ComfyUI-RemoteModels

<p align="center"><img src="docs/hero.png" width="880" alt="Remote Models: run the text encoder and VAE on another machine's ComfyUI"></p>

ComfyUI Remote Models

Why: save VRAM. Text encoders take gigabytes and VAE decodes spike; put them on a spare GPU and keep yours for the diffusion model.

How: add Load Remote CLIP or Load Remote VAE, point it at the other ComfyUI, and use it like the normal loader.

<p align="center"><img src="docs/nodes.png" width="880" alt="Load Remote CLIP and Load Remote VAE wired into a workflow"></p> <p align="center"><img src="docs/settings.png" width="880" alt="Remote Models settings"></p>

Install

Clone into ComfyUI/custom_nodes on both machines and restart ComfyUI:

git clone https://github.com/Mr-Quin/ComfyUI-RemoteModels

Server

• Start ComfyUI with --listen and allow its port (8188) through the firewall.
• Put the model files in its models/text_encoders and models/vae. The client doesn't need them.
• In Settings > Remote Models, turn on Serve models to other machines.
• For VAE encodes of long videos, raise ComfyUI's request limit with --max-upload-size (in MB, default 100).

| Setting | Default | | | --- | --- | --- | | Serve models to other machines | off | Off answers every request with 403. | | Unload idle models after (minutes) | 10 | 0 keeps them loaded. | | Reload recently used models on startup | on | Avoids a slow first request after a restart. |

There is no authentication, same as ComfyUI itself, and the serve switch is not access control: anyone who can reach the port can change settings. Limit access to the port at the network level.

Client

1. Optionally set Default server URL in Settings > Remote Models.
2. Add Load Remote CLIP or Load Remote VAE (loaders/remote) and enter the server URL, e.g. http://192.168.1.50:8188. It connects on its own and fills the file list.
3. Wire CLIP into your encode nodes and VAE into VAE Decode or Encode, tiled or not.

On the server machine itself, set the URL to local to share the already-loaded models with local workflows.

Reference workflows are in example_workflows/. Dragging docs/icon.png into ComfyUI loads the workflow it was made with (the background removal needs BiRefNet).

| Button shows | Fix | | --- | --- | | Unreachable | Check the URL, --listen and the firewall on the server. | | Server not serving | Turn on serving in the server's settings. | | Pack missing or old on server | Install or update the pack on the server and restart it. | | Version mismatch, update both | Update the pack on both machines. | | Not on server: <file> | Add that file to the server, or pick another. |

Limitations

• Load Remote CLIP works with encode nodes that just tokenize and encode a prompt (CLIP Text Encode, Qwen image, MOSS-Image, Krea 2, Anima). SDXL, SD3 and Flux dual-prompt nodes, CLIP LoRAs and hooks are not supported.
• Load Remote VAE covers decode and encode, tiled or not. Nodes that reach into the VAE weights directly do not work with it.
• A remote VAE is only as fast as the server's GPU, and decoded images cross the network, so it trades speed for VRAM. Output can differ slightly from a local decode when the server has a different GPU.
• Remote requests wait while the server runs its own prompts, but a prompt queued on the server while a request is already running is not held back, and the two can disturb each other's models. Avoid generating on the server while clients use it.
• Embeddings in prompts come from the server's embeddings folder.
• After replacing a model file on the server, restart its ComfyUI.
• Settings are read from the default user, so servers started with --multi-user need serving turned on as that user.

License

MIT