ComfyUI and GPU setup

The guide is part of Settings

Open Settings → Installation guide for the complete Korean or English GPU playbook. The same guide is available from Settings → Providers → GPU setup → Read the installation guide and remains readable offline.

Opening a launcher does not immediately download a model or create a billable GPU. 9Miho opens a separate terminal and stops at each license, download, installation, and cloud-cost boundary. Credentials stay in official CLI or hidden terminal input; never paste them into Canvas or agent chat.

Choose the correct execution path

Goal	Supported systems	Choose in Settings
ComfyUI and selected models on this computer	Windows with NVIDIA GPU	Local ComfyUI · Windows NVIDIA
Private GCP VM connected through IAP	Windows, macOS, Linux	GCP ComfyUI through IAP in Settings
New remote Runpod ComfyUI	Windows, macOS, Linux	Runpod remote ComfyUI (Alpha)
Inspect, start, or stop an existing Runpod Pod	Windows, macOS, Linux	Runpod · existing Pod control

The launcher list is server-owned. A card is disabled when its required setup component or terminal is not available on the server machine. If a terminal start is left unclaimed, Settings can offer Prepare safe retry after the recovery window instead of opening duplicate setup processes.

Approval words are deliberate

• DOWNLOAD permits approved model files to be downloaded.
• INSTALL permits pinned software to be installed in an isolated folder.
• UPLOAD permits verified files to be copied to paid cloud storage.
• DEPLOY permits billable cloud resources to be created.
• START or CONNECT permits a specific local service, VM, tunnel, or Pod to run.
• STOP permits an existing Pod to stop.
• REPLACE permits a conflicting 9Miho profile to be changed.

Closing or declining before an approval prevents later steps from running. Setup approval is still separate from the per-Run spend card shown before a generation dispatch.

Local ComfyUI on Windows

Start 9Miho, open Settings → Providers → GPU setup, then choose Local ComfyUI · Windows NVIDIA. The reviewed installer is included in the signed 9Miho runtime installed by Kumiho Desktop; no repository access or separate installer download is required.

The installer checks GPU and VRAM, disk space, target paths, model terms, profile-name conflicts, and the installation plan before changing the machine. Choose H3, Ltx25, or None. Model download and runtime installation require separate approvals.

After installation, the installer creates or updates a non-default Local ComfyUI profile. If you approve START, it launches loopback-only ComfyUI and the 9Miho local inference service, verifies required nodes and model slots, then makes the profile default only if concurrent Settings changes have not selected something else. The owning terminal keeps the process alive; closing it stops ComfyUI while downloaded model files remain.

GCP ComfyUI through IAP

Choose GCP ComfyUI through IAP in Settings → Providers → GPU setup. 9Miho prepares the required pinned inputs and opens the reviewed flow; there is no supported standalone command for this path.

The launcher verifies a fixed public setup-kit revision and file hashes before opening the terminal. It checks pwsh, gcloud, the active account, project, quota, zone, and model access; provisions the specifically approved private VM; verifies model files; stops the VM after bootstrap; then asks before starting or connecting.

The currently bundled public IAP kit supports the ltx23-hdr setup. Use another supported execution path for other model families until a matching pinned kit is available.

The resulting ComfyUI profile records the non-secret project, zone, and immutable instance ID and is marked externally billable. Keep the terminal open while rendering because it owns the local service and IAP tunnel. A normal exit requests a VM stop; if shutdown cannot be proven, check the exact instance ID in Google Cloud rather than stopping a VM by name.

This is not the Vertex AI ADC path. IAP connects 9Miho to your own ComfyUI GPU VM; ADC authenticates hosted Google model APIs.

Runpod remote ComfyUI

For a new deployment choose Runpod remote ComfyUI (Alpha) in Settings → Providers → GPU setup. The installer included with the signed runtime opens a separate terminal for publisher review, authentication, and billable approval.

The installer shows the digest-pinned image and model plan, uses hidden input for the Runpod key, separates controller and Hugging Face secrets, asks before creating a billable Pod, verifies controller health and model slots, and only then attempts to make the profile default. The profile uses the authenticated 9Miho controller on port 9808, not a raw public ComfyUI :8188 URL.

For an existing Pod, choose Runpod · existing Pod control, then select Status, Start, or Stop and enter the exact Pod ID in the terminal opened by 9Miho.

The API key is requested through hidden input and is never a command-line argument. Stopping a Pod releases GPU compute but retained or network volumes can continue to cost money.

Recovery and readiness

• Keep the setup terminal open until it reports success or a specific recovery action.
• Do not open a second installer for the same profile while the first terminal is active.
• If a profile save response is uncertain, inspect Settings → Providers before deleting or recreating anything.
• Verify local service health and every required model slot before selecting a ComfyUI Variant.
• For GCP or Runpod, verify the exact VM or Pod state and retained-resource costs in the provider console after a failure.
• An UNPRICED external-cost card means the manifest cannot know your real GPU, disk, storage, network, discount, or Spot price. It does not mean free.