🐠RH批量LLM提示词@炮老师的小课堂

Here's the workflow this node exists for: you have a folder of reference images and you want an LLM to write one prompt per image for a downstream image generator or editor. Now add the twist that makes it genuinely hard - the reference images come in aligned groups (A/B/C/D), where A is the image you're writing for and B/C/D are the same row's context: the style reference, the product angle, the edited version. DapaoRHBatchLLMPromptNode is the RunningHub-pack node that does exactly this: reads A/B/C/D groups from folders or wired IMAGE inputs, aligns them, and emits a prompt per A-image, with concurrency, retries, and a failure policy so a 200-image batch runs unattended.

The inputs that matter

• 🌐 API渠道 + 🔑 API密钥 - the RH rule: domestic/international keys don't cross. This node calls the RH LLM endpoint (llm.runninghub.cn), which is separate from the image/video API - so the key is a RunningHub LLM key.
• 🤖 模型ID - 72 model IDs, default google/gemini-3.1-flash-lite-preview. The flash-lite default is the right call for a job where you want volume, not a flagship's price tag.
• 🎯 系统角色 - defaults to "you are a professional batch image-editing prompt expert" who strictly writes one prompt per aligned group. This is the prompt-discipline contract; leave it unless you know why.
• 🧾 元指令 - the meta-instruction block describing what each group's images mean and how to fuse them. The default is well-written - a good template to adapt rather than replace.
• 📂 A组文件夹 / B组 / C组 / D组文件夹 - read entire folders of images. This is the scale path: point at folders, not individual wires. A/B/C/D组图像 (IMAGE inputs) are the small-scale alternative.
• 🧩 缺失处理策略 - what happens when groups have unequal lengths (strict error default vs. tolerant fills).
• 🔢 无图默认数量 - how many items to produce when a group is empty.
• 🛡️ 多图模式最大提示词数量 - cap on prompts in multi-image mode (0 = no cap).
• 🚦 多图推理模式 - 并发逐条请求 (concurrent per-item requests, default) vs the serial alternative. 🚀 并发数 - 4 default. 🛟 失败重试次数 - 1. 🧪 推理失败策略 - 失败占位继续 (placeholder-and-continue) default.
• 🖼️ 发送图片最长边 (1024) and 🗜️ 发送图片JPEG质量 (85) - the compression that keeps token costs sane on image-heavy batches.
• 🧠 推理强度 - four levels, default none. Some RH LLM models support a reasoning pass; this dials it.
• Standard 🌡️ 温度 0.7, 📝 最大输出令牌 2048, 🎲 Top_P 1, 🎲 随机种 (cache-only), ⏱️ 超时时间 180, plus ➕ 额外参数JSON.

Outputs

📝 提示词列表 - a list of STRING, one prompt per aligned A-image (this is the deliverable), plus 📄 完整响应 and ℹ️ 处理信息 for auditing.

The honest read

This is a niche-but-real pattern - "aligned image groups → one downstream prompt each" is exactly what you'd hand-build with a loop and a VLM node, and this node packages the alignment, concurrency, retry, and failure policy so you don't have to. The llm-in-comfyui.md framing applies: the LLM is a translation layer between your aligned reference set and the downstream generator, and subject drift is the standing risk - which is why the default 元指令 insists on "only output the current item's prompt, one-to-one with image A." Keep that discipline in whatever you write.

Two cost realities: every item is a separate paid LLM call (72 model IDs means a wide price range - flash-lite exists so you don't pay flagship rates for 500 prompts), and the JPEG/1024 compression is your real money saver on image-heavy batches. Don't crank 发送图片最长边 up and wonder why the bill doubled.

Gotchas

• Unequal groups - folder lengths that don't match will trip the strict error strategy; set 缺失处理策略 to tolerant if you're running ragged folders.
• Placeholder-and-continue hides failures - the 提示词列表 will contain placeholder strings; audit 处理信息 before feeding the list downstream.
• Key mismatch - image/video RH keys vs LLM keys may be different credentials; make sure the key actually has LLM access.
• Chinese-only labels.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git → pip install -r requirements.txt → restart. RunningHub account + API key required.