OpenAI

IsekaiOpenAI is the rare node in this pack that tells the truth about itself. Unlike the image effects, which are all local PIL math, this one really does phone home: it calls OpenAI's Chat Completions API and hands the reply back to your workflow as a STRING. If you've ever wanted a GPT step mid-graph - rewriting a prompt, generating a title, producing a caption you'll feed to a save node - this is the simplest way to get it without building your own API wrapper.

How it works. The node POSTs to https://api.openai.com/v1/chat/completions with a standard messages array. Your prompt becomes the user message; system_prompt (optional) goes first as a system message if you provide one. temperature (default 0.7) and max_tokens (default 1000) map straight onto the API. Nothing fancy, no streaming, no vision - text in, text out.

The one mechanism worth knowing is the API key handling. The node checks the OPENAI_API_KEY environment variable first, and only falls back to the api_key input field if the env var is missing. When you use the field, it prints a warning that keys typed into node inputs get saved in your workflow JSON. It's right - and it's worth remembering that any node you load executes arbitrary code on your machine, so a key sitting in a workflow you drag in from the internet is a key you just gave away.

On failure, the node doesn't crash your graph. It returns an Error: ... string as the output. That's actually friendly for debugging, just don't wire it into a text encoder and then wonder why your prompt says "Error: Authentication failed."

The inputs that matter. You'll touch four of these, max:

• prompt - the actual thing you're asking. Multiline, so paste freely.
• model - a hardcoded dropdown: gpt-4o, gpt-4-turbo, gpt-3.5-turbo, gpt-4o-mini. This is the node's biggest weakness: you can't type an arbitrary model name, so when OpenAI ships the next thing, you're stuck on this list until the pack updates.
• system_prompt - set the persona ("you are a terse caption writer").
• temperature / max_tokens - leave alone until the output is too boring or too short.

The single output is response (STRING), which wires straight into a CLIP text encode, a text display node, or a save-text node. A natural setup: batch a bunch of images, pass a per-image description in, and get back captions you can write to disk.

Install. Same as the rest of the pack: open ComfyUI Manager, search "isekai", install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/isekai-sh/isekai-comfy-node
cd isekai-comfy-node
pip install -r requirements.txt

Then restart ComfyUI. requirements.txt is just Pillow, requests, and the torch/numpy you already have - no model downloads, nothing heavy. Set OPENAI_API_KEY in your shell (or a .env you load before launching ComfyUI) rather than pasting it into the node.

Common issues. Empty prompt returns an empty string rather than an error - it's a silent skip, so if you wired a possibly-empty value in, that's what's happening. A wrong or expired key returns an error string in the output (check the [Isekai] OpenAI lines in the console for the real message). Rate limits get retried three times by the shared client before failing, but on a busy account you can still eat an error - bump max_tokens down and batch less aggressively. And one more honest note: this pack has essentially no community footprint - a search for it turns up nothing on the usual forums - so you're the early adopter. The node works, but "works" here means "the README and the code agree," not "hundreds of threads complain when it breaks."