🐠GPT-LLM智能对话@炮老师的小课堂

DapaoGPTLLMChatNode is the "just let me talk to a frontier LLM" node in ComfyUI-dapaoAPI. No Skills, no material library, no state loop - you give it a system prompt, a user message, up to 8 images, and an API key, and it returns the model's answer as plain text. It's the node you reach for when you want a quick vision-language job - caption this image, summarize this screenshot, rewrite this prompt - without assembling the whole multi-turn workbench.

It sits in the pack's "大炮AI主力维护" (primary-maintained) category, which tells you the author considers it stable enough to lean on.

The inputs that matter

• 🔑 API密钥 + 🤖 模型 - dapaoAI key and the 11-model menu (default gemini-3.7-flash). Same relay as the rest of the pack: requests go to api.dapaoai.com, and the key is never written to a config file.
• 🎯 系统角色 - defaults to "you are a professional, friendly, accurate AI assistant." Change it per job; this is where 90% of the node's power lives.
• 💬 用户输入 - your message. Reference the images in prose; the node just packs them alongside.
• 🖼️ 图像1–图像8 - optional IMAGE inputs. Plug any of them in and it becomes a vision call. Eight is a lot of context for a caption job, but for comparing multiple screenshots or product angles it's genuinely useful.
• 🌡️ 温度 (0.7), 📝 最大输出令牌 (2048), 🎲 Top_P (1), 🎲 随机种 (cache-only), ⌛ 请求超时 (300) - the standard tuning set.
• 📋 额外参数JSON - escape hatch for OpenAI-compatible parameters the node doesn't expose as fields. Handy, but it's a free-form string, so a JSON typo surfaces as a request error.
• 🚫 出错时跳过 - on failure, emit an empty/skip result instead of failing the whole queue. Good for batch pipelines where one bad image shouldn't kill the run.

Outputs

• 💭 AI回复 - the answer, as a plain STRING. This is what you wire downstream.
• 📄 完整响应 - the raw API response, for when you need metadata the node strips from the friendly output.
• ℹ️ 处理信息 - the node's own status/telemetry, for debugging.

Where it fits

This is the classic "LLM as a tool in the graph" pattern from llm-in-comfyui.md - the model is not your prompt encoder, it's a worker that runs before or alongside generation. Common setups: feed it a rendered image from your KSampler and have it write a caption or a follow-up prompt; chain it into a text node that feeds another generator. It has no notion of conversation history - every call is stateless. For actual multi-turn work you want the DapaoAPIMultiTurnChatNode workbench instead.

Gotchas

• Random seed is cache-only. Bump it and ComfyUI re-runs, but the API result isn't pinned to it - matching behavior across the whole pack.
• Not every model in the list is multimodal. If you wire in images and pick a text-only model you'll get an error or a refusal-style answer, depending on the relay. The gemini-3.x and gpt-5.x entries are the safe picks for image input.
• It's a paid call every run. No auto-retry of paid POSTs - the pack philosophy is that a flaky connection costs you a manual re-run rather than risking a double charge.
• Chinese-only labels, as with the entire pack. The defaults are sane enough that you can ignore most of them and just edit the system prompt and user message.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git into ComfyUI/custom_nodes/, pip install -r requirements.txt, restart. Get a dapaoAI key at api.dapaoai.com (register, buy a redeem code, apply for a key in the default group) and paste it in.