Ideogram P-Image

If the image has text in it, Ideogram has been the name to beat for years - accurate typography is its whole reputation. P-Image is Ideogram's fast, cheap tier of that, now callable straight from a ComfyUI graph as a partner node. You give it a prompt, it hands you one image, and the magic trick is that it can spell. For logos, posters, packaging mockups, or any job where the letters have to be exactly right, this is the node you reach for when you don't want to run the full open-weight Ideogram 4 locally.

What it is. A text-to-image call to Ideogram's hosted P-Image model, proxied through Comfy's partner-node infrastructure. One prompt in, one image out. Nothing downloads, nothing installs - the node is part of ComfyUI core, and the actual generation happens on Ideogram's side, billed against your Comfy account credits.

How it works. Beyond plain prompts, P-Image accepts an Ideogram 4.0-style structured JSON caption: exact colors as #RRGGBB hexes, exact text strings, and bounding-box layout. That's the power move here. Instead of hoping the model reads your mind, you hand it a caption that literally says "the headline is 'GRAND OPENING' in #FF4B2B at position (x,y,w,h)" and it renders it. The only catch is the prompt_upsampling toggle, which defaults to AUTO and rewrites short prompts into a detailed caption before generation. When you're supplying your own JSON or exact wording, set it to OFF so nothing mangles your carefully-built caption.

The inputs that matter. quality runs VERY_LOW/LOW/MEDIUM/HIGH, defaulting to MEDIUM - the everyday tier, and the floor for anything with difficult text; below MEDIUM, typography falls apart, which kind of defeats the point. resolution is 1K or 2K (16:9 gives 1280x720 at 1K, 2560x1440 at 2K) - combine HIGH with 2K for crisp letters. aspect_ratio covers everything from 1:3 to 3:1, and prompt_upsampling is AUTO/ON/OFF. The optional seed matters only when upsampling is OFF.

The outputs. Two: the IMAGE itself, and final_prompt - the prompt the image was actually generated from. That second output is a reproducibility cheat code: with upsampling ON/AUTO the rewrite varies every run, so the same seed gives you different images. Capture final_prompt, feed it back with upsampling OFF and the same seed, and you get the same image back.

How you get it. Built into ComfyUI core, no extension. You do need a Comfy account logged in with credits on it, and the cost scales with your quality tier and resolution - HIGH + 2K is noticeably pricier than MEDIUM at 1K, which is exactly how Ideogram nudges you toward the fast tier the model is named after.

Where people get burned. Forgetting upsampling is ON and then wondering why your seed is useless - check final_prompt. Supplying a JSON caption without flipping upsampling OFF, and watching it get rewritten anyway. And Ideogram's content-safety filter can simply refuse a generation; the node raises a clear error when it does, so don't read it as a settings bug. One more: this is a hosted API, so your prompt travels to Ideogram's servers - fine for marketing drafts, worth remembering if you're working on something sensitive.