📦大炮API多轮对话素材库@炮老师的小课堂

Ever seen those chat boxes where the model is supposed to look at your reference images but there's no clean way to say which one, this round? DapaoAPIChatMaterialLibraryNode is the dapaoAI workbench's answer: a sideboard where you park up to 20 images, 5 videos, and 5 audio files, then reference them from the chat itself with @图片1, @视频1, @音频1 style mentions. Only the materials you tag in the current message get processed and sent. Untagged materials never leave your machine - which is both a privacy win and a token-cost win, because nothing is uploaded that the model isn't asked about.

How it works

You wire up to 20 🖼️图片N inputs (each expecting a single IMAGE), 5 🎞️视频N, and 5 🎵音频N - the tooltip for each video and audio slot spells out the deal: video is only keyframe-extracted and sent when the current chat round mentions @视频N, audio only gets compressed and sent on @音频N. The node hands the chat node a single 📦多轮对话素材库 output of custom type DAPAO_API_CHAT_MATERIAL_LIBRARY.

Then the chat node's message parser scans your message for @图片N / @视频N / @音频N references, matches them to the library, and builds the actual request parts. Two rules the README hammers home:

• Materials only travel in the round that tags them. If you want the model to re-examine something next round, you @ it again. This is deliberate - it keeps every call cheap and keeps history from silently carrying megabytes of image data.
• Images get Lanczos-scaled to a max 2048px long edge and sent as PNG, per-image, right at the request boundary (that cap is actually governed by the settings node). Videos get representative frames pulled by model capability; audio only goes to models that explicitly accept audio input.

There's also one optional input, 🏷️素材别名 - a JSON map like {"图片1":"产品正面","视频1":"开场镜头"} that gives your materials friendlier names. The tooltip notes fixed numbers are still the stable internal IDs, so aliases are just labels for the LLM's benefit, not renumbering.

Why you'd use it

The whole workbench is aimed at people building chat-driven generation pipelines - a model that looks at a product shot and emits an image prompt, or a video prompt, or answers questions about your assets. This node is what makes those workflows cheap to iterate on. Compare it to the blunt alternative (dump every reference into every call): the @-mention model means a 20-image library costs you nothing until you actually ask about an image. That's the difference between a toy and a pipeline that won't empty your wallet by accident.

Where it slots in

It's optional in the stack - DapaoAPIMultiTurnChatNode has a 📦素材库 input that accepts its output. When a material is selected but the mapped model can't take images, the chat node throws a clear "switch to a multimodal model" error rather than silently dropping it - one of the more thoughtful touches in this pack.

Gotchas

• One IMAGE per port, tooltip says so explicitly. Batches will confuse it - split them.
• The @图片N mention must match the port number. @图片1 is port 🖼️图片1. Aliases don't change which number triggers what.
• If the model doesn't support audio input, tagging audio either gets skipped or errors depending on model - stick to image/video refs unless you know your model takes audio.
• Re-tagging every round is easy to forget, and the model will cheerfully answer without re-looking at the reference. That's the feature, but it does mean your prompt discipline matters.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git into ComfyUI/custom_nodes/, pip install -r requirements.txt, restart. You need a dapaoAI API key (register at api.dapaoai.com, top up via redeem code) in the config node upstream. Like the rest of the pack the menus are Simplified Chinese - the numbers in the port names make it survivable.