Ideogram 4.0 open-weights model hits 0.97 English OCR

Ideogram 4.0 is positioned less as a generator of mood images and more as a workspace for outputs close to usable design assets. The Japanese page frames the workflow around deciding the use case, writing subject, text, typeface, composition, and ratio into the prompt, reviewing multiple options, then finishing in Photoshop, Figma, or Illustrator. In that framing, Ideogram 4 is the starting point and pre-delivery standardization stays in later steps.

The official pipeline fixes text placement through JSON text elements and normalized-coordinate bounding boxes, with a color palette of up to 16 colors (up to 5 per element) guiding by color values rather than description. Because the same schema is validated at inference, the input format matches training. Ideogram also lists separate 4.0 and 4.5 roles, with a site note that the linked service may not run the native 4.5 model; this article covers the 4.0-generation uses described on the release page.

On quality, official figures and independent evaluation answer different questions. The official five-axis comparison reports English OCR 0.97 and 7Bench mIoU 0.69, while independent blind testing by designers focuses on whether work can be shown to clients, where differences opened up in typographic craft and client-submission readiness. The authors themselves caution that the result is directional and do not claim statistical power for significant differences.