多模态｜Gemini 3.5 Flash 推理

ComfyUI has a whole genre of "the LLM as a node" plumbing - captioning, prompt enhancement, workflow decision-making - and this is the closed-model API end of that spectrum, which the KB's LLM-in-ComfyUI doc covers as its §3 local-vs-API decision. TikpanGemini35FlashNode wraps gemini-3.5-flash (a long-document / complex-reasoning tier) through the Tikpan relay, and it's the most input-rich reasoning node in the pack: six image slots, plus URL lists for images, video, audio, and files, plus local file paths, plus web search and sandboxed code execution. If your graph needs "look at this video and these PDFs and tell me what to do next," this is the node.

How it works

The node builds a Gemini request (/v1beta/models/{model}:generateContent native, or OpenAI-compatible /v1/chat/completions if you flip 接口模式), attaches whatever media you supplied, POSTs to the relay, and returns the model's answer. Outputs: 回答文本 (the answer), 优化提示词 (a prompt-optimization pass), 结构化JSON, 用量 (usage - your billing receipt), and 状态日志. Small local files get base64-inlined; big ones should go on a CDN first and be passed by URL, per the tooltip.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 用户问题 - your question, combined with any attached media.
• 任务类型 - nine presets (通用问答, 图片理解分析, 视频/音频理解, 长文档/PDF总结, 代码与架构审查, 复杂任务规划, 提示词优化, 结构化JSON提取, 自定义) that swap the system-prompt template for you. This is the biggest time-saver in the UI.
• 思考预算 - the thinking depth: auto / off / 1024 / 4096 / 8192. Deeper = better at complex reasoning, slower and pricier. The README highlights the 8192 budget as the selling point.
• 最大输出Token - 8192 default, up to 65536.
• 创意温度 - 0.4 default; lower for fact work, higher for writing.
• 启用搜索工具 / 启用代码执行 / 启用URL上下文 - three booleans that unlock live web lookup, sandboxed Python, and URL reading. All off by default, all billable when on.

Optional: 图片1–6, 图片/视频/音频/文件 URL lists (one per line), 本地文件路径, and 高级自定义JSON - a deep-merge escape hatch that injects extra fields into the payload for relay/upstream features that haven't been surfaced in the UI yet.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart (or Manager → "Tikpan"). Pure network client; no local models.

Where people get burned

• Tool toggles are spend toggles. Web search, code execution, and URL context each add tokens and latency. Flip them on only when the task needs them.
• URL errors kill the run by default. URL错误处理 defaults to 严格报错 - one dead link aborts. Switch to 跳过坏链接并写日志 for batch robustness.
• Thinking budget is a cost knob. 8192-depth on every question is the fastest way to watch relay credits evaporate. Let auto handle most calls.
• Big local files. The inline path is for small files; the tooltip's advice (upload to OSS/CDN, pass the URL) is there because giant base64 payloads time out.

It's the workhorse multimodal-reasoning node of the pack - if your pipeline analyzes anything more complex than a single image, this is the one to start from. Just treat the power toggles as what they are: a meter with extra buttons.