Ideogram’s Open-Weights Image Model and the Future of AI Design

Ideogram released its first open-weight image generation model at 9.3 billion parameters, supporting up to 2K output resolution and capable of running on a single consumer GPU. Previous state-of-the-art models ran at approximately 80 billion parameters, making the size reduction roughly ninefold. CEO Mohammed Nourouzi attributed the efficiency to architectural innovation rather than compute scaling, noting that even ten times Ideogram's current funding would not allow the company to match Google's chip resources.

The model is trained exclusively on JSON-structured prompting, where every scene element, including position, bounding box, and layout, is fully specified before pixel generation begins. Submitting a one-word prompt returns a safety-blocked image not primarily due to content filtering but because the input is not a valid JSON specification. Nourouzi contrasted this with OpenAI and Google, which silently expand short prompts into richer inputs on the back end without showing users what the model actually receives. He argued that exposing the real input fosters more user control, consistency, and innovation, and that changing one JSON element while leaving the rest unchanged produces highly consistent output across generations.

Early image models trained on alt text scraped from the internet, which is often short or inaccurate. Ideogram's core training innovation involves using AI to convert images into detailed structured descriptions with bounding box and element information, then training a model to reconstruct images from those descriptions. The intermediate representation in the current architecture is approximately 4,000 tokens. Ideogram has also chosen HTML over a proprietary JSON structure for editable design elements, reasoning that large language models are already extensively trained on HTML and therefore handle it more reliably.

Accurate text generation within images was Ideogram's original differentiator when the company launched roughly three years ago, at a time when garbled text was a near-universal failure of image generation systems. Nourouzi said text generation quality maps directly to the graphic design and storytelling industries and remains central to Ideogram's identity. He described the 9.3 billion parameter model's text generation as comparable to much larger models including NanoBanana and GPT image despite its smaller parameter count.

Frontier models that score highly on leaderboards tend to produce the same visual style repeatedly because heavy reinforcement learning training pushes them toward average aesthetic preferences. Ideogram's open-weight model has undergone very little reinforcement learning, making it a raw model with more style variation but requiring more precise prompting to control. Nourouzi said image model benchmarks often fail to correlate with pixel fidelity or realism that professional designers care about, so Ideogram instead uses human designers conducting side-by-side comparisons between model versions and competitors to assess taste, which he defined as going outside the norm rather than conforming to average opinion.

The most significant unreleased feature Nourouzi cited is editable text and layout control, which had not yet shipped at the time of recording. He said design and marketing workflows require editable outputs rather than flat images, and that Ideogram's editing models, when released, will also use the JSON prompting approach. Ideogram has also not yet released version 4.0 of its custom model training application but confirmed it is planned.

Three tiers of customization exist. The first is an open-source quantized model available through Hugging Face, the open-source community, and Comfy UI. The second is a custom model training application where artists with at least 50 pieces of work can upload images to train a personal style model, priced at 60 dollars per month for two training sessions. The third is high-touch enterprise work in which Ideogram annotation teams curate data, write detailed prompts, and learn company-specific keywords and mascot names. Enterprises reported that generic models failed to meet their design bar or follow brand guidelines, and after Ideogram trained custom models for them, clients said the model understood their brand DNA and used it for design ideation and marketing. Nourouzi noted that visual brand identity varies far more across companies than written communication does, making image model customization more valuable in enterprise contexts than language model customization. Scaling the 9.3 billion parameter architecture ten or one hundred times using mixture-of-experts approaches was described as a new frontier, with the note that mixture-of-experts does not necessarily increase storage requirements but makes models significantly more capable.