Gemini 3 聊天

Most LLM nodes are one-shot: prompt in, answer out, memory wiped. Gemini3Chat is the one node in this pack that's built for a conversation. It takes a message and a chat_history, returns an answer and an updated history, and if you wire that history back into the same node you get a real multi-turn loop - the kind of thing that's surprisingly rare in ComfyUI, where everything defaults to stateless plumbing.

You'd reach for this when your workflow is a dialogue, not a single translation. A workflow that asks Gemini to critique a prompt, then asks follow-up questions about its own critique. An agent-style step where the model needs to remember what it decided three turns ago. The whole point is that the memory lives in the updated_history string, so the graph itself doesn't have to.

How it works

The node expects chat_history as JSON - an array of {"role": "user" | "model", "parts": [...]} entries. It parses that, appends your new message, sends the whole conversation to Google's generateContent endpoint, then rebuilds the history: your message goes in as a user turn, the model's reply goes in as a model turn, and critically, any thoughtSignature in the response is preserved in the history. That thought-signature carryover is what keeps the reasoning context coherent across turns instead of each message being a fresh start.

The loop is manual but simple: connect updated_history to the chat_history input of the next copy of the node, or run the node once per turn in a pipeline and thread the string through.

Inputs and outputs that matter

Required: message (what you say this turn), plus the usual api_provider, api_key, model, and thinking_level - same defaults as the rest of the pack (gemini-3-pro-preview, high thinking).

Optional:

• chat_history - the JSON from a previous run. Leave it empty for the first turn.
• system_instruction - set the persona once and it applies to every turn.
• max_output_tokens - cap on the reply, default 8192.

Outputs: response (the model's text), updated_history (the JSON to feed back in - this is the one that matters), and usage_metadata (token counts, useful when a long conversation is burning through your quota).

Install

Same pack as every Gemini 3 node here - ComfyUI Manager (search "ComfyUI-Gemini-3") or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI-Gemini-3
cd ComfyUI-Gemini-3
pip install -r requirements.txt

No models to download. You need a Google AI Studio key via GEMINI_API_KEY, the pack's config.json, or the api_key field on the node.

Common issues

• History never grows - if you don't wire updated_history back into chat_history, every turn starts fresh. That's the most common "why doesn't it remember" report, and it's expected behavior, not a bug.
• Corrupt history JSON - if chat_history isn't valid JSON, the node silently drops it and treats the turn as the first one. If a long conversation suddenly forgets everything, check that the history string wasn't truncated somewhere in the plumbing.
• It's a chat, not an agent. This node keeps a conversation alive, but it won't call tools on its own - for that, you want the pack's function calling node, and the two compose fine (a chat history can carry function-call parts).

The one real gotcha is cost: every turn re-sends the whole history plus a thinking-level high model's deliberation, so a 20-turn session is 20 billable round-trips. If you're iterating on a long persona doc, consider caching the system context instead - that's what the pack's context cache nodes are for.