DALL-E 3 Generator

This node does exactly one thing: it sends your prompt to OpenAI's DALL-E 3 API and drops the returned image into your graph as an IMAGE. That's it. The pack author says as much in the README - "nothing special, it was more for my own use a while back than for anything else." No local model runs here, no VRAM gets touched, no ComfyUI sampler is involved. You're paying OpenAI per call and getting back a finished image you can then pipe into the rest of your workflow - upscale it, composite it, use it as an img2img source, whatever.

How it works

The node reads your OPENAI_API_KEY from the environment - there's no key field on the node itself, so it has to already be set wherever ComfyUI is running before you hit Queue. It builds a request from your prompt plus the three options below, sends it to OpenAI, and downloads the resulting image into a standard ComfyUI IMAGE tensor. Worth knowing: OpenAI's DALL-E 3 API silently rewrites your prompt for safety and consistency before generating - you're not always getting exactly what you typed, that's an OpenAI-side behavior baked into the API itself, not something this node adds or can turn off.

The inputs and outputs that matter

• prompt - multiline text, your actual description. This is the one field you'll spend time on.
• image_size - one of 1792x1024 (landscape), 1024x1024 (square), or 1024x1792 (portrait). DALL-E 3 only supports these three; you can't request arbitrary dimensions.
• image_quality - standard or hd. HD costs more and takes longer per OpenAI's own API, for finer detail and more consistent output.
• style - vivid or natural, and this node defaults to natural. Vivid pushes toward hyper-real, dramatic, saturated images; natural aims for a more realistic, less amped-up look. If your outputs feel oversaturated and "AI-generated" in that glossy way, that's vivid doing its job - switch to natural.
• Output is a single IMAGE - wire it wherever any generated image would go.

How to install it

Via ComfyUI Manager: search "comfy_PoP", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/picturesonpictures/comfy_PoP

The clone gets you the node; it doesn't get you API access. Separately, set OPENAI_API_KEY as an environment variable on the machine ComfyUI runs on - an OpenAI account with billing enabled and a generated API key. There's nothing to download model-wise; every generation happens on OpenAI's servers, not yours.

Common issues & troubleshooting

Node errors out immediately, no image. Almost always the API key. If OPENAI_API_KEY isn't set in the environment ComfyUI actually launched from - not just a terminal you had open once - the request never authenticates. Set it, then restart ComfyUI so the process picks it up.

Prompt got rejected or came back different than expected. DALL-E 3's API has its own content moderation and prompt-rewriting layer on OpenAI's side. If a request fails outright, it's usually a policy rejection; if it succeeds but doesn't match your wording, that's the automatic prompt rewrite. Neither is something this node controls.

It's slow, or costs more than you expected. Every single generation is a paid, network-bound API call - there's no caching, no local fallback, and hd quality costs more than standard. Check OpenAI's current API pricing before you run this in a loop or a batch; it adds up fast compared to a local checkpoint.

No internet, or airgapped ComfyUI install. This node simply won't work - it needs to reach OpenAI's servers for every call, unlike the rest of comfy_PoP, which is all local, offline logic.