Cloud Offload Partition Output

Despite the display name - "Cloud Offload Partition Output" - this node is the return half of the cloud box's handshake, and it runs on your machine. CloudPartitionExtract is hidden and dev-only, compiled in by the pack rather than placed by you. Its one job: take the opaque CLOUD_PARTITION_RESULT that the gateway produces and pull one ordinary Comfy value back out of it, so your local graph picks up exactly where the box ended.

Where it sits

A boxed partition compiles into a hidden gateway plus one extract node per boundary output. Each extract is wired to the same partition result and produces the value that feeds whatever came after the box - as if the cloud run had happened locally. Your Save Image, your upscaler, your downstream sampler: they never know the pixels took a detour through a rented GPU.

Inputs and output

• result - the CLOUD_PARTITION_RESULT from the gateway.
• boundary_key - a string naming which output to pull out of the result bundle.
• type_name - the boundary type, which the node validates before doing anything.

The output is value of type * - a wildcard, because the extracted thing can be an IMAGE, a MASK, a LATENT, a string, a tensor, or a file-backed mesh restored to a fresh local file. That wildcard is exactly how one node can return different types depending on what crossed the boundary.

How it works

It validates the type, then looks up boundary_key in the result's values map. A missing key means a hard, specific error - "Cloud partition returned no output for X" - which tells you the box's output side never wrote that value, not that the whole cloud run failed. File artifacts (meshes and FILE_3D values) are written to a temp partitions directory under your system temp and handed back as a local path object, so 3D results survive the trip too.

Install and the fine print

Install is the standard pack path - ComfyUI Manager (search "Cloud Offload") or clone into custom_nodes, restart, and make sure the coordinator is running:

python -m cloud_offload serve --host 127.0.0.1

The pack's own deps are just requests and safetensors; there's nothing to download on top.

The is_dev_only flag means you won't find this node in the normal picker, and you should never hand-wire one. If you open a saved workflow and see CloudPartitionExtract nodes sitting between the gateway and your downstream nodes, that's the plumbing working as designed - the value they hand forward is just a normal Comfy value, so the rest of your graph treats it like any other wire.