Ideogram 4.0q

For three generations Ideogram was the closed benchmark nobody could touch on text-in-image. Posters with legible paragraphs. Brand logos that don't melt. When it finally released weights as Ideogram 4, the catch was the size: the only checkpoints that shipped are quantized fp8/nf4, and they still want a serious GPU to run comfortably. This node is the end run around all of that - it's Ideogram 4 running on Runware's cloud (the q is the quantized checkpoint, ideogram:4@q), so you get the best text renderer in the business with zero local VRAM and zero download. You just pay per image.

There's one thing you need to understand before you queue anything: Ideogram 4 is not prompted like a normal model. It was trained on structured JSON captions, not prose, and plain natural-language prompts underperform badly. That's baked into this node's interface. The one required input isn't a prompt - it's structuredPrompt.high_level_description, a one-or-two-sentence description of the full image including subject and style. That's the heart of the structured schema, and the rest of the structuredPrompt.* fields let you slice the description finer.

The inputs that matter

• structuredPrompt.high_level_description (required) - your core sentence(s). This is the spine of the structured prompt; spend real time here.
• structuredPrompt.style_description.* - aesthetics, art_style, lighting, medium, photo. Fill in the ones that matter: art style for non-photo work, lighting and mood, photographic style for realism. These become the JSON nodes Ideogram 4 was actually trained on.
• structuredPrompt.compositional_deconstruction.background - environment, framing, and lighting context.
• positivePrompt / negativePrompt - free-text extras. The negative is ignored when guidance is off (see CFG below).
• width / height - 64 to 4096, step 16, default 1024. Ideogram 4 handles wild aspect ratios from one set of weights.
• CFGScale - a toggle, off by default, because this model has usage rules around it. Flip it on to expose the CFGScale_value (default 7).
• settings.renderingSpeed - TURBO, DEFAULT, or QUALITY. Your speed/quality tradeoff; the obvious place to start tuning.
• numberResults - up to 20 variations, each with its own seed.
• outputFormat / outputQuality - JPG/PNG/WEBP, quality 20–99.
• advanced_json - the escape hatch for the structured fields this node doesn't expose as widgets (compositional_deconstruction.elements, style_description.color_palette). If you want bounding boxes and typed text elements, this is where the schema goes.

The output is a single image (IMAGE), straight into Preview or Save.

Why you'd run it here instead of locally

The KB's history is worth remembering: Ideogram 4 is a non-commercial release with a refusal image trained into the weights, and the community's fix was routing prompts through the structured JSON schema. On Runware you're calling the hosted model, not the downloaded weights - which sidesteps the quantized-only download, the VRAM question, and the license paperwork in one move. You trade that for per-image cost and an API key.

Installing it

This node is part of the Runware/ComfyUI-Runware pack. Install via ComfyUI Manager (search Runware) and restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model files, light deps (runware-sdk, pillow, soundfile). You do need a Runware API key - Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Where people get burned: they treat it like a normal text-to-image node and write a prompt, then wonder why the output ignores half of it. If your image has text or tight composition, work the structured fields. And remember it's a billed API - the title bar shows each run's cost, and the QUALITY rendering speed plus higher resolutions cost more.