🍭llama反推额外选项@炮老师的小课堂

Dapao_LlamaCaptionOptions is the simplest node in this pack, and that's fine - simple is exactly what it's for. It's a bundle of 17 boolean toggles that becomes the 🍭反推额外选项 input on Dapao_LlamaCaption. Flip a switch, and the captioner's instruction prompt grows by one requirement. Nothing else happens in this node; there's no model, no inference, no VRAM. Its entire job is to hand the caption node a LLAMA_CAPTION_OPTIONS object.

The name is a lie, sort of - it doesn't "reverse-prompt" anything itself. It's a pre-flight checklist for the reverse-prompt node. Think of it as the difference between asking "describe this image" and asking "describe this image, and also tell me the lighting, the camera angle, the composition, the depth of field, whether there's a watermark, and rate how safe it is."

The switches that matter

All 17 default to off, and off is a fine starting point. Turn on only what your job needs, because each one makes the model spend tokens and sometimes contradict itself:

• 👤 包含人物信息 - people details; the flip side is 🚫 排除不可改变特征, which tells the model to skip fixed traits you can't prompt away.
• 💡 包含光照信息, 📐 包含相机角度, 📷 包含相机详情 - the photography cluster. Great for img2img seeds where you want the shot recreated.
• 🎨 包含艺术质量, 📊 包含构图信息, 🌈 包含景深信息 - composition and look.
• 🔍 排除性感内容 - the one that matters if your captioner is feeding a prompt box you'll re-roll: keep the caption SFW.
• 📝 不提及文字, 🔇 不提及分辨率, 🏷️ 包含水印信息, 🖼️ 包含JPEG伪影 - cleanup switches for dataset work. If you're captioning a training set, "don't mention watermark artifacts" is usually the one to leave off (you want it flagged as junk) - but for prompt seeding, on is better.
• 🌍 不使用模糊语言, ⭐ 描述重要元素, 🔒 包含安全性 - the rest of the control surface.

How to actually use it

Wire its 🍭llama反推选项 output into the caption node's 🍭反推额外选项 input. That's the whole hookup. Because every option is just an appended instruction, you can stack them - "人物信息 + 光照 + 构图" is a common combination for turning a still into a rich img2img prompt. And since it's a separate node, you can leave it on the canvas permanently with your preferred set and swap caption styles without redoing the checkboxes.

Installing it

It ships with the rest of paolaoshi/ComfyUI-llama_Dapao:

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

then install requirements and restart. No models needed for this one - it's pure config - but it's useless until you've also got the caption node working with a GGUF + mmproj in ComfyUI/models/LLM/. If your toggles don't seem to change the captions, first check that the LLAMA_CAPTION_OPTIONS wire is actually connected; an unplugged options node silently does nothing.