ComfyUI Node

Cloud VAE Decode

Where the cloud chain finishes decoding — and your GPU genuinely does nothing

By Dobidop·Created 5 months ago·Updated 5 months ago· 0
Cloud VAE Decode
  • samples
  • vae
  • images

Every local workflow has that satisfying moment where the sampler hands off to VAEDecode and pixels come out. Cloud VAE Decode is that node, with one twist that trips people up on first use: it doesn't hand you pixels. It appends a VAEDecode node to the assembled cloud workflow and returns a CLOUD_IMAGE handle. The decoding happens on the cloud instance; you still have to finish the chain with a terminal node that submits the job and pulls the images back down.

How it works

Two inputs, both cloud handles: samples (CLOUD_LATENT - whatever your sampler produced) and vae (CLOUD_VAE - from Cloud VAE Loader, or a checkpoint's bundled VAE). It merges their accumulated workflow JSON, appends a VAEDecode node, and returns a CLOUD_IMAGE. Nothing is submitted yet.

The terminal is Cloud Fetch Images. That node appends a SaveImage, submits the whole assembled workflow to /api/prompt, polls until it's done, downloads every output PNG, and returns them as a single local IMAGE batch. (The README's "How it works" prose says Cloud VAE Decode is the terminal - that's stale; the node list at the bottom of the README and the code both make Cloud Fetch Images the submitter.) So the practical chain is: sampler → Cloud VAE Decode → Cloud Fetch Images.

Why you'd reach for it

Because everything upstream in this pack is just JSON assembly until this point, Cloud VAE Decode is effectively where your "workflow" becomes a real cloud job. It's also where you should think about timeouts: the terminal node owns the whole job's timeout, and long video runs will punish a small one - the README quotes 8–10 minutes for Wan 2.2 14B I2V without the 4-step LoRA.

Video caveat

For video workflows the IMAGE batch you get back is N frames, not a single image. The README's advice: wire it into VHS_VideoCombine to mux locally. Don't expect a video container to appear on its own.

Installing it

Install via ComfyUI Manager (search "ComfyUI-CloudAPI-worker") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker
cd ComfyUI-CloudAPI-worker
cp config.json.example config.json

Paste your API key from https://platform.comfy.org/profile/api-keys into config.json and restart ComfyUI. Just requests, Pillow and safetensors locally - no model downloads, everything decodes on the cloud.

Common issues & troubleshooting

  • Forgotten terminal - Cloud VAE Decode on its own does nothing. If you queue and see no network activity, you've missed the Cloud Fetch Images (or Cloud Save Video) step downstream.
  • Job dies at the timeout wire - raise timeout on the terminal node, not this one.
  • Long executing looks like a hang - watch the console status prints (queued_waiting → allocated → preparing → executing → success). On a 14B video model, minutes of executing is normal.
Categorycloud

Inputs (2)

NameTypeDefaultDescription
samplesCLOUD_LATENT
vaeCLOUD_VAE

Outputs (1)

NameTypeDescription
imagesCLOUD_IMAGE