Skip to content
AI Primer
workflow

Ideogram 4.0 supports bounding-box layouts in ComfyUI

A creator showed Ideogram 4.0 using structured JSON and drawn bounding boxes in ComfyUI to lock composition and text placement, then extended it through Claude Code and ComfyCloud MCP for exact-pixel variations. The posts say the model needs 12 turbo steps and the Kijai prompt-builder node is still being updated, so users should expect an evolving workflow.

3 min read
Ideogram 4.0 supports bounding-box layouts in ComfyUI
Ideogram 4.0 supports bounding-box layouts in ComfyUI

TL;DR

  • Ideogram's model card says Ideogram 4 was trained around a structured JSON prompt format with explicit bounding-box and color-palette controls, and hellorob's ComfyUI demo shows that schema being used as a layout tool instead of a normal text prompt.
  • In hellorob's original post, the workflow hits usable results in 12 turbo steps, with composition locked by drawn boxes and text placement that holds up even in a zoomed demo.
  • hellorob's follow-up workflow pushes the same JSON prompt into Claude Code and then through ComfyCloud MCP to generate subject and color variations while keeping the layout at the same pixel positions.
  • The stack is still moving fast: hellorob's node update reply says Kijai was updating the node daily, while hellorob's ComfyCloud reply said the full workflow would be shared once the node lands in ComfyCloud.

You can read Ideogram's technical post, check the model card, and see that Hugging Face's Diffusers docs already treat structured prompt expansion as the native path. There is also an active KJNodes issue around bounding-box behavior, which matches the feeling in the tweets that this workflow is stabilizing in public.

Bounding-box composition

Ideogram 4's official materials frame layout control as a first-class feature. The technical post says the model was trained on structured JSON captions with optional per-element bounding boxes, and the model card calls out explicit bounding-box layout control and multilingual text rendering.

That shows up in the ComfyUI workflow as a design surface. Instead of nudging composition with prompt phrasing, hellorob's original post describes drawing boxes to place elements exactly, then iterating seeds on the turbo path.

JSON as the prompt builder

The interesting bit is not just that Ideogram accepts JSON. The Diffusers documentation says even a plain text prompt is meant to be expanded into Ideogram's native structured schema, either through Ideogram's hosted magic-prompt API or a local prompt-expansion path.

In hellorob's variation workflow, that schema becomes portable. hellorob tweaks the bbox layout in ComfyUI, brings the structured JSON into Claude Code, then uses ComfyCloud MCP to swap subjects, descriptions, and hex colors while the layout stays fixed.

Turbo steps and moving nodes

Ideogram's own available models page pitches 4.0 for posters, interface assets, and other design-heavy work, but the tweets add the practical constraint: this particular ComfyUI path is still being assembled in public.

One reply says Kijai was updating the node every second, and another says the workflow would be shared on ComfyHub once the node becomes available on ComfyCloud. A GitHub issue on KJNodes from the same weekend shows users already discussing how the bounding-box canvas should map to image dimensions.

Share on X