Generation Parameters Hub

Every serious workflow ends up with the same value typed in five places. Steps, CFG, sampler, scheduler, denoise - you set them on the KSampler, then again on the refiner pass, then again on the hires-fix stage, and the first time you forget to update one of them, you get a render that disagrees with everything else in the graph. Generation Parameters Hub is the "define once, fan out" answer: one node holds the full sampling recipe and hands it to every consumer that needs it.

This is the plumbing pattern the KB calls "one source, many consumers" - the exact thing value nodes exist for, except here the value is a whole parameter set instead of a single number. For a model-testing workflow, where you sweep the same settings across a dozen models and want every render comparable, a single hub is what keeps the comparisons honest.

How it works

The node does no generation and loads no model. It's a router: you set the values, it returns them from seven outputs plus three string variants, and you wire those into your samplers, schedulers, and checkpoint loaders. Two details make it more than a toy:

• The sampler_name and scheduler dropdowns come straight from comfy.samplers.SAMPLER_NAMES and SCHEDULER_NAMES - the exact same lists the core KSampler uses. No hand-maintained list that drifts out of sync with your ComfyUI version; what the hub offers is what your install supports.
• The ckpt_name output is typed as a combo slot, and the sampler/scheduler outputs are combo-typed too, so they plug directly into the matching sockets on a Checkpoint Loader, KSampler, or SamplerSelect rather than just being text that happens to look right.

The inputs you actually set

• steps (20) - the base denoising step count.
• step_refiner (10) - the refiner pass's step count, or the threshold for multi-pass pipelines. If you don't run a refiner, leave it unwired.
• cfg (7.0) - classifier-free guidance; how strictly the model follows the prompt.
• sampler_name / scheduler - algorithm and noise curve.
• denoise (1.0) - 1.0 for fresh txt2img, 0.2–0.6 for img2img or upscaling passes.
• ckpt_name - which checkpoint the recipe applies to, passed along for downstream loaders.

The three *_str outputs (ckpt_name_str, sampler_str, scheduler_str) are the same values as plain text, for metadata and save filenames - the "what did I run this with" record.

Installing it

Part of MarwanDSAI/comfyui-mdsnodes. ComfyUI Manager: search "ComfyUI-MDSNodes", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/MarwanDSAI/comfyui-mdsnodes

Restart, and no pip install - the pack's requirements.txt is empty.

Gotchas

Remember what this node isn't: it doesn't load a checkpoint, so you still need a Load Checkpoint somewhere, and the ckpt_name output is there to drive it, not replace it. And a hub is only as trustworthy as the wiring - if one sampler in a multi-pass setup gets a steps value from somewhere else, you're back to the silent-mismatch problem this node was built to kill. Wire every consumer from the hub and nothing else, and the recipe stays single-sourced.