Protoface Ideogram V4 · Text to Image

Posters, logos, and the word "Grand Opening"

Text in generated images used to be a punchline. Through three generations, Ideogram was the lab that took it seriously - the benchmark for typography, logos and graphic design, founded by the people who wrote the diffusion papers everything else is built on. If your output has to contain a legible headline, a brand name, or a menu, this is the model family you'd reach for.

Ideogram 4.0 also opened its weights in June 2026, non-commercial and quantized - which creates a genuinely confusing situation this node sits in the middle of. The downloadable weights are licensed for research and personal use only. The hosted API is the commercial route. So: same family, two very different permissions, and you're on the rented side here, which is the legal-clean one and the one that comes with Ideogram's moderation attached.

This is the Text to Image task from Protoface, the beta pack that exposes hosted models as native ComfyUI nodes. Paid per generation through your Protoface account.

How it works

Schema-driven node, shared adapter: payload validated against the provider's public request, submitted against your key, polled, downloaded, returned as native IMAGE + MASK. Nothing local, no VRAM.

The one design choice worth flagging is that this same node does both text-to-image and image-to-image. There's no separate edit node: you get an image input, a strength control that only applies when an image is connected, and off you go. Simple, and slightly easy to miss.

generation_id is your paid-request identity as usual - fixed to reuse or resume, incremented to buy a new generation.

The inputs that matter

prompt - the tooltip says something useful: plain language is fine, and Ideogram expands it into its own structured caption format internally. So write "a matchbook for a jazz club called The Blue Room, 1960s, printed on cream card, centered" rather than learning a JSON schema. Put the actual text you want rendered in quotes.

quality defaults to medium and maps to sampling steps - low for fastest, high for best detail. This is where your money goes. Text rendering is the first thing to degrade at low quality, so if you're doing typography, don't cheap out here.

resolution is a tier by total pixels: 1k ≈ 1 MP, 2k ≈ 4 MP. aspect_ratio defaults to 1:1. output_format defaults to jpeg - switch to PNG if you're going to composite or upscale, because JPEG artifacts around sharp type are exactly what you don't want.

image + strength (0.8 default) are the image-to-image pair: lower strength stays closer to your source, higher changes more. Only meaningful with an image connected.

Outputs

image (native IMAGE, list), mask (native MASK, list, inverted-alpha semantics), files (PROTOFACE_FILE list of originals) and run id (string, for Protoface Resume Run).

The files output is worth knowing about: it holds the original download, so if the provider returns a transparent PNG or an artifact ComfyUI's conversion would flatten, that's where the untouched version lives.

Install

Manager → search Protoface → install protoface-nodes → restart → reload. Manual, from the ComfyUI directory:

git clone --branch v0.1.0 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git custom_nodes/protoface
python -m pip install -r custom_nodes/protoface/requirements.txt

ComfyUI 0.41.0+, Python 3.10+, installed with ComfyUI's own interpreter. Then click Protoface in the lower-right corner and enter a key from your Protoface account - this is billed directly and does not use Comfy credits.

Where people get burned

Confusing this with the open weights. Running the Ideogram 4 checkpoint locally is a different thing with a different licence (non-commercial, no commercial render without an agreement). This node is the API product; the licence question is Ideogram's to answer, not yours.

Cheaping out on quality for a logo. Low/medium will give you a beautiful mark and a slightly wrong letterform. If the text matters, pay for the tier and inspect at 100%.

Expecting a text-only poster to come back clean. Filters are real on the hosted path, and there's no local patch for a closed endpoint. If a prompt gets refused, it's the provider's policy, not the node.

Rerolling with the same generation_id. Cached job, same output, and you'll think the node ignored you. Increment to spend; keep it fixed to recover.