Model Skills

---
name: ideogram4
description: Teaches an LLM to write effective image prompts for Ideogram4.
---

# Ideogram4

## When To Use

Use this skill when the user asks for an image or an edit made with Ideogram4 and you need to write or revise the prompt. Apply the rules below when composing the prompt text; do not paste the rules into the prompt itself.

## Prompting Rules

- Always verify that the bbox provides enough pixel space for the described subject to avoid cropping.
- Image references can be supplied alongside the prompt to maintain consistency when generating product imagery with accurate nutritional-information text.

## Text To Image

### Prompting Rules

- Ideogram 4.0 is trained exclusively on structured JSON captions, so prompts must be formatted as valid JSON to match the training distribution.
- Prompts should be structured as JSON captions that include per-element styling and optional bounding boxes and color palettes.
- Use size qualifiers such as 'huge', 'big', 'normal', 'small', and 'minuscule' to define the relative proportions and importance of elements within the image.
- Using descriptive precision in prompts helps refine results and increases the likelihood of generating the desired image.
- Remove unnecessary descriptive details from prompts to prevent the model from overriding specific style or framing instructions with literal interpretations of those details.
- Use JSON structured prompts to achieve precise, maximum, and repeatable control over text placement, positioning, color palettes, style, and composition, whereas plain text prompts leave layout and text positioning to the model's discretion.
- Specify a color palette using hex codes, for example: `color palette #0A2540, #00D1B2, #F5F7FA`, to achieve on-brand work.
- Structure prompts by leading with the subject, then layering details in the order: subject, style, lighting, composition, and mood.
- When generating infographics with multiple text blocks, use bounding boxes to define the position of each element, as this is critical to prevent text overlaps.
- For social media graphics, use wide bounding boxes for headline text and keep the element count low to ensure bold text with brand colors is rendered effectively.
- Use the 'photo' mode with specific camera specifications to achieve photorealistic results.
- Use the 4-part prompt structure: [Visual subject] + ["EXACT TEXT in quotes"] + [Design style] + [Layout / format].

### Anti-Patterns

- Never let the subject bounding box touch or exceed the frame edges vertically when generating full-body shots.
- Do not add full-body language to the description when the user specifies a crop such as knee-up, waist-up, or bust-up.
- Never place a full standing figure in a 2:3 portrait with ymin greater than 100 or ymax less than 850.
- Plain text prompts can suffer from ambiguity regarding spatial relationships and color-object associations, whereas JSON prompts resolve this by explicitly defining positions and colors.
- Passing plain-text prompts directly to the model will not work and will likely trigger a safety warning.

### Parameters And Constraints

- The inference pipeline validates prompts against a JSON schema and rejects inputs that do not parse, requiring strict adherence to the JSON format.
- The `color_palette` field is optional in the `style_description` object, but if included, it must remain in the final position of the key order.
- Use the `photo` field for camera and lens details (e.g., "35mm, f/1.4, bokeh") or the `art_style` field for non-photo styles (e.g., "flat vector illustration"), but do not use both fields simultaneously.
- The bounding box parameter controls product placement and negative space in the composition.
- Colors in the `color_palette` array must be uppercase `#RRGGBB` strings (e.g., `#1B1B2F`), not lowercase or shorthand formats.
- Per-element `color_palette` in `compositional_deconstruction` is optional and supports up to 5 hex entries.
- In the `style_description` field, you must choose either `photo` (for camera/lens specs) or `art_style` (for illustration/design style), but never both.
- In `compositional_deconstruction`, keys must follow the order: `type`, `bbox`, `desc`, `color_palette` for object elements (`type: "obj"`), and `type`, `bbox`, `text`, `desc`, `color_palette` for text elements (`type: "text"`).
- The `compositional_deconstruction` field is required in the JSON caption schema and must contain both `background` and `elements` sub-fields, with `background` specified before `elements`.
- Ideogram 4 requires prompts to be provided as structured JSON captions rather than plain text to achieve good results and enable controllability.
- Color palettes in the JSON schema accept up to 16 hex codes globally and up to 5 hex codes per individual element.

### Capabilities

- Ideogram 4 uses typed text elements that carry a literal string to render and a separate visual description for styling, enabling multi-line and multi-font in-image text.
- Ideogram 4.0 supports explicit color palette control by including hex color codes (e.g., #1B3A5C #5B8FB9) in the prompt to define the image's color scheme.
- Ideogram 4 supports color palette conditioning using up to 16 hex colors per image and 5 per element to steer dominant colors directly, bypassing descriptive language.
- Ideogram 4.0 can handle complex scenes with up to twenty-eight bounding boxes marking items, alongside text bounding boxes for labels.
- Ideogram 4.0 allows precise typography and text control by using six text bounding boxes to anchor specific text elements such as titles, subtitles, dates, taglines, credits, and RSVPs.
- Elements can be placed using bounding boxes defined as [y_min, x_min, y_max, x_max] in 0–1000 normalized coordinates with the image origin at the top-left.
- Ideogram 4.0 enables granular text styling by using six text bounding boxes to set each word at its own scale, position, and rotation.
- Ideogram AI versions 1.0 and up are very observant of the order of instructions in the prompt.
- Ideogram is particularly sensitive to prompt construction, including specific word choice, semantics, and punctuation.
- Ideogram 4.0 was trained exclusively on structured JSON captions, so prompts should be written as exhaustive descriptions of every element in the image.
- Use the 'Magic prompt' feature to have an LLM automatically expand a casual plain-text prompt into a full JSON caption for generation.
- Ideogram 4.0 supports structured prompting using JSON with bounding-box and hex palette definitions.