Hi everyone,
I wanted to share a ComfyUI custom node I recently developed for Ideogram 4. This node helps you visually design layouts and arrange prompts more intuitively right inside ComfyUI.
I took a lot of inspiration from KJ Nodes and LTX Director while making this.
Key Features
Ideogram 4 JSON Prompt Builder: Generates structured JSON prompts optimized for Ideogram 4.
Visual Bounding Box Editor: Draw, resize, and modify layout zones directly on the canvas.
Element Organization: Easily organize elements like text, logos, signatures, and titles by specific areas.
External JSON Loading: Import prompt drafts from other nodes, like the Local LLM Loader.
Safe Mode: Asks for confirmation before replacing an existing board with a new JSON, preventing accidental overwrites.
Error Resilience: Displays a warning instead of crashing your workflow if an invalid JSON is entered.
Translate On/Off Helper: Assists with converting prompts into English.
Text Preservation: Keeps the actual text content inside the TEXT field exactly as it is, even during translation.
How to Use
Install Deno Custom nodes via ComfyUI.
Set your desired canvas aspect ratio and resolution.
Arrange your visual zones by placing bounding boxes on the canvas.
Input the role, description, and exact text content for each box.
Use the generated Ideogram 4 prompt and bbox data in your workflow.
(Optional) Connect it with a Local LLM Loader to automatically bring in prompt drafts.
I'm also sharing the workflow used in the demo video. If you're using it for the first time, just load this up, tweak the box positions and text, and you’ll get the hang of it pretty quickly.
The main focus of this node isn't about writing "longer prompts"—it’s about visually mapping out your image components and organizing the structured data for Ideogram 4.
Let me know if you have any feedback or ideas for improvement!
-------------------------------------
Description
I've updated the workflow.
FAQ
Comments (17)
can something like this (prompting for specific regions of the image) work on anima & sdxl models or just ideogram?
thank you for your hard work
How do you write the prompt ? it looks different than when i use KJ node
You can either use an LLM or enter the prompt directly in the Director node to generate it. The usage is basically the same as the KJ node.
Oh man.... I look at everyone's workflows and nodes. This is Top Notch work. excellent job 👍
The Deno local LLM Loader offloaded to my cpu. Do you by chance know how to ensure it stays on the 16gb graphics card?
Hi. Which backend are you using? This doesn’t look like a workflow or node issue, it looks like the LLM model you’re using is just too large.
The LLM I’d recommend is Gemma 4 12B QAT.
@Deno2026 I'm using ollama with huihui_ai/Qwen3.6-abliterated:27b
The reason why CPU offloading takes place is because you chose a model that is too large.
Choose another model.
@Deno2026 just gave Gemma 4 12B a shot and it is much faster! Thank you!
@Deno2026 Have you encountered this error while utilizing it?
Local LLM run failed. The local model returned thinking text but no final result. Turn Thinking off, or ask the model to write a final answer after thinking.
@JBsharp It looks like the LLM model you downloaded doesn't support think mode. Try downloading a different model that officially supports it.
Local LLM run failed. Local LLM server returned HTTP 400: {
"error": {
"message": "Model 'gemma-4-12b-it-qat- ' does not expose reasoning configuration.",
"type": "invalid_request",
"param": "reasoning",
"code": "invalid_value"
}
}
It looks like the LLM model you downloaded doesn't support think mode. Try downloading a different model that officially supports it.
Amazing!