आइडियोग्राम 4.0

Ideogram's fourth-generation model, built around typography and graphic design. It accepts only a fixed set of output dimensions, all of them 2K or larger.

Ideogram made its name on text inside pictures, and 4.0 is the release where the lab turned that into an argument about the whole shape of a design model. It is a 9.3B single-stream diffusion transformer of 34 layers, trained from scratch and published as Ideogram's first open-weight foundation model, with a commercial licence for teams that want to fine-tune it on their own brand material and run it inside their own infrastructure.

The decision that matters most for prompting is that the model was trained exclusively on structured JSON captions, and the reference pipeline validates every prompt against that schema and rejects anything that does not parse. Training and inference share one format, so the model is never asked to interpret an input shape it has not seen. Three things fall out of that. Colour is conditioned directly: up to sixteen hex values per image and five per element, rather than a description the model has to guess at. Layout is addressable: any element can be placed by a bounding box in normalised coordinates, and Ideogram reports 0.69 mean intersection-over-union on the 7Bench layout evaluation for how tightly generated objects land inside the boxes they were given. And a text element is typed - it carries the literal string to render plus a separate description of how it should look, which is the mechanism behind multi-line, multi-font in-image type rather than a single rendered phrase.

Typography is still the headline capability. Ideogram reports 0.97 English OCR accuracy on the X-Omni text-rendering evaluation, and puts that figure alongside parameter count to make its actual claim: that a 9.3B model beats every larger open-weight release on text, including ones several times its size. On the other axes it reports 0.76 on the SpatialGenEval spatial-reasoning split and 0.89 on the Prism-bench alignment track for following long compositional prompts. In its own blind arena, where graphic designers pick the better of two images without being told what made either, 4,366 votes across nine pipelines put 4.0 second overall and first among open-weight models.

The direction Ideogram says the model starts is layers rather than flat frames. Its reasoning is that production design does not end at a single pixel layer - headlines change before launch, cutouts move to new backdrops - and the lab already ships those as separate steps: a Background Remover that returns a clean alpha cutout from any generation, and a Layerize step that pulls headlines, body copy and graphic elements back out as editable layers. The stated plan for the next 4.0 release is to return alpha channels and editable text layers straight from inference, with no second pass.

Two limits are worth carrying into a prompt. Ideogram's own designer arena puts a closed model ahead of it overall, so 4.0 is the best open model in that ranking rather than the best model in it. And the schema is not optional: the reference pipeline rejects a prompt that does not parse as valid JSON rather than interpreting it, so the structure has to be produced by something - the lab's own tooling, or yours - before the model ever sees the request.