💓Llama批量提示词@炮老师的小课堂

Dapao_LlamaBatchPrompt is the pack's answer to a boring-but-real problem: you've got a folder of images and you need one prompt per image, all generated by a local model, without babysitting it. It's the same llama.cpp/GGUF inference core as the pack's chat node, but reorganized around bulk work. Feed it up to four aligned image groups (A/B/C/D), tell it what to do, and it walks the batch and hands you a list of prompts plus diagnostics.

The "对齐" (alignment) thing is the interesting part. The A/B/C/D groups are meant to be corresponding - A is the main image, B/C/D are supporting views or references for the same item. The node matches them by index, so image 3 in group A gets reasoned about together with image 3 in groups B/C/D, and the system prompt drives a per-item prompt. If your folders are ragged, a 🧩缺失处理策略 dropdown picks the behavior: strict error, reuse the single image, pad to the longest group, or ignore missing groups.

The three modes hiding in here

• With images: the A/B/C/D folder paths (or 🖼️A组图像…🖼️D组图像 IMAGE inputs) drive the batch. 🛡️多图模式最大提示词数量 and 🚦多图推理模式 (逐条推理 vs 单次批量请求) control whether each item is one inference or many images packed into one call.
• No images: 🔢无图默认数量 lets it generate N prompts from pure text. That's a freebie - a local batch text-to-prompt expander.
• The retry logic: 🛟失败重试次数 (0–5) and 🧪推理失败策略 (失败占位继续 / 跳过失败继续 / 任一失败中断). If you're batch-running 200 items overnight, "占位继续" keeps the list index-aligned even when a call dies, so the output list still lines up with your inputs. That's the detail that makes this node production-usable instead of demo-ware.

Inputs and outputs that matter

The 📝系统提示词 default is a real meta-prompt ("你是专业的批量图像编辑提示词专家…") and 🧾元指令 is the per-item instruction - both editable, both the key to making output match your base model's prompt style. 📏图像最大边长 (default 1024), 🎲随机种子, 🌡️温度 (0.7), 🔁重复惩罚 (1.1), 🧠思考模式 and the Qwen3.8推理强度 control all carry over from the chat core.

Outputs are 📝提示词列表 (a STRING list - wire it to a batch text node or splitter), 📄完整响应, and ℹ️处理信息 (the diagnostic output that tells you what got aligned, reused, skipped, or failed).

Where people get burned

The alignment strictness. If groups A/B/C/D have different file counts and you left the strategy on 严格报错, the node will refuse to run - that's not a bug, that's the safety net. Set the strategy deliberately, and know that "单图复用" means every item in the short group gets the same supporting image, which is fine for style refs and wrong for per-item identity refs.

Installing it

Same pack, one install covers everything:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-llama_Dapao

then pip install -r requirements.txt and restart. The recurring pack gotcha applies: llama-cpp-python is pinned to the JamePeng fork at 0.3.47+ (stock PyPI 0.3.35 chokes on Qwen3.8 GGUFs with missing tensor 'blk.64.ssm_conv1d.weight'), and NVIDIA users should overlay a matching GPU wheel from the JamePeng releases. Models live in ComfyUI/models/LLM/. For a vision batch you'll also want the mmproj matching your model - without it, only the no-image mode will work.