Run selected ComfyUI nodes on rented cloud GPUs as a visible, editable Cloud Offload box. Draw a box around any part of your graph and it executes on a remote runner while the original nodes stay expanded and report live progress, previews, and errors in place.

This pack is a thin client. All provisioning, queueing, provider credentials, and remote execution live in the separately built Cloud Offload coordinator service (Python package cloud_offload). RunPod is the default provider; Vast.ai is the alternative, and support for studio fleets and pooled compute is designed and on the roadmap.

Requirements

A running Cloud Offload coordinator service. The pack never imports it as a library and never handles provider credentials — it only speaks HTTP to the coordinator's client-facing routes.

Installation

Start the Cloud Offload coordinator separately (see the cloud-offload repo), for example:
```
python -m cloud_offload serve --host 127.0.0.1
```

Install the pack — from the Comfy Registry:

comfy node install cloud-offload

or clone into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/jethac/ComfyUI-Cloud-Offload.git

Restart ComfyUI.

Coordinator discovery

The pack discovers the coordinator in this order:

CLOUD_OFFLOAD_URL environment variable;
~/.cloud-offload/service.json (a JSON file with a url and optional token_path);
the localhost default http://127.0.0.1:11435.

Port 11434 (Ollama's reserved port) is never used. For a non-local or authenticated coordinator, set CLOUD_OFFLOAD_TOKEN, or let the pack read the token path recorded in the service file. The request itself is the authoritative availability check; discovery does not health-gate every call.

Nodes

| Node | Category | Description | |------|----------|-------------| | Cloud Status | Cloud Offload | Show queue counts, active workers, and RunPod/Vast.ai balances as JSON | | Cloud Workflow | Cloud Offload | Preflight and confirm a whole API-format ComfyUI workflow, then return its first image and artifact result JSON |

The four partition bridge nodes are compiler-generated and hidden (is_dev_only); users never place them by hand. They live under Cloud Offload/Internal:

| Node | Role | |------|------| | CloudPartitionGateway | Local proxy that submits the boxed subgraph and pauses until it completes | | CloudPartitionExtract | Restores one ordinary Comfy value from the partition result | | CloudPartitionInput | Runner-side bridge that restores an uploaded boundary value | | CloudPartitionOutput | Runner-side bridge that writes a typed boundary bundle |

Cloud Offload box

Select one or more nodes and choose Cloud Offload selection from ComfyUI's selection toolbox. A visible box appears around them and owns the provider, GPU type, minimum VRAM, timeout, and warm-runner policy. The nodes stay expanded and receive incremental remote progress: the currently executing node is highlighted, completed/cached nodes are marked, the box title shows percent, and live previews appear.

At queue time the box is compiled into a hidden gateway plus typed input/output bridges (see PARTITION_PROTOCOL.md). The runner image must contain every custom node and model the boxed subgraph uses. The default runner is model-agnostic: a pinned ComfyUI plus the partition bridge nodes, so any node installed in that image can ride inside the box and report normal node-level progress.

GPU recommendation and rental confirmation

After the gateway uploads the final boundary artifacts, it runs free preflight before it submits paid work. The default confirmation shows the recommended provider, GPU, region, hourly price, estimated total-cost and time ranges, prepared-cache coverage, rationale, confidence, and meaningful uncertainty. It starts automatically after the server-controlled ten-second countdown.

The panel also provides Start now, Cancel, Choose another GPU, and Don't show this confirmation again. Opening details or changing the GPU pauses automatic start. A material price, cost, capacity, region, or storage change always opens a new mandatory confirmation. The coordinator enforces the countdown, so no paid job can start early from the browser.

Use the Cloud Offload action-bar button to restore confirmation or change the countdown, recommendation policy, hard hourly, total-cost, and paid-runtime limits, allowed regions, or material-change tolerances. Hiding normal confirmation does not disable these hard limits or mandatory change notices.

The Cloud Workflow node uses the same free preflight and rental confirmation flow. It sends a canonical comfy.workflow.capsule.v1 closure. Input images are uploaded as content-addressed artifacts before preflight. Completed images, meshes, and files return as artifact records, not base64 job data. Cancellation continues into the remote ComfyUI execution.

Cloud Jobs panel

Use the Cloud Jobs action-bar button or the fixed Cloud Jobs button to open the persistent job panel. It loads independently of the canvas and rebuilds from the coordinator journal after a page reload or reconnect. Active jobs refresh once per second. Idle history refreshes every five seconds.

Each job shows its current stage and operation, elapsed time, progress basis, measured transfer bytes and throughput, ETA range and confidence, GPU and Pod identity, rate and estimated spend, prepared-cache work, cancellation state, billing state, durable resource lease, provider closure time, and recent safe event history. A terminal job never claims that billing stopped until the coordinator has a provider termination receipt for the exact resource.

The same-origin ComfyUI proxy returns only the coordinator's allow-listed job projection. Workflows, prompts, local paths, signed URLs, raw events, and raw coordinator job rows do not enter the browser panel.

Example

[Load Image] → [ ☁ Cloud Offload box: [KSampler] → [VAE Decode] ] → [Save Image]

Only portable boundary values may cross the box edge (IMAGE, MASK, LATENT, CONDITIONING, AUDIO, tensors, JSON-compatible values, byte buffers, and file-backed mesh / 3D-file artifacts). Live objects (MODEL, CLIP, VAE, control nets, samplers, …) are rejected before a paid runner is provisioned; move their loader or producer inside the box.

License

Apache-2.0