KIE Gemini 3 Pro (LLM) [SL]

The odd one out in this pack, in the best way. Where the other nodes generate pixels, this one thinks. It's a full Gemini 3 Pro chat node that can read the images, videos, and audio already sitting in your ComfyUI graph - which makes it the prompt-generator, captioner, or workflow-decision-maker for everything around it. Want a Kling prompt written from the frame you just generated? This is the node that writes it.

It calls KIE's OpenAI-style /gemini-3-pro/v1/chat/completions endpoint. You get three outputs: text (the reply), reasoning (Gemini's thinking trace, empty if you disable thoughts), and raw_json (the last response chunk for anyone who wants to parse the full payload). The reasoning_effort combo (low/high) trades thinking time for speed, and include_thoughts decides whether that chain-of-thought even comes back.

The inputs that matter

• prompt - your message, multiline.
• role - the message's role (developer, system, user, assistant, tool), defaults to user.
• images / video / audio - the party trick. Drop any ComfyUI media in and it's uploaded and attached to the message. Audio gets coerced to WAV, video to MP4, images to PNG.
• messages_json - a raw JSON array of message objects. If you set this, it overrides prompt, role, and all media inputs. Power-users who want multi-turn conversations or tool-call histories live here.
• enable_google_search - gives the model a search tool. It's mutually exclusive with response_format_json; the node errors if you set both.
• response_format_json - a JSON schema to force structured output, which pairs beautifully with the pack's Parse Prompt Grid JSON node.

The gotchas

stream defaults to true (SSE), and the node waits for completion before returning - so you don't get tokens trickling into the graph, you get the full answer at the end. If you want fewer surprise tokens, drop reasoning_effort to low. And the key-in-a-node caveat applies doubly here: this is an LLM, so your prompts and media genuinely leave the machine, and your key rides along in the saved workflow JSON. Fine for your own graphs; scrub it before sharing.

Install

Standard for this pack: ComfyUI Manager search "Kie-API-Serverless", or git clone https://github.com/v3devv/ComfyUI-Kie-API-Serverless into custom_nodes, then restart. No models to download - the intelligence lives on Google's (via KIE's) servers. The experimental flag in the source just means the model family moves fast; the node itself is straightforward.