Float To Sigmas

Float To Sigmas does one small, specific job: it takes a list of float values and packages them into a SIGMAS tensor - the noise-level schedule a sampler steps through during denoising. If you've ever wanted to hand-write your own sigma schedule instead of picking one off a scheduler dropdown, this is the node that lets you.

Why you'd want a hand-built schedule at all

Sigmas are the sequence of noise levels a sampler walks down from (high noise, mostly random) to (low noise, nearly the final image) over the course of your steps. Normally you don't touch this directly - a scheduler node (normal, karras, simple, and so on) generates that curve for you based on your step count and a couple of parameters. But schedulers are curve shapes, and sometimes what you want doesn't fit any of the built-in shapes: a custom front-loaded or back-loaded schedule, a schedule that matches something a paper specified exactly, or a manually tuned curve you arrived at through trial and error and want to lock in and reuse exactly. Rather than fighting a scheduler's parameters to approximate that curve, you can just type the sigma values you want directly and feed them straight to a custom sampler.

This isn't an exotic idea - it's the same underlying concept behind why sigma-curve shape matters so much on flow-matching models in the first place, where the community has found that different curve shapes suit different image types and some samplers are far more sensitive to the schedule than others. This node just gives you the most direct possible control over that curve: you write the numbers yourself.

The inputs and outputs that matter

• float_list (FLOAT) - the values that become your sigma schedule, in order, highest noise first and descending toward zero, matching how every other sigma schedule in ComfyUI is structured.
• SIGMAS out - wire this into any node that accepts a sigmas input, most commonly a custom/advanced sampler node in place of whatever a scheduler node would normally provide.

There's genuinely nothing else to configure - this node's entire value is that it doesn't impose any shape on your input, it just converts it to the right tensor type.

Installing it

Ships with the pack, nothing extra:

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt, restart.

No dependencies beyond the base pack requirements - it's a type conversion, not a feature that pulls anything in.

Common issues & troubleshooting

Sampler errors or produces garbage. The most common cause with any hand-built sigma schedule is getting the order or the final value wrong - sigmas should descend from a high starting noise level down toward (usually) zero at the last step, not ascend or sit flat. If your list doesn't follow that shape, the sampler is working from a schedule that doesn't correspond to a real denoising trajectory, and results will reflect that.

This is more advanced than I need. It probably is, honestly - for the vast majority of workflows, a normal scheduler node picking a karras/simple/normal curve is the right call, and reaching for a hand-typed schedule is solving a problem you likely don't have yet. This node earns its keep specifically when you're chasing a schedule shape the built-in schedulers can't produce, or reproducing an exact published curve - not as a everyday replacement for the scheduler dropdown.

Not sure what values to put in. There's no generic answer here; it depends entirely on what curve you're trying to build. If you're just experimenting, start by outputting a known scheduler's sigmas (some custom-sampling setups let you preview them) and tweak individual values from there rather than typing a schedule from scratch with no reference point.