Cloud Sampler Custom Advanced

If you've ever poked at SamplerCustomAdvanced in a local workflow, this is the same node, just running on ComfyUI Cloud instead of your GPU. The name is a mouthful, but the job is simple: instead of one KSampler box hiding steps, CFG and denoise behind a wall of defaults, you hand Cloud Sampler Custom Advanced four explicit pieces - noise, guider, sampler, sigmas - and it runs exactly that combination. It's the node you reach for when the simple path isn't good enough.

Why you'd reach for it

The pack ships Cloud KSampler and Cloud KSampler Advanced for the 95% of cases where you just want "sample this latent, please." This node is the other 5%: flow-matching setups (Wan, LTX) where you want a flat CFG 1.0 guider and a custom sigma schedule, or any two-stage sample where the default sampler is doing something you don't want. If you already know why you'd use SamplerCustomAdvanced locally, you know why you'd use this. If you don't, start with Cloud KSampler and come back when you have a specific reason.

How it works

Here's the pack's core trick, worth internalizing once: none of the "Cloud" nodes run a model locally. Each one appends a node spec to a growing workflow JSON, and Cloud Sampler Custom Advanced is a straight translation of ComfyUI's SamplerCustomAdvanced into that JSON. It merges the partial workflows carried by its five input handles, adds one SamplerCustomAdvanced node to the pile, and returns two cloud latent handles. The actual sampling happens on the cloud instance when a terminal node finally submits the whole chain.

The inputs and outputs that matter

All five inputs are handles from other cloud nodes, so there's nothing to type:

• noise (CLOUD_NOISE) - from Cloud Random Noise.
• guider (CLOUD_GUIDER) - from Cloud CFG Guider.
• sampler (CLOUD_SAMPLER) - from Cloud KSampler Select.
• sigmas (CLOUD_SIGMAS) - from Cloud Manual Sigmas.
• latent_image (CLOUD_LATENT) - from Cloud Empty Latent, Cloud VAE Encode, or Cloud Wan Image To Video.

It returns two latents: output (the diffused result) and denoised_output (the model's final denoised prediction). The pack's own docstring notes that LTX 2.0 uses the first, and advanced workflows sometimes use both. Either way they're CLOUD_LATENT, so they feed the next step in the chain - ultimately Cloud VAE Decode, then a terminal like Cloud Fetch Images to pull pixels back into your graph.

Installing it

Same story as every node in this pack - tiny dependencies (requests, Pillow, safetensors), no model downloads (everything model-side lives on the cloud), just an API key. 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

Then paste your API key from https://platform.comfy.org/profile/api-keys into config.json and restart ComfyUI. Model dropdowns populate in the background on startup.

Common issues & troubleshooting

• Nothing runs - you've probably forgotten the terminal. This node is lazy; the whole chain only fires when a terminal node (Cloud Fetch Images or similar) submits it. If you queue and see no network activity, that's the miss.
• "Stuck" jobs - watch the console for the pack's status prints (queued_waiting → allocated → preparing → executing → success). Long executing is normal on a video model, not a hang.
• Proof of concept - the README is honest that ControlNet, CLIP Vision conditioning and cloud-side video encoding aren't wired up yet. Don't assume every local sampler feature has a cloud twin here.