Ideogram 4.0

Ideogram 4.0 is a 9.3 billion parameter open-weight text-to-image model from Ideogram, released on June 3, 2026. It focuses on accurate in-image text rendering, bounding-box layout control, color palette conditioning, and output up to 2K, with weights available to download under a commercial license.

Ideogram 4.0 is open-weight rather than fully open source. The weights, inference code, and prompting guide are public on Hugging Face and GitHub, and commercial deployments are covered by a license that scales with usage. You can download, fine-tune, and self-host the model.

Text rendering is the model's headline strength. Ideogram reports a 0.97 score on the X-Omni English OCR benchmark, which measures whether text inside a generated image is actually readable and correctly spelled, and the model handles multilingual text as well as English.

You attach bounding boxes to elements in the prompt, each coupled to a plain-language description, and the model places those objects inside their requested regions. Ideogram reports a 0.69 mIoU score on the 7Bench layout benchmark, which measures how tightly generated objects sit inside their boxes.

Instead of one long sentence, an Ideogram 4.0 prompt is a JSON object: a high-level scene description, a style block for aesthetics and lighting, individual elements with optional bounding boxes, typed text elements with the literal string to render, and an optional palette of up to 16 hex colors. The reference pipeline validates each prompt against the schema before generating.

Ideogram 4.0 generates from 256 to 2048 pixels per side with flexible aspect ratios. A single set of weights covers 1024 square output, 1536 by 1024 landscape and portrait, 1920 by 1088 widescreen, 2048 by 768 ultrawide, phone wallpapers, and 1584 by 396 social banners.

They are three sampling presets that trade speed for polish: Turbo runs 12 denoising steps, Default runs 20, and Quality runs 48. A common workflow is drafting compositions on Turbo, then re-running the chosen prompt on Quality for the final asset.

Yes. Ideogram publishes fp8 and nf4 quantized checkpoints that fit on a single 24 GB GPU, alongside the full weights and inference code. Teams can also fine-tune the model on their own brand or product data and deploy it inside their own environment.

On Ideogram's designer-preference ELO leaderboard, Ideogram 4.0 scores 1062, second behind the closed-source GPT Image 2 at 1141 and ahead of every other open-weight model. The practical difference is access: GPT Image 2 is API-only, while Ideogram 4.0 can be downloaded, fine-tuned, and self-hosted.