Gemini 聊天节点

The payoff node

This is the one that does the actual generation in this pack. Feed it the api_instance from GeminiApiLoader, a prompt, and optionally an image or a file, and it returns a plain STRING you can wire anywhere - into a CLIP text encode, a text display, a filename, a caption field.

Out of the box it's a prompt optimizer. The shipped default system_prompt is a prompt-rewriting instruction ("rewrite user inputs into high-quality Prompts… less than 200 words") with example outputs. That makes sense for 2026: the new LLM-encoder checkpoints take natural-language instructions, not tag soup, so a Gemini pass that polishes "big dragon, night" into an actual scene description can genuinely lift results. The default prompt is also opinionated - it tells the model not to dodge NSFW or gore and to describe it directly. Swap it out if that's not the vibe for your machine.

How it works

The node builds a contents array from your inputs and POSTs it to Google's generateContent endpoint (or streamGenerateContent if you flip stream), via raw httpx - no SDK. Every reply gets appended to a running messages list, which is what makes multi-turn work: wire history_json_out into history_json_in and the next run continues the conversation. The pack's multiturn.json example shows the exact wiring.

The inputs worth touching

• user_prompt - the main text. Leave it empty if you're only sending an image or file.
• system_prompt - the "人设" (persona), per the tooltip. Replace the default optimizer with whatever role you want.
• content_part_1/2/3 - three slots for GEMINI_CONTENT_ITEMs (text blocks, inline images, uploaded files). Chain a GeminiContentConnector if you need more than three parts.
• temperature / top_p / top_k - sampling. 0.7/0.95 by default; drop the temperature for deterministic rewrites.
• max_tokens - 2048 default. Thinking models chew through it fast; raise it for long output.
• thinking_budget - -1 disables thinking, 0 allows it with no budget, >0 sets one. Only meaningful on 1.5/thinking models.
• response_schema_json - structured output. Give it a JSON Schema and you get JSON back instead of prose. Three formats are accepted: a full schema object, a simplified {"name": "string"} object, or bare "name": "string" pairs. The companion comfyui-aichat-tools repo has schema helpers.
• filter_thoughts - strips the model's internal "thought" parts from the returned text. Leave it on; you usually want the answer, not the chain of thought.
• should_change - the sneaky one. True (default) means the node re-runs on every queue, because IS_CHANGED returns NaN so ComfyUI never caches it. Flip it false and the node caches until an input actually changes. Do that for slow or expensive calls.
• retries - 0–5 retries with backoff on transient API errors.

Outputs

response_text - the model's answer as a STRING. history_json_out - the full conversation as JSON, for multi-turn.

Troubleshooting

• Timeout on thinking models: raise the loader's timeout toward 600.
• Empty result: if the prompt and all content parts are empty, the node bails early with an empty string.
• JSON mode failing: the provider/model has to support structured output, and your schema has to match one of the three accepted formats - the error message ("无法解析 Schema") is actually specific.
• Your key and your prompts go to Google's servers. Keep live keys out of shared workflow files.