OpenAI DALL·E 3

DALL·E 3 is the model that made text-in-image almost boring - OpenAI's image model was the mainstream baseline for two years, and this node is how you reach it from a ComfyUI graph: a synchronous API call that returns a normal IMAGE tensor, no local weights, no install, billed per call against your Comfy account. It's one of the original partner nodes from the first cloud batch in May 2025, and it's showing its age next to the newer OpenAI GPT Image nodes - which is exactly why it's still useful to know about.

The honest take: DALL·E 3 is no longer OpenAI's headline image model, and ComfyUI's own GPT Image nodes are the ones people actually reach for now. But DALL·E 3 is cheap, fast, and famously good at following a written prompt to the letter. If you want the model that defined "type a sentence, get a correct image," it's right here.

How it works

The node posts your prompt to OpenAI's images/generations endpoint through Comfy's proxy and returns the image synchronously. The interface is deliberately small, which makes it a great first partner node to play with.

The inputs that matter

• prompt - the text for DALL·E. This is a prompt-following model, so natural language wins. Ask for exactly what you want, including the words you want rendered in the image.
• quality - standard or hd. HD roughly doubles the price per image and buys finer detail. Standard is fine for most jobs.
• style - natural or vivid. Vivid leans hyper-real and dramatic; natural stays, well, natural. The tooltip says it better than I can.
• size - 1024x1024, 1024x1792, or 1792x1024 (portrait and landscape crops of the same budget).
• seed - here's the funny one: the tooltip says "not implemented yet in backend." It's a real field that the node genuinely doesn't use. Set it and it changes nothing about the output - it only controls whether the node re-runs, like every partner node.

Output is IMAGE, ready to save or chain into the graph.

Where people get burned

• The seed does nothing. It's not a bug in your setup; the backend literally ignores it. If you're trying to reproduce an image by holding the seed, stop - DALL·E 3 has never been seed-reproducible through this endpoint.
• Expecting the newest OpenAI model. You'll get DALL·E 3, full stop. If you want GPT Image (or editing with an input image), that's the GPT Image node in this same family.
• No negative prompt, no references. This is generation-only. For edits you're in the wrong node.
• Cost discipline. HD plus a large size is the expensive corner; standard 1024 is the cheap one. For iteration, stay standard.