多模态｜GPT-5.4 Mini 推理

If you want OpenAI's current reasoning model inside a ComfyUI workflow without your own OpenAI key, TikpanGPT5MiniResponsesNode is the relay wrapper for it: it calls gpt-5.4-mini through Tikpan's /v1/responses endpoint. The "Responses" part matters - this is OpenAI's newer unified API (the one that fuses chat, tools, and file inputs into a single response object), not the older chat/completions shape. The node's job is to surface the core parameters as ComfyUI fields and hide the fixed host, so you get image inputs, video frame analysis, file reading, web search, and a reasoning-strength dial without touching the API.

How it works

You supply a question, optionally images (图片1–4, up to 16 per the tooltip), image URLs, video frames (视频帧_IMAGE - it analyzes by frame-sampling), or file URLs/local file paths (small files inline as input_file). The node builds a Responses payload with your system prompt, reasoning strength, and detail settings, POSTs to the relay, and returns the usual five outputs: 回答文本, 优化提示词, 结构化JSON, 用量, and 状态日志. The 任务类型 presets (9 of them - 通用问答 through 安全合规检查) swap system-prompt templates so you don't write one from scratch.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 用户问题 - the question, combined with any media.
• 推理强度 - minimal / low / medium / high reasoning effort. This is the cost lever: higher = better at hard reasoning, slower, pricier.
• 回答详细度 - low / medium / high (default), a separate knob for answer length.
• 最大输出Token - 4096 default, up to 32768.
• 图片细节 - auto / low / high; resolution detail is a token cost.
• 抽帧策略 and 最大抽帧数 - the video-analysis controls (mixed-smart default, 12 frames max default), same family as the Gemini analyst node.
• 启用联网搜索 - off by default; flipping it on lets the model fetch current info, at extra cost.

Optional: 图片1–4, 图片URL列表, 视频帧_IMAGE, 文件URL列表, 本地文件路径.

Installing it

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

restart (or Manager → "Tikpan"). No models to download - pure relay client.

Where people get burned

• Reasoning strength is where the bill hides. Every "high" run on long input is a noticeably bigger charge. The 用量 output is there so you can actually watch it - check it after a few runs and calibrate.
• File size rules. Small files inline; the tooltip explicitly says big files should go on a CDN/OSS first and be passed by URL. Huge base64 payloads will time out.
• URL errors abort by default. URL错误处理 defaults to 严格报错 - switch to skip-bad-links for batch.
• Cache and retries. 复用本地缓存 (default on) avoids re-billing the same analysis; POST重试策略 (幂等键轻重试 by default) is the network-resilience knob.

If your pipeline wants OpenAI-grade reasoning on documents, images, and video frames with a per-call meter, this is the node - just treat 推理强度 like a budget slider rather than a quality slider, because that's what it economically is.