Cloud CFG Guider

If you've ever built a SamplerCustomAdvanced chain locally, you know the drill: a CFG Guider, a sampler select, a sigma schedule, a noise source, and then the custom sampler itself, wired together like a miniature engine. Cloud CFG Guider is the version of that first piece that runs on Comfy Cloud instead of your GPU. It takes your model, your positive and negative conditioning, and a cfg number, and hands back a CLOUD_GUIDER handle - a promise that a CFGGuider node will exist in the workflow JSON the cloud actually executes.

Like every node in this pack, it doesn't compute anything. The name is honest about the inputs and silent about the mechanism: nothing runs locally, no weights load, no tensors move. What you get out is a lightweight reference that carries a fragment of workflow JSON around, waiting for a terminal node to stitch everything together and ship it off to cloud.comfy.org.

Why you'd reach for it

Honestly? Most people shouldn't. If you just want text-to-image or image-to-image on a rented GPU, Cloud KSampler (the "Quick" one) is one node and it's done. Cloud KSampler Graph wraps a whole stock KSampler and is the better default for ordinary graph-style chains. This node exists for the niche but real case where you're porting an advanced local workflow that splits sampling into its primitives - guider, sampler, sigmas, noise - so you can do things like two-stage sampling or custom sigma schedules. If your local graph has a CFG Guider in it, this is its cloud mirror.

The inputs and output that matter

• model (CLOUD_MODEL) - from a Cloud Checkpoint Loader or Cloud UNET Loader.
• positive / negative (CLOUD_CONDITIONING) - from two Cloud CLIP Text Encode nodes.
• cfg (FLOAT, default 8.0, 0–100) - the guidance strength, same knob as the stock node.
• Output: guider (CLOUD_GUIDER) - wire it into Cloud Sampler Custom Advanced, the pack's sibling node that consumes it.

One trap: all three inputs must be cloud handles, not local tensors. If you try to feed a locally-encoded conditioning into a cloud guider, the socket types won't match and the graph won't even type-check. Keep the whole chain cloud-side until you hit a terminal node.

How to install

Install once for the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Then copy config.json.example to config.json inside the pack folder and paste an API key from https://platform.comfy.org/profile/api-keys, and restart ComfyUI. Easier: search "ComfyUI-CloudAPI-worker" in ComfyUI Manager and let it handle the clone. Dependencies are just requests, Pillow, and safetensors - no model downloads, because the models never come near your machine.

Common issues

• Wrong handle types. Cloud CFG Guider only accepts cloud handles. Double-check the sockets you're wiring from.
• Nothing happens. A graph of cloud nodes produces no output until you attach a terminal like Cloud Fetch Images or Cloud VAE Decode. This node is mid-chain by design - it's not a submission point.
• It's a proof of concept. The author ships it with the caveat that the pack is "very hacked together" - fine for image and video diffusion, rougher for exotic setups. Expect rough edges and check the GitHub issues before you bet a production run on it.

And the pack-wide reminder: this is an API-wrapper node that holds your key and phones home by design. Clone from the official repo, and remember your prompt text is part of what leaves the machine.