Ideogram 4 Prompt Builder KJ

When Ideogram 4 released its open weights in mid-2026, it came with a catch: it wants prompts in a structured JSON format, not the plain sentence you'd type into any other model. Writing that JSON by hand - regions, bounding boxes, text elements, color palettes - is miserable. Ideogram4PromptBuilderKJ is the node the community reached for within about a day of the model landing. It gives you a visual canvas where you draw the regions of your image, set what goes in each, and it assembles the correct JSON caption for you. It's a GUI for a format that was clearly designed for machines.

Why it exists

Ideogram 4's standout is text rendering and precise layout - putting words and objects exactly where you want them. To do that it consumes a structured caption describing regions on a normalized grid rather than a free-form prompt. This builder lets you lay those regions out visually: drag on the canvas to draw a box, tell it whether that box is an object or text, describe it, and the node handles converting your layout into the [ymin, xmin, ymax, xmax] coordinates on the 0-1000 grid the model expects.

The inputs and the canvas

The real interface is the on-node canvas, described in detail in the node's own tooltip: drag to draw a region, click to select, double-click to edit its description, right-click for a region list you can reorder (top = front). You can set each region's type, text, and color palette, drop in color swatches, and even use the last generated image as a live background to trace over.

The widget inputs back that up:

• width / height - the canvas aspect ratio, and the dimensions carried into the output.
• high_level_description - the overall scene description.
• background - the background prompt.
• style, plus aesthetics, lighting, and medium - the optional style fields that flesh out the look.
• bg_brightness - how bright the traced background shows on the canvas (a display aid, not part of the image).
• output_format, coord_mode, bbox_order - control how the assembled JSON is written (compact vs not, normalized coordinates, yx ordering).

There are optional inputs for an image to use as the canvas background, an import_json to load an existing caption back in, and bboxes. The outputs are the assembled prompt string (the JSON you feed the model), a preview image, the bboxes, and the width/height.

Installing KJNodes

In ComfyUI Manager, search "KJNodes for ComfyUI" in the Custom Nodes Manager, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt (portable: python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-KJNodes\requirements.txt), restart. The builder node itself needs nothing extra - but you'll separately need the Ideogram 4 model weights set up in your ComfyUI to actually generate.

Worth knowing

There's a bonus reason to use the JSON format that has nothing to do with layout: Ideogram 4 ships with a safety filter trained into the weights, and prompting through the structured schema - the format the vendor told you to use in the first place - is the community's noted way around a lot of its refusals. So the builder isn't just convenience; the format it produces is the one the model behaves best on.

The main friction is the canvas learning curve. It's a real little editor with its own drag/click/modifier vocabulary, so read the tooltip before you fight it. And the coordinate output only means something to Ideogram 4 - this is a purpose-built node for one model's prompt format, not a general prompt tool, so don't wire its JSON into an unrelated model and expect sense.