NIX_ImageInference

NIX_ImageInference is the "tell me what this image is" node. It takes your generated image, sends it to an OpenAI vision model, and returns the model's answer as a plain text string you can feed anywhere - into a prompt, a filename, a comparison. If you've ever wanted a reverse-prompt loop ("look at what I just made, describe it, then I'll nudge the prompt and try again"), this is the node that closes the loop.

One honest caveat before you get excited: it's the only node in the NIX pack that needs the openai Python library, and its default endpoint is a third-party proxy, not OpenAI's official API. More on that below, because it's the thing most likely to confuse you.

How it works

Give it an image, a model_name, a base_url, an api_key, and a user_prompt (multiline). The node takes the first image in the batch, encodes it as a base64 PNG, and calls OpenAI's Chat Completions API with your prompt plus the image as a vision input. Whatever the model says comes back as prompt_word (STRING). It's a one-shot call - there's no retry, no batching, no image out.

The default model_name is o3-2025-04-16. Change it to whatever your endpoint supports; if you're using a different provider that speaks the OpenAI API format (most do these days), the node doesn't care.

The default proxy thing

Here's the trap, and it's straight from the README: the default base_url is https://api.openai-proxy.org/v1, a third-party proxy, not https://api.openai.com/v1. The README itself recommends switching to the official endpoint for stability and compliance, and says the proxy may fail. Treat the proxy default as "the author's convenience," not a recommendation. If you're sending real images anywhere, you should also think about where they're going - this node ships your pixels to whatever endpoint you set. For work you can't send to a cloud API, local captioners (Florence2, Joy Caption) are the private alternative.

Also note there's no signal on this proxy in the community corpus - nobody's talking about it, good or bad - so you're relying on the author's word that it works. Set your own base_url and api_key and test.

The few inputs that matter

• api_key - required. Empty key means a failed call, and the node returns the error as the output string rather than crashing, which is nice but easy to miss.
• user_prompt - this is what the model is told to do with the image. "Describe this image in one sentence" is a reasonable start; "output a comma-separated tag list" works great if you're feeding tags back into a positive prompt.
• image - remember, only the first image of a batch is used. Feed one at a time.

The output prompt_word is just text, so wire it into a text concat, a CLIP text encode, or a save node's filename. It doesn't show up anywhere automatically.

Getting it installed

The pack ships all nodes together, but this one has an extra dependency:

cd ComfyUI/custom_nodes
git clone https://github.com/J-ChenX/ComfyUI-NIX
cd ComfyUI-NIX
pip install -r requirements.txt    # installs openai, numpy, Pillow

Or use ComfyUI Manager (search "NIX" / "NIX ComfyUI Plugin", or install by Git URL https://github.com/J-ChenX/ComfyUI-NIX) - Manager handles requirements for you. Restart after installing. If the node returns "未安装 openai 库" (roughly "openai not installed"), you've skipped the requirements step - run pip install "openai>=1.0.0" in your ComfyUI environment.

Gotchas worth knowing

• It's not a batch node. One image in, one text out. Feed a batch and you only get the first frame described.
• Network failures return a string, not an exception, so a workflow "succeeds" with an error message as the prompt_word. Check your output before you build a whole pipeline on it.
• Third-party proxy by default. Swap base_url to https://api.openai.com/v1 and set your real key before you trust results.
• Zero community footprint. This pack has no meaningful Reddit presence as of mid-2026, so if the API call misbehaves, the error string it returns is your only breadcrumb - plus the openai library's own docs.