Ideogram 4.0 works best with JSON prompts

Ideogram has published a guide on prompting its 4.0 image model with JSON. The company says the model was trained on structured captions that describe an image as labeled data, not as one long sentence. Users who carry over prompts from earlier versions may get layouts that shift between generations, generic styles and garbled text.

According to the guide, 4.0 is "more literal than you might expect." Vague words like "moody" or "editorial feel" have less effect than specific instructions, such as "warm directional light from the upper left."

The recommended format has three top-level fields:

• high_level_description: a scene summary of under 50 words.
• style_description: aesthetics, lighting, medium and an optional colour palette.
• compositional_deconstruction: a background plus an array of elements. Layout, text placement and bounding boxes go here.

Ideogram says the model has no default style. The keys inside style_description must also appear in a fixed order. For photography, that order is aesthetics, lighting, photo, medium, then color_palette. For illustration and design, it is aesthetics, lighting, medium, art_style, then color_palette. An open-source tool called CaptionVerifier checks the key order before generation.

The company also says plain-text prompts trip its safety filter far more often by mistake than structured prompts do. It may address this in a future checkpoint update, and for now it suggests adding more detail to the JSON. Users who don't want to write JSON can use Magic Prompt, which uses an LLM to turn natural-language input into an editable structured caption. In the live version, the model decides automatically whether Magic Prompt is on.

The source text cuts off at this point, and the guide gives no benchmark figures.