Gemini 3 Pro (openai)

The Pro card is the expensive member of the Gemini family in this pack, served over an OpenAI-compatible chat-completions route. You don't reach for it to rewrite a caption. You reach for it when the text step is the part that has to be correct - a long brief compressed into one structured prompt, a validation pass that has to hold a schema, a piece of reasoning that decides what happens downstream.

What's different about the Pro card in practice

Two things you can see from the card itself. It's a Pro tier, so expect proportionally more per call and a slower round trip than the Flash cards. And KIE's own documentation for this page includes a googleSearch function alongside the usual example tool definitions - a search passthrough exposed as a tool. If your enhancement step is "find the actual correct name of this thing before writing the prompt," that's a real difference in kind, not just speed.

Everything else is the family pattern: text in, text out, one node per model version, no model dropdown to get wrong. The KB's LLM-in-ComfyUI section is worth a skim on one point - it argues reasoning-heavy models leak scratch-work into output. A Pro card is where you accept that trade knowingly: give it a system prompt that forbids commentary, and then read the text before wiring it into conditioning.

Inputs

prompt - required, multiline, the message.

system_prompt - here it's load-bearing. Long-context structured output is the reason to buy Pro, and the format contract is what makes it usable.

history_json - a JSON array of prior messages for multi-turn work. Nothing is stored between runs.

tools_json - function definitions. This is where you'd put a googleSearch-style tool, or your own function schema; the endpoint is OpenAI-shaped, so existing definitions drop in.

stream - off by default; on, the response arrives as a stream the pack normalises into one document. No incremental display in ComfyUI.

expert_override_json - raw body override, debugging only.

No image socket on this card.

Outputs

text is the STRING you wire into a prompt field or a text saver. raw_json is the entire provider response - go here first when the text output is empty or when tool calls happened and you want to see what was invoked. usage_json carries token accounting, which matters more on a Pro card than a Flash one. credits_consumed and credits_left are floats, and the node draws its own credits receipt after a successful run.

Install

Search the pack title in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart. Dependencies: requests, Pillow, numpy, PyYAML. Nothing runs on your GPU, nothing needs downloading - it's a thin Python client for a paid API.

Key once: Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. Validated before it replaces a working key; masked fingerprint plus CONNECTED when it takes. KIE_API_KEY in the environment overrides the saved key for servers and headless runs.

Gotchas

Don't leave a Pro node in a loop. This pack paints a credits receipt directly on the node after each run and exposes credits_consumed / credits_left as sockets. Use them. A Pro node firing on every queue inside a batch is how people discover their balance the hard way.

A 200 response with an empty text output usually means the model answered somewhere the extractor doesn't look - tool calls, for instance. raw_json settles it in seconds.

Generated node library. The pack builds its cards from KIE's API documentation at startup and re-checks daily, so the menu shifts with the provider, not with the plugin. If this card or a sibling version vanishes after a restart, that's upstream churn - the cached catalogue keeps you running when KIE's docs are briefly down, but it won't resurrect a retired model.