Lam 智能体

LamAgent (the UI calls it 智能体, "agent") is the brain of the Lam pack. You give it a topic and it returns a structured, ready-to-use generation plan - the kind of output that feeds straight into a video model. It does this by calling an LLM over an API, not by running anything locally, and it's clearly designed to be the "director" in front of the MiniMax H3 video node this pack also ships.

How it works

LamAgent is a thin wrapper around a cloud LLM. Look at the defaults and you'll see the real story: server_url points at https://ark.cn-beijing.volces.com/api/coding/v3, which is Volcengine ARK - ByteDance's cloud AI platform (the same infra behind Doubao). The default model_name is deepseek-v4-pro, and the output is a single STRING named 结果 ("result"). So the node is: send my text to a model on Volcengine ARK, get a response back, hand it to the rest of the graph.

The interesting part is skill. There are nine preset skills, and they're all Chinese-language video-production roles:

• 3D动画短片生成器 - 3D animation short-film generator
• h3-prompt-writing - H3 prompt writer (the one you want before the H3 node)
• 双人游戏开场视频生成器 - two-player game opening video generator
• 品牌宣传短片生成器 - brand promo short generator
• 手绘实拍融合视频生成器 - hand-drawn / live-action fusion generator
• 极简产品广告生成器 - minimalist product ad generator
• 纸拼贴讲解动画生成器 - paper-collage explainer generator
• 纸艺定格科普视频生成器 - paper-art stop-motion science explainer
• 音乐MV动态字幕生成器 - music MV dynamic-subtitle generator

Each skill is essentially a system prompt that makes the model produce output in the right shape for that production style. You pick a role, type a topic into text, choose language (默认 "自动(中文优先)" - auto, Chinese-first - or force 中文/英文), and the model returns a generation plan in 结果.

The inputs that actually matter

For a beginner, only three things matter:

• api_key - you must get one from the Volcengine ARK console and paste it in. Without it, nothing happens. This node is a paid API call; there's no free local inference.
• skill - pick the production style. If your target is MiniMax H3, h3-prompt-writing is the one that produces a prompt the H3 node will actually like.
• text - your topic or instruction, in plain language.

server_url and model_name you'll almost certainly leave at their defaults, though you can point model_name at any model your ARK key can access.

Wiring it into a workflow

The single 结果 STRING output is the whole payload. In a Lam workflow it usually flows one of two ways: into a text display so you can eyeball the plan, or straight into the prompt input of LamMiniMaxH3RefVideo. The pattern is: IdentifyingQR decodes a QR with your brief → LamAgent expands it into a full H3 prompt → LamMiniMaxH3RefVideo conditions on it → sample, decode, save. The agent is the piece that turns a sentence into something a video model can act on.

Installing and the real gotchas

Install is just the pack install - ComfyUI Manager search ComfyUI_Lam, or:

cd ComfyUI/custom_nodes
git clone https://github.com/yanlang0123/ComfyUI_Lam.git

then restart. The pack README covers install scripts and model files but says nothing about LamAgent; the API details live entirely in the node defaults.

Where people get burned:

• No API key, or a key without the model enabled. Volcengine ARK keys only work for models you've actually enabled on your account. If you get an auth or "model not found" error, that's where to look.
• It's a China-region cloud endpoint. If you're not in a region with good connectivity to cn-beijing, calls can time out. Point server_url at your own ARK endpoint if you have one.
• Language expectations. The default is Chinese-first output. If you want English results for downstream tools that expect English, set language to 英文.

One honest caveat: this is a fully cloud-dependent node for a pack that's otherwise about local generation. If you'd rather not hand your prompts to a Beijing API, you can get the same "LLM writes my prompt" effect from local nodes like DiffusionGemma or any local-LLM wrapper. LamAgent's edge is that it's zero-setup if you're already paying for ARK, and the skills are tuned for this pack's video workflow.