📦大炮本地多轮对话素材库@炮老师的小课堂

DapaoLocalChatMaterialLibrary ("大炮本地多轮对话素材库") is the staging area for the chat workbench. It registers up to 20 images, 5 videos, and 5 audio clips as numbered materials, and hands the whole bundle to DapaoMultiTurnChatV2 as a DAPAO_LOCAL_CHAT_MATERIAL_LIBRARY object. That's the whole job - and the job has one rule you must internalize, because it's the difference between this node being useful and this node being a memory hog.

Connecting media to the library does not put it in context. The README is explicit: the library is just the prep area. Only the materials you explicitly reference in this turn's chat message with an @ - @图片1, @视频2, @音频1 - get processed and sent to the model. And history stores only material numbers plus text analysis, never the raw media again; to have the model re-look at something, you re-@ it.

That design exists for a reason. If all 20 images auto-loaded into context every turn, you'd blow your context window in one message. Selective injection is what keeps the workbench usable on a consumer card - the same VRAM-budgeting logic that runs through the whole pack.

What each input does

• 🏷️素材别名 - an optional JSON dictionary mapping names to friendlier labels, e.g. {"图片1":"产品正面","视频1":"开场镜头"}. The fixed numbers are the internal stable IDs; aliases are just for you and the model.
• 🖼️图片1…🖼️图片20 - each accepts a single IMAGE, per the tooltips. One wire per image; no batching in one port.
• 🎞️视频1…🎞️视频5 - VIDEO inputs. Frames are only extracted when that video is @-referenced this turn.
• 🎵音频1…🎵音频5 - AUDIO inputs. Only compressed to 16kHz mono WAV when referenced.

Mechanically, the chat node does the heavy lifting downstream: images are Lanczos-scaled to a max edge (the 2K default from chat settings) and sent as PNG, videos get representative frames, audio becomes the 16kHz mono WAV. Whether audio actually reaches the model depends on whether your GGUF + llama.cpp ChatHandler supports input_audio - if not, the node tells you clearly instead of failing silently.

Where people get burned

Two traps. First, the "it's connected but the model can't see my image" panic - that's the @ rule. Second, image/video chat requires a matched mmproj on the model loader; without it, an @图片1 turn errors out with a message telling you to load the vision projection. If you're doing pure text chat, none of this matters - leave the library unwired and skip the node entirely.

Installing it

Standard pack install, once:

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

then requirements + restart. Pack-wide gotchas apply: llama-cpp-python from the JamePeng fork (0.3.47+ for Qwen3.8; stock PyPI 0.3.35 errors on its GGUFs), GPU wheel overlay for NVIDIA, models in ComfyUI/models/LLM/. The library itself holds no model, but it's dead weight without DapaoLocalModelLoader and DapaoMultiTurnChatV2 on the canvas beside it.