Cloud KSampler

Cloud KSampler (display name "Cloud KSampler", class CloudKSamplerGraph) is the graph-style sampler of the pack, and its whole design goal is that it should look nothing like "the cloud" and everything like the stock KSampler you already know. Same shape: model, positive and negative conditioning, and a latent in; a sampled latent out. Same dials: seed, steps, cfg, sampler, scheduler, denoise. If you can read a local KSampler workflow, you can read a cloud one.

The difference from the "Quick" Cloud KSampler is the difference between assembling a graph and calling a monolith. The Quick node bakes the entire checkpoint→prompt→latent→sample→decode chain into one node and hands you an image. This node is a mid-chain piece: it takes handles, it returns a CLOUD_LATENT, and it does not submit anything. It only appends a KSampler spec to the accumulating workflow JSON. The job actually ships when you hit a terminal node downstream - Cloud VAE Decode to go back to pixels, Cloud Fetch Images to bring them home, or Cloud Materialize Latent if you want the raw latent locally.

The inputs that matter

• model (CLOUD_MODEL) - from Cloud Checkpoint Loader or Cloud UNET Loader.
• positive / negative (CLOUD_CONDITIONING) - from two Cloud CLIP Text Encode nodes.
• latent_image (CLOUD_LATENT) - from Cloud Empty Latent (images) or Cloud Empty Latent Video (video chains).
• seed / steps / cfg / sampler_name / scheduler / denoise - the standard controls, with familiar defaults (steps 20, cfg 7). 22 samplers and 7 schedulers are on offer.

Output: latent (CLOUD_LATENT) - into Cloud VAE Decode, another sampler, or a materialize node.

When to use it

Whenever you want a graph, not a shortcut. Adding a LoRA between the checkpoint and the sampler? Need a custom scheduler? Two-stage sampling? The graph-style nodes are how you do that, and this is the sampler they all funnel through. Single-stage text-to-image with no LoRAs? The Quick node is one node and done - use that instead and save yourself the wiring.

How to install

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

Copy config.json.example to config.json, paste an API key from https://platform.comfy.org/profile/api-keys, restart. Or install via ComfyUI Manager. Dependencies: requests, Pillow, safetensors.

Common issues

• "I queued it and nothing happened" - expected; this node never submits. Attach a terminal (Cloud VAE Decode → Cloud Fetch Images) and the whole chain ships together.
• Local nodes won't connect - this sampler only speaks cloud handle types. Keep the chain cloud-side until the terminal.
• Want two-stage? Use Cloud KSampler Advanced instead - the plain KSampler has no start_at_step/end_at_step controls for a reason.
• PoC caveat - single-author experimental pack. The plain single-pass image workflow is one of the pack's solid paths; expect roughness the further you wander from the example workflows.