Limits of the assistant
The assistant is a language model with a few tools. It is good at explaining and at building common graphs from your nodes. It is not a person who has seen your pictures, and it can be wrong. This page lists what it cannot do today.
What it cannot do
- Run the graph or see the result. It never runs anything in ComfyUI and never sees the picture or video your graph makes. Tell it in words what came out.
- Install or download. It cannot add a node pack or a model; it can only tell you which one you need.
- Work inside a subgraph. It builds on the main canvas; the inside of a subgraph is yours to edit.
- Use what you do not have. It may only add node classes your ComfyUI has; others are refused.
- Answer two questions at once. One answer at a time; a question asked meanwhile waits.
- Remember other projects. The conversation belongs to this project; a cleared or switched-off history is gone. There is no button to start a new conversation; Clear chat history in Settings ▸ Privacy removes the old one.
- Change your work by itself. In a build it changes only what it placed; a change to your own nodes comes as a grey draft. See Grey drafts.
- Open a link. A link in an answer only copies its address.
Where it can be wrong
- Its words are not facts. What the assistant wrote carries the badge Assistant: "It may be wrong: the built-in knowledge and the documentation come first." Facts read from ComfyUI or a file carry Fact. When it matters, check.
- Small models. A small model on your own server may misread a long graph, pick wrong nodes, or stop making sense in a long build. A larger model or a cloud model helps.
- A small context window makes it forget the question. See The context window.
- The web. Pages may be out of date. Answers from the web carry their link.
What helps
Ask about one object (right-click it), say what you want in plain words, check the graph after a build, and use the Tour to read what was built.