Vector Engine GPT

Vector Engine GPT is the sibling node in the same pack as Vector Engine Gemini, and it does the same trick for a different model family. Where the Gemini node proxies Google's image API, this one fronts OpenAI's images API through the same third-party service at api.vectorengine.ai. One prompt, up to five reference images, and you get a finished image back as a normal ComfyUI tensor - no local GPU work, no model downloads.

You'd reach for it for the same reason you'd reach for GPT Image anywhere: it's stubbornly good at text rendering and at following finicky instructions, and the edit path (feed it images, tell it what to change) is one of the few places a closed model genuinely outclasses local diffusion. If you already run the Gemini node from this pack, adding GPT support costs you literally nothing except a second environment variable.

How it works

The mechanism is the OpenAI images API, shape and all. With no input images, it POSTs a JSON body to /v1/images/generations with your model, prompt, size, and n: 1. Give it any of image_1 through image_5 and it switches to /v1/images/edits - a multipart/form-data upload, images encoded as PNG and downscaled to 2048px. It reads the response as either a URL or b64_json, decodes it to a IMAGE tensor, and hands you an info STRING with the same diagnostics the Gemini node provides: model, size, resolution, per-stage timings, status, and the error message when things go sideways.

Inputs and outputs that matter

• model - gpt-image-2 (default), gpt-image-1.5, or gpt-image-1. Default is fine; step down to the older models only if you're chasing a specific behavior or price point.
• prompt - multiline. This is the whole game for editing: be explicit about what stays and what changes.
• size - auto, 1024x1024, 1536x1024, or 1024x1536. Note the difference from the Gemini node: no aspect-ratio presets, just four fixed canvas sizes.
• seed - same trap as the Gemini node: display-only. It never reaches the API, so don't use it to reroll.

Outputs: image → wire it to Preview Image / Save Image; info → read it when you get a black result (more below).

Installing

It ships in the same pack, so there's nothing new to fetch if you already have ComfyUI-Vector-Engine (Manager: search "ComfyUI-Vector-Engine", or git clone https://github.com/dzy1128/ComfyUI-Vector-Engine into custom_nodes, restart). No requirements.txt, no model files - just stdlib plus the torch/numpy/Pillow ComfyUI already ships.

The catch the README won't warn you about: this node isn't in the README at all. The pack's docs only cover the Gemini node, and the GPT node is code-only. Which is exactly why people trip on the key. It doesn't use VECTOR_ENGINE_API_KEY like its sibling - it reads a separate VECTOR_ENGINE_API_KEY_GPT, and throws a VECTOR_ENGINE_API_KEY_GPT environment variable is not set error if you launch without it:

export VECTOR_ENGINE_API_KEY_GPT="sk-your-other-key-here"

Common issues

The same failure pattern as its sibling: a black 512×512 image means it failed, and the reason lives in the info string - authentication, network, or "no image data in response." Check that before touching anything else. 20–30s generation time is normal API latency, not a hang. And the usual expectations apply: input images get downscaled to 2048px before upload, and the node is just a wrapper around a reseller proxy that has essentially zero community footprint - so your key and prompts are in this service's hands. Fine for occasional text-heavy edits; worth knowing before you build a whole pipeline on it.