ComfyUI-Ideogram-Autoprompter

ComfyUI Ideogram 4 Autoprompter

An AI autoprompter node for Ideogram 4's structured JSON caption format. Describe an idea (and/or drop a reference image) in the node, click Generate, and a vision LLM builds the full caption — background, placed elements with bounding boxes, descriptions, rendered text, and color palettes. Everything stays editable on a visual black-and-white canvas afterward. At workflow run-time the node only passes the assembled caption JSON string through (plus a preview image).

Install

Copy this folder into ComfyUI/custom_nodes/ and install requirements into the ComfyUI Python environment:

pip install -r requirements.txt

torch is provided by ComfyUI. For the optional 4-bit local-model toggle, also pip install bitsandbytes (CUDA only); without it the node simply loads in full precision.

Engines

• Local (default): huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated, downloaded automatically on first generate via transformers. The model is unloaded after each generation by default (toggle "unload after"). You can point the model-id field at any compatible Qwen3-VL checkpoint.
• Gemini (free API key): paste a key, click Fetch models, pick one, then Generate. The API key is kept in memory for the session only — it is never saved into the workflow.

Usage

1. Add the Ideogram 4 Autoprompter node (category Ideogram/text).
2. Pick an engine, type an idea and/or drop a reference image, click Generate.
3. The canvas and fields populate from the generated caption. Edit freely: draw/resize/move regions, set type (obj/text), descriptions, text, and palettes.
4. The prompt output is the caption JSON string; preview is a rendered layout image.

Generation happens on the node UI (via the button), not during graph execution.