OpenAI GPT Image 2

The node's display name says "OpenAI GPT Image 2," its class says OpenAIGPTImage1, and ComfyUI's own source marks it deprecated. Welcome to the joys of fast-moving partner integrations. What matters for you: this is the node that talks to the GPT Image family - the natively multimodal successor to DALL·E - and despite the confusing name and the deprecation flag, it's still the straightforward way to get OpenAI's current image models into a ComfyUI graph. ComfyUI just shipped a newer, tidier GPT Image node (the "V2" of the node, naturally), so think of this one as the older door to the same room.

Why GPT Image at all? It's OpenAI's answer to everything the newer hosted image models do: it reads your image, follows a natural-language edit instruction, renders text legibly, and doesn't need a mask to edit - though it does support one. It's also the "image generation that went viral inside ChatGPT" technology, so the quality bar is real.

The inputs that matter

• model - gpt-image-1, gpt-image-1.5, or gpt-image-2 (default). Newer is generally better; 1.5 was the December 2025 refresh, and 2 is current.
• prompt - generation or edit instruction. With no image attached it's text-to-image; with an image it edits.
• quality - low / medium / high. This is a cost dial more than a quality dial: higher quality means more tokens, more money, more time. Start low, go high for the final.
• background - auto / opaque / transparent. Transparent output is a genuinely handy feature - generate a logo or asset straight onto alpha instead of removing a background later.
• size - presets from 1024² up to 3840×2160, or Custom to unlock custom_width / custom_height (both multiples of 16, GPT Image 2 only). This is how you get 4K-ish output from OpenAI.
• n - 1–8 images per run.
• image + mask - optional. Attach a mask and white areas get replaced (same contract as DALL·E 2, but both are optional here - a plain image alone is fine for editing).
• seed - again "not implemented yet in backend." Ignore it.

Cost, setup, and the honest read

Ships with ComfyUI core, needs a Comfy account with credits and a current ComfyUI, login from localhost. GPT Image bills by API tokens rather than flat per-image, and output-image tokens are the pricey part - roughly 7,600 credits per million output tokens on gpt-image-2, which lands a typical image in the low tens of credits. The price badge on the node approximates it by quality; watch it before you queue high + custom 4K.

The deprecation flag deserves one sentence: it means ComfyUI considers this node legacy and will eventually stop updating it, not that it's broken today. If you're building a workflow from scratch, the newer GPT Image node is the better bet going forward. But if you've got a saved workflow that references this class, it'll keep running - and honestly, for getting a current GPT Image render into a graph, it does the job without ceremony.