NodeWise Studio
RUDownload for Linux

Known limitations

What the app does not do, and where it is weak today. NodeWise Studio is an experimental project; this list says plainly what to expect. Wishes go to the project's Discussions, problems to its Issues.

By design

  • It never runs a workflow. It builds and explains; ComfyUI runs. Running the result in ComfyUI
  • It never changes ComfyUI. It does not install nodes or models, download files or queue prompts. ComfyUI is only read
  • Two languages: English and Russian.

Systems and packages

  • Linux only, 64-bit x86: .deb, .rpm, AppImage and a .tar.gz. There is no ARM package and no Flatpak.
  • Windows and macOS are not supported. You may try running the app from source; nothing there has been tried. Windows, macOS
  • No automatic updates: install a new version from the project's Releases yourself. Settings ▸ Privacy ▸ Check for new versions can tell you when there is one. Updating
  • Not tried yet on a real desktop: installing the .deb with apt, the .rpm with dnf on Fedora, the final AppImage file. These are checked on CI only.

Connection and reading

  • Not tried yet on a real install: the connection was checked against test servers that answer the way ComfyUI does, not against a real ComfyUI, ComfyUI on Windows, the ComfyUI desktop app or a real MCP server.
  • Straight from ComfyUI there are no descriptions next to model files, no pack READMEs and no fingerprints, so Find on the internet searches by file name only. An MCP server that reads files adds them. What each way shows
  • An older ComfyUI lists model names only, without sizes or pictures.
  • The official comfy-mcp of Comfy Org runs as a local command and cannot be connected yet; an MCP server whose tools the app does not know is not used.
  • The context window of a model on your own server is read only from Ollama. Other servers do not report it; the app then notices a request that did not fit only by the token count of the answer.
  • Model lookup: Civitai finds a file exactly by its fingerprint; Hugging Face is searched by the file name, and the card says it is a guess. Computing the fingerprint of a large file takes the server minutes, so sync does not do it.
  • A node pack is described from its pyproject.toml and the start of its README; a pack without them shows only its name and version.
  • When the cloud model has no web search of its own, the assistant searches through DuckDuckGo's HTML page, which can stop working when that page changes.

Size of a graph

  • About a thousand nodes is slow. On a graph of about 1000 nodes, dragging and zooming run at about 5 frames per second, and the first drawing takes many seconds.
  • The cap is 1500 nodes and 3000 links, subgraphs counted. A larger workflow or project is refused before anything changes: "This graph is too large to open…". The largest of 1335 real workflows the app was tried on has 603 nodes.
  • If the canvas did not finish drawing a graph at the last start, the app asks once whether to start with an empty canvas (the graph stays in the history) or to open the graph again.

The graph

  • Inside a subgraph you cannot make a new subgraph or a group, and a nested subgraph cannot be dissolved: do that on the main canvas. Groups that came inside a subgraph from ComfyUI are kept on save but not shown.
  • A workflow saved in the older format (0.4) cannot hold a subgraph; the app says so. Open it in a current ComfyUI and save it again, or copy its nodes into a new project (which uses format 1). Subgraphs
  • Legacy group nodes of older workflows are shown as one unknown node; they are saved back as they were.
  • Inputs that grow as you connect them, and list marks on inputs, are not shown yet.
  • Native reroute points of ComfyUI are drawn and saved back, but cannot be moved. Add reroute in the menu of a link adds a Reroute node instead.
  • A group frame follows its nodes. You cannot resize it by hand: add nodes to it or take them out (Add to group, Take out of group).
  • A picture or video opens only when ComfyUI wrote the workflow into it. A picture from elsewhere, or one sent through a chat app, usually has none.

The assistant and lessons

  • The assistant builds only on the main canvas. If a subgraph is open, the canvas returns to the main level and says so. Lessons also run only on the main canvas.
  • A pause at a pause point (the step limit) does not survive a restart of the app: ask again after the restart.
  • The quality of answers depends on the model. Small local models with a small context window lose track of long builds. The context window
  • The video lessons on WanVideoWrapper cover a part of that pack; the rest of its features have no lesson.
  • A lesson the assistant drafted is gone after a restart unless you kept it in My lessons.

Accessibility

  • The arrow keys do not move the focus between nodes: they move the selected node, and Tab steps through the nodes.
  • Not tried yet with a real screen reader.
  • At 200 percent interface size on a 1080p screen the window does not fit.

More: Accessibility.

See also

This page is imported from the application repository's wiki (docs/wiki/reference/Known-Limits.md, revision b8e3f71). The same guide is built into the app: Help ▸ User guide.

To refresh it, run npm run import and then npm run build in the website folder.