Cloud Model Sampling SD3

Since SD3, most new models aren't diffusion-in-the-classic-sense - they're flow-matching, and one of the few tuning knobs left on a flow model is shift. It controls how sampling effort splits between the compositional early steps and the detail late steps. Cloud Model Sampling SD3 is the cloud version of ComfyUI's ModelSamplingSD3 node: it takes your CLOUD_MODEL, applies a shift, and returns the patched model. The source docstring names the family it serves - SD3, Flux, Wan, and friends.

Where it fits: after your cloud checkpoint or UNET loader, before your sampler. The cloud model handle flows in, a ModelSamplingSD3 entry gets appended to the workflow JSON, and the patched handle flows out. If you're replicating a local workflow that used ModelSamplingSD3 - very common on Flux-family models - this is the node that keeps the port faithful.

Inputs

• model - a CLOUD_MODEL handle.
• shift - FLOAT, default 3, range 0–100. The higher the shift, the more sampling effort goes to low-noise detail at the end; lower pushes effort toward composition. ComfyUI's own default is 3, which is a fine starting point for most models.

Output: model, a patched CLOUD_MODEL.

Installation

Same pack:

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

Copy config.json.example to config.json, add an API key from https://platform.comfy.org/profile/api-keys, restart. In ComfyUI Manager under "ComfyUI-CloudAPI-worker"; deps: requests, Pillow, safetensors.

Gotchas

• Shift is model-specific. The KB's flow-matching panel is blunt about this: some models want a very different shift than the default (Z-Image Turbo, for instance, is known to want 7 against ComfyUI's 3). Copy the value from the model card or a known-good local workflow rather than improvising.
• It only matters on flow-matching models. Apply it to a DDPM-style checkpoint and it's noise at best, harmful at worst. Check what your checkpoint is built on before adding this node.
• This is a cloud-side JSON-builder like the rest of the pack; the actual sampling still happens when the chain reaches the terminal node. And the whole pack is a proof of concept, so if the cloud's ModelSamplingSD3 support ever drifts, the failure shows up as a cloud API error, not a friendly local one.