Sigmas Start

Sigmas Start is a schedule-splitter. You hand it a full sigma schedule and a step number, and it gives you back just the front portion - the early, high-noise steps. It's most at home in "chainsampler" setups, where you run part of a generation on one sampler and hand the rest off to another, which is one of the workflow tricks the RES4LYF docs specifically show off.

Some grounding first. The sigma schedule is the ladder of noise levels your sampler steps down, built by a scheduler node, high at the start and near zero at the end. Normally one sampler walks the entire ladder. But sometimes you want the first chunk of steps done by, say, a fast multistep sampler that converges cleanly toward the composition, then the tail done by a slower substep sampler that nails detail. To do that you have to cut the schedule in two - and Sigmas Start gives you the front half.

How it works

Feed it a SIGMAS schedule and set sigmas_until. It returns the sigmas from the start of the schedule up to that point, dropping the rest. Pair it with a step or seed offset on the second sampler so the tail picks up exactly where this slice left off, and the two samplers form one continuous denoise. The output is still a normal SIGMAS list, so anything that eats a schedule will accept it.

The inputs and outputs that matter

• sigmas (SIGMAS) - the full schedule you're slicing.
• sigmas_until (INT, default 10) - where to cut. As the name and this cutoff suggest, you keep the leading steps up to this index. On a 20-step schedule, 10 gives you the first half to feed the first sampler; the second sampler then runs the remainder.

Output is one SIGMAS, which you wire into the first sampler's sigmas input. Feeding an external schedule overrides that sampler's internal scheduler and step count, which is exactly what you want here - you're driving it deliberately.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF/, activate your venv and pip install -r requirements.txt (portable ComfyUI: use the embedded python's pip), then restart and hard-refresh with F5.

The pack's nested sampler menu needs rgthree-comfy installed with "Auto Nest Subdirectories" enabled - worth having anyway.

Common issues

Chainsampling is genuinely fiddly, and this node is only one piece of it. The single most common mistake is a mismatch between where you cut the schedule and where the second sampler resumes. If the handoff step counts don't line up, you get a visible seam mid-denoise - a sudden shift in the image - because the second sampler is starting from a noise level it doesn't expect. Keep the cut point and the resume point consistent.

Also worth saying plainly: for most people, most of the time, a single sampler over the whole schedule is the right call, and chainsampling is an optimization you reach for when you have a specific reason (a fast explore pass plus a slow finishing pass, or blending two sampler characters). If you're just trying to make one image look better, start there before splitting schedules. And because the whole point is precision, hold your seed fixed while you tune the split - otherwise you can't tell a bad cut from a different roll of the dice.