Add a machine
- Start ComfyUI on the other machine with
--listenso it accepts connections from the network, for examplepython main.py --listen 0.0.0.0 --port 8188. - Open the settings menu and click LAN instances….
- Fill in a name (optional), the address and the port, then click Add.

Choose where jobs run
Generate on in the settings menu sets the default for every job:
Each generator card can override the default under Run on, next to the model. Default follows the settings menu; any other choice applies to that card only and is kept when you regenerate or remix its results.

How Auto picks a machine
For every job, Auto asks each enabled machine for its queue and checks that it has every node in the workflow and every model file the workflow names. Machines that are off, unreachable or missing something are skipped. Among the rest it picks the shortest queue; on a tie it picks this computer. If no machine can run the workflow, the job goes to this computer and fails there with ComfyUI’s own error. A batch of several images in one card (Count 2–4) is one job and runs on one machine. To spread work across machines, start several cards.Files and models
- Input images, video and audio are uploaded to the machine’s
input/comfyart/before the job is queued. - Results are downloaded into this computer’s
output/comfyart/, with the machine name and job id added to the file name, for exampleimage_00001_Spark-28138bd7.png. - Model files are matched by file name, so they may sit in different subfolders on each machine.
- Qwen Image 2512 loads with the ComfyArt Precise UNET Loader, so it only runs on machines that have ComfyArt installed. Auto skips the others; picking such a machine by hand gives a “node not found” error.