ComfyUI

ITS INCOM AI runs ComfyUI on its own GPUs. You build the workflow, our machines do the computing, and you pay for the seconds a machine works on your job. There are three ways in, with the same account, the same models and the same price:

The page https://ai.itsincom.org/comfyui has the examples, the models with their licences and the price per second.

The editor in the browser#

Sign in to the dashboard and open ComfyUI. Workflows are saved in your account and results stay in the cloud for 30 days. The models are already installed: the template browser's "missing models" notice does not apply.

ComfyUI Desktop with the add-on#

The add-on puts a button next to Run. It decides where Run runs the workflow: on this computer or on our GPUs. In cloud mode the whole graph is sent, progress and previews come back while the machine works, and the results are saved in this computer's output folder.

Install it once per computer.

macOS and Linux:

curl -fsSL https://ai.itsincom.org/comfyui/install.sh | sh

Windows (PowerShell):

irm https://ai.itsincom.org/comfyui/install.ps1 | iex

The installer looks for ComfyUI's folder, downloads the add-on and unpacks it into custom_nodes/comfyui-cloud-gpu. Where it does not find the folder, pass it: curl -fsSL https://ai.itsincom.org/comfyui/install.sh | sh -s -- /path/to/ComfyUI. By hand, download the zip and unpack it into custom_nodes. Then restart ComfyUI, press the button next to Run, paste an API key and choose ITS INCOM AI.

What to expect in cloud mode:

• The node search and the model lists show what our machines have. A workflow that uses a node we do not have is stopped before it starts, with the node's name, and nothing is charged.
• Files the workflow loads from the input folder are uploaded before the job, once per content.
• Cancel stops the job on the GPU.
• When the job is done a notice says the seconds, the machine and the cost.

The add-on is at version 0.1 and was tested with ComfyUI 0.44 on Linux. Tell us what you see on Windows and macOS at segreteria@itsincom.it.

ComfyUI's API#

The service speaks ComfyUI's API under https://my.ai.itsincom.org/comfy/api, with your key as a bearer token. In the editor, Workflow → Export (API) gives the graph in the form the API takes.

Queue a job:

curl https://my.ai.itsincom.org/comfy/api/prompt \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": { ... }}'

The answer carries a prompt_id. Read the job with it:

curl https://my.ai.itsincom.org/comfy/api/jobs/PROMPT_ID -H "Authorization: Bearer $KEY"

status goes from pending to in_progress and ends as completed, failed or cancelled. A completed job lists its files under outputs; download each one with its filename, subfolder and type:

curl "https://my.ai.itsincom.org/comfy/api/view?filename=FILE&subfolder=cloud&type=output" \
  -H "Authorization: Bearer $KEY" -o result.png

Each user sees only their own queue, files and history.

Choosing the machine. Without a choice the job goes to the machine with the shortest queue. To pick one, send its nome from /api/siati/macchine in the X-Siati-Macchina header of POST /api/prompt.

What the job cost. GET /api/jobs/{id} has a siati object with the machine, its price per second and the seconds it worked. The cost is their product, booked when the job completes.

Limits#

• 3 jobs at a time per user, queued or running.
• Input files up to 50 MB each.
• Results stay in the cloud for 30 days.
• ComfyUI's stock nodes only: no third-party node packs and no models of your own for now.

Content checks#

The texts of a workflow are checked before it is queued, and every result is checked before it is delivered. A workflow whose texts are refused is not run and costs nothing. A result that is withheld is not delivered: the job's siati object says how many files were withheld and why. The job did run, so its seconds are counted.

Errors#

Errors have ComfyUI's shape, {"error": {"type": "...", "message": "..."}}.