
Ideogram's own prompting guide opens with a sentence most people using the model have never read: plain-text prompts "will not work and will likely trigger a safety warning."
Not work less well. Will not work.
That line sits at the top of the official guide, and it reframes a complaint I have heard from a dozen designers. Ideogram is the model people reach for when a picture has to carry readable type, and the recurring gripe is that it renders beautiful gibberish. Headline looks gorgeous. Words are wrong. Try again, get different wrong words.
That is not a model that is bad at spelling. That is a model being handed the wrong kind of input and doing its best.
What it actually eats
Ideogram 4 was trained on structured captions. Not sentences with structure in them. Actual nested objects, with fields in a fixed order.
There are three parts, and only the third is required.
The first is a one or two sentence summary of the whole picture. Strongly recommended, the guide says, and it is the only part that looks like a normal prompt.
The second describes treatment: the mood words, the lighting, the medium, and then either camera details for a photograph or a style description for anything drawn. Not both. You pick one lane. There is also an optional palette here, up to sixteen colours written as full uppercase hex codes. Shorthand gets rejected. #FFFFFF, never #fff.
The third part is where it stops being a prompt and starts being a layout. You give a background description, then a list of elements. Each element is either an object or a piece of text. Each one can take a bounding box, given as four numbers on a grid running 0 to 1000 from the top left. Each one can carry its own palette, up to five colours.
And a text element holds the words you want rendered in a field of their own, separate from the description of how they should look.
That separation is the whole ballgame.
Why the typography breaks
When you write "a minimal poster reading GOOD MORNING in bold navy sans-serif," the model has to work out on its own which of those words are glyphs to draw and which are instructions about drawing. It is a guess, and the guess is made inside the same soup as every other word in your sentence.
Get the boundary wrong and you get a poster with the word "BOLD" on it, or the right words in a style you never asked for, or something that looks like English from across a room.
In a structured caption there is no boundary to guess. "text": "GOOD MORNING" is the string. "desc": "large bold rounded sans-serif title in deep navy" is the treatment. "bbox" is where it goes. Three separate fields, three separate jobs, no interpretation required.
I would bet most of the people currently blaming Ideogram for bad typography are typing sentences at it.
The part nobody mentions: you get coordinates
The bounding boxes deserve their own paragraph, because they are the thing that moves this from prompting into something closer to actual design work.
[y_min, x_min, y_max, x_max], normalised 0 to 1000, origin top left. So the top third of a square canvas is roughly y 0 to 333. A logo in the lower left corner is somewhere around [850, 20, 980, 200].
The official worked example is a Formula 1 paddock shot with seventeen elements in it, most of them small pieces of text: sponsor logos on a sleeve, a number on a cap, a watermark in the corner. Every one placed by coordinates. You can read it as a spec sheet for a photograph that already exists, which is exactly what it is, and it tells you how much control is on the table.
For anyone who has fought an image model for twenty generations trying to get a headline to sit above the product instead of across it, this is the fix. You stop asking and start placing.
Put this into practice
Start here, today, in about ten minutes.
Take a layout you already made. A poster, a social card, a title slide. Something where you know where everything goes, because you put it there.
Write the third block first. Ignore the mood words for now. Describe the background in one honest sentence, then list your elements in order, background things before foreground things. For each one: is it an object or text? What does it say, if it says anything? What should it look like? Where does it sit?
Add the treatment block second. Aesthetics, lighting, medium, then camera details for photographic work or a style description for illustration. Pull five to eight hex codes straight out of your existing palette and put them in. Include your background colour and both your highlight and your shadow. The guide is specific that leaving the background colour out of the palette is a common way to not get the background you wanted.
Mind the key order. It is strict, and the pipeline runs a verifier that warns you when keys are missing, unknown, or out of sequence. For a photo: aesthetics, lighting, photo, medium, colour palette. For anything else: aesthetics, lighting, medium, art style, colour palette. Palette always last.
Or skip all of that. The repo ships a magic prompt that expands a plain sentence into a valid caption for you. The default path runs server-side, is free, and only needs an Ideogram API key. Ideogram are upfront that it is not the same expander their hosted product uses, so results differ. Use it as a scaffold: run your sentence through it, read the caption it produces, then edit that instead of writing from a blank file. Reading a good caption teaches the format faster than any guide.
Then change one thing. Move one bounding box. Swap one hex code. Regenerate. The composition should mostly hold. That stability is the real gain here, and it is what makes this usable for client work rather than slot-machine pulling.
If you want to run it on your own card
Four shrunk-down copies of Ideogram v4 went up on September 16, from rootonchair. Shrinking here means storing each number in the model with less precision, which costs a little quality and saves a lot of memory. There are two levels of shrinking, offered for both the Instant and the Fast version of the model. The smaller one comes to roughly 16 GB, against an original of 9.28 billion numbers at full size, which is what puts it inside a 24 GB card.
Two settings the card insists on: guidance at 1.0, with the guidance schedule off. Leave the defaults in place and the output is garbage, and it will look like the model's fault.
Before you download anything, know the licence. The upstream weights are gated behind Ideogram's Non-Commercial Model Agreement. This is a thing to learn on, not a thing to bill for.
Where this breaks
Small text is still imperfect. The card for the shrunk-down build says so directly, and says it is slightly worse at small type than the original. Set your type at poster scale. Do not try to render a paragraph of body copy and then be surprised.
Key order is fussy in a way that will annoy you. Palette has to sit last. You use photo or art_style, never both. Hex must be uppercase and full length. None of this is hard, all of it is the kind of thing you get wrong at eleven at night and spend forty minutes debugging.
The safety filter has false positives. You get a grey card reading "Image blocked by safety filter" on prompts that are entirely ordinary. Ideogram acknowledge it in the guide and say the rate is high specifically for non-JSON prompts, which is one more argument for writing the caption properly, but it still happens. Re-roll the seed.
Bounding boxes are a strong suggestion, not a guarantee. The guide frames the required-field language as what the caption verifier checks, not as a hard constraint on the pipeline. Expect placement to be close rather than exact, and design a layout that survives a bit of drift.
And the format is a real cost. Hand-writing a seventeen-element caption is work. For a quick idea, a sentence through the magic prompt is the right call. The structured version earns its keep when you know what you want and need it to come out that way twice.
What I think this means
Most prompting advice is folk knowledge. People trade phrasings that worked for them, and half of it is superstition dressed as craft.
This is the opposite. It is a published schema with field names, coordinate systems, colour formats and a verifier that checks your work. The model has an interface and the interface is documented. Anyone typing sentences at it is not prompting badly. They are using the wrong door on a building that has a perfectly good front entrance with a sign on it.
Which raises a question worth sitting with for whatever tool you use most. How much of what you have learned about coaxing results out of it is craft, and how much is you working around an interface you never read the documentation for?
Go read the guide. Then go move a bounding box and watch what happens.
Medium metadata
Title: Ideogram 4 Does Not Want Your Sentence. It Wants a Layout Spec.
Subtitle: The model was trained only on structured JSON captions with coordinates and hex colours, and Ideogram's own guide says plain text will not work. Here is how to write one.
Tags: Ideogram, AI Image Generation, Graphic Design, Prompt Engineering, Typography
Suggested read time: 7 minutes