KoO Sigma Schedule Designer

Pick "normal" or "karras" from the scheduler dropdown and you're picking one of a handful of curves somebody else wrote. Those curves decide where your steps go - how much of the denoising budget gets spent inventing composition versus defining materials versus polishing texture. Same 30 steps, same seed, different spacing, visibly different image.

KoOSigmaEqualizer - it shows up in the menu as KoO Sigma Schedule Designer - hands you that curve as an editable object. Sigmas in, sigmas out, with four bands of the schedule you can re-weight without touching your step count. It's the difference between "use karras" and "spend three more evaluations on microdetail and two fewer on large forms."

Where it fits

In a stock SDXL graph you never see a sigma tensor; the sampler makes one from the dropdown. Move to SamplerCustomAdvanced and sigmas become a real wire - the wire this node feeds.

The interesting case is a two-stage handoff: a first sampler ends around sigma 1.6, where composition and big forms are settled, and a second one takes the low end and does detail work. The Designer has a mode for each role and a numeric handoff_sigma to bind them together.

One honest caveat. Reshaping is a big win on SD1.5/SDXL-class noise schedules and a weak tool on rectified-flow models, which are already trained on a straight trajectory; the better sampler sweeps keep landing on the same conclusion. Nudge the curve, don't reflavor it.

How it works

The order of operations explains most of its behavior.

1. Source. If usable SIGMAS are wired, it uses them (range_source defaults to AUTO / INPUT). Otherwise it fabricates a reference from steps and max_sigma.
2. Range. sigma_range_mode picks FULL RANGE, HIGH SIGMA (stops at the handoff) or LOW SIGMA (starts there). IMG2IMG REFINE mode crops by denoise_active_range on top of that.
3. Shape. The profile is stored as normalized (progress, remaining sigma) anchors, so one shape scales to a 14.6 full schedule or a 1.6 refinement tail. The band sliders move where those anchors sit; curve_strength decides how far the target departs from the reference.
4. Smooth and validate. smoothness flattens kinks, then the toggles fire: monotonic lock, duplicate repair, endpoint pinning, clamping to the input range.
5. Blend back. reference_blend pulls the result toward the original schedule, then it validates again.

The inputs you'll actually touch

mode first: AUTO works out from the incoming schedule whether you're a first or second sampler, and FIRST SAMPLER / SECOND SAMPLER / IMG2IMG REFINE / CUSTOM override it. Older workflows using Auto, TXT2IMG or Manual still load - legacy names get migrated.

profile is the vocabulary: 15 current presets (Balanced Generation, Photoreal Base, Detail Refiner, Microdetail, Turbo Tail, Identity Refine…), the legacy names they replaced, and a Custom slot. Choosing one writes its band values into the sliders; from there you fiddle.

Then curve_strength and reference_blend, the two knobs most people should be turning. High strength with low blend is a big departure from what the model expects. High blend with moderate strength is a gentle re-weight that rarely breaks anything.

steps and max_sigma are pure fallbacks - AUTO / INPUT ignores them whenever incoming sigmas are usable, so don't tune them while wired up. Everything else lives behind ui_complexity; leave it on BASIC until you have a reason.

The outputs

sigmas goes into SamplerCustomAdvanced (with a KSamplerSelect for the algorithm), or into SplitSigmas if you're building a handoff. A plain KSampler won't take it; that node makes its own schedule.

sigma_info is a long STRING of diagnostics: source used, whether it fell back, first and last values in and out, tensor length, finite/monotonic flags, every repair applied. Wire it into any string consumer. It answers "why did nothing change" in one line, because it literally reports input_equals_output.

Install

Manager, searching koo-nodes-pack (publisher KoO.Ai; registry availability depends on moderation approval). Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/kevintronchin-blip/koo-nodes-pack

Restart and hard-refresh. requirements.txt only declares torch, numpy, Pillow and aiohttp, all shipped with ComfyUI. No downloads; the pack never pip-installs at runtime.

Where people get stuck

The curve editor ignores your clicks. Manual point editing unlocks only in CUSTOM mode with ui_complexity on PRO. The status line says so; clicking on a preset profile does nothing until you switch.

Your input was ignored. If incoming sigmas aren't strictly decreasing - a cropped tail, a split range fed back in - the node treats them as unusable and falls back to steps/max_sigma. sigma_info says exactly that.

Nodes 2.0. The Designer draws its own curve canvas, and the README is upfront that node-level canvas drawing and mouse hooks need the Legacy/LiteGraph renderer. On Nodes 2.0 you get a node with no editor.