(Deno) Ideogram Director

Here's the thing nobody warns you about Ideogram 4: it doesn't prompt like any other model you own. It was trained on structured JSON captions - bounding boxes, hex color palettes, typed text elements - and plain natural-language prose underperforms badly. (Deno) Ideogram Director is a visual builder for exactly that format: you draw and edit bbox regions directly on the ComfyUI canvas, describe each region, and the node minifies the whole board into the JSON prompt Ideogram 4 actually wants. No hand-authoring JSON, no guessing the schema.

How it works

The board is your working surface. You set high_level_description (short whole-scene summary) and background, then place boxes on the canvas with their own descriptions. A style_mode of none, photo, or art adds a structured style block, with aesthetics, lighting, medium, photo, and art_style fields feeding whichever mode is active. The backdrop input is deliberately just a layout backdrop to trace over - it is not used for generation. Style and layout preset galleries give you starting points with lightweight previews.

Two features matter for real use. First, the Language view: you can read and edit board descriptions in your own language while the final output stays model-ready English - literal TEXT box words like signs and headlines are preserved exactly. translate_output and view_language control this, with Google as the default translation engine (or a custom LibreTranslate server). Second, import_json: you can feed in a JSON prompt from an upstream source - the README explicitly pairs this with the (Deno) Local LLM Loader - and import_mode decides whether it asks before replacing your existing board.

Outputs: prompt (the minified Ideogram JSON, wired into a positive CLIPTextEncode), plus width, height, seed, and bboxes (your drawn boxes as downstream BBOX data). There's also a seed_lock Fixed/Random control and optional auto_save with save_prefix for turning the latest generation into a saved file.

Why the format matters (and the filter thing)

Prompting Ideogram 4 through its documented JSON schema isn't just better quality - it's also the community's noted workaround when the model's trained-in safety filter objects to a prompt. The guard keys on prompt format rather than content, so a workflow that always emits structured JSON rarely trips it. Worth knowing if you run into a grey "refusal image" and wonder what broke.

Install and caveats

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI

The node builds the prompt; it doesn't run Ideogram 4 itself - you need the Ideogram 4 weights (non-commercial license, so check the terms before any commercial use) and a ComfyUI path for the model. And set expectations on speed: Ideogram 4 is slow - roughly 3–4 minutes per image on consumer hardware - so this node is a composing tool, not a sketchpad for fast iteration. Build the board carefully, then generate. For quick exploration you'll probably still reach for something faster and come back here to compose the keeper.