Nodes/⚗ AlchemicaSampler/⚗ Alchemica Options · Schedule
ComfyUI Node

⚗ Alchemica Options · Schedule

The step-scheduler control panel, for when the calibrated profile isn't enough

By Nessuno76·Created about a month ago·Updated 21 days ago· 4
⚗ Alchemica Options · Schedule
  • opts
  • opts
◄scheduleralchemica_dual►
◄steps8►
◄dual_high1.00►
◄dual_width_high0.10►
◄dual_low8.0►
◄dual_width_low0.08►
◄calibrated_exponent0.50►
◄calibrated_mix0.15►
◄shift_override0.00►
◄flux_q1.00►
◄plunge_sigma0.00►
◄calibrated_v2false►

The Alchemica samplers deliberately hide most of what they're doing. This node is the hatch: it exposes the step-scheduling machinery as widgets, and its single output overrides every scheduling setting on the sampler it feeds.

Which is the first thing to understand about it. Connect opts to a sampler's opts input and the option node wins - including over the sampler's own steps, contraction and cfg. That's not a bug, it's the design, and it's how people confuse themselves for an afternoon.

What it exposes

scheduler - the default is alchemica_dual, and this widget decides which of the other knobs mean anything.

  • alchemica_dual - two lobes of extra steps: one at high sigma for composition, one at low sigma for the final detail. dual_high (1.0) and dual_low (8.0) are how much extra each lobe adds; dual_width_high (0.1) and dual_width_low (0.08) are how far each lobe reaches. This is the schedule you get when there's no calibration profile - a sensible shape, not a measured one.
  • alchemica_calibrated - the profile measured on your checkpoint, with two dials on top: calibrated_exponent (0.5 is the theoretical optimum for Euler; higher puts more steps where the model changes most) and calibrated_mix (0.15 = the share of steps spread uniformly rather than concentrated).
  • The fixed curves - flux_shift (flux_q above 1 = more steps at the top, below 1 = more at the bottom) and linear_flow, plus anything ComfyUI registers.
  • shift_override - flow-mu. 0 means "the model's own", and Krea 2's is 1.15. Raise it and you get more of the budget at high noise; lower it and you get more at the end, which is where detail lives.
  • plunge_sigma - 0 = off. Set it and structure stops at that sigma, with the last step jumping straight to 0. It's a curiosity, not a default; the calibrated path handles the tail better.
  • calibrated_v2 - the interesting one. On, the calibrated schedule gets the solver-order step law (cube root of the third derivative for second-order solvers like euler_2m and heun, instead of the square root Euler-optimal law), smoothing in log-SNR, and half-resolution curves for stage 1 of two-stage sampling. It needs a profile calibrated after 2026-10-10; on older profiles only the smoothing applies. Flex turns this on by default, everything else leaves it off so 1.0 schedules stay byte-identical.

Input and output

steps (8 by default) sits on the node too, because a scheduler without a step count is half a schedule. The optional opts input is a chain: option nodes stack, each one inheriting from the last and overriding what it sets. That's how you get a schedule node plus a restart node plus a two-stage node all feeding one sampler through a single wire.

Output: opts (an ALCHEMICA_OPTS). One wire into the sampler.

When you actually want it

Three honest use cases.

  1. You want more steps at the end without touching the profile. Drop shift_override slightly below the model's own value.
  2. You want to try the v2 step law on a sampler that doesn't default to it. Flip calibrated_v2 on here and let the option chain carry it.
  3. You're building something the sampler's presets don't cover - a linear flow schedule at a step count nobody shipped, a dual-lobe shape with a different balance. This is the node for that.

If none of those describe you, you don't need this node. That's a real answer, not a dismissal; the calibrated profile with mode on balanced is doing more than any hand-tuned lobe.

Installing

# ComfyUI Manager: search "AlchemicaSampler"
cd ComfyUI/custom_nodes && git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

ComfyUI 0.36+, no pip dependencies, restart the server. It lives under ⚗ AlchemicaMente/sampling/options, and the 1.1 node is the English twin of the 1.0 options node with the Italian widget names.

Where people get burned

Forgetting the override. Your sampler says 8 steps and detail 0.12, and you're queueing, and nothing changes, because the option node is upstream and overriding the schedule. Disconnect the opts wire to A/B it.

Expecting calibrated_v2 to do anything without a fresh calibration. It says so on the tooltip. On a pre-October profile you get the smoothing, which is not nothing, but it isn't the step law.

Scheduler names that aren't there. The list is read at startup from the pack plus ComfyUI and its other packs. Restart after installing anything that adds a scheduler, or it won't appear.

Category⚗ AlchemicaMente/sampling/options

Inputs (13)

NameTypeDefaultDescription
schedulerCOMBOalchemica_dual13 options: alchemica_dual, alchemica_calibrated, flux_shift, linear_flow, comfy:simple, comfy:sgm_uniform, +7
stepsINT81–200Total steps (single stage)
dual_highFLOAT1.000–20alchemica_duale: extra steps on composition (high sigma)
dual_width_highFLOAT0.100.01–1alchemica_duale: how far the high lobe reaches down
dual_lowFLOAT8.00–40alchemica_duale: extra steps on the final detail (low sigma)
dual_width_lowFLOAT0.080.01–1alchemica_duale: how far the low lobe reaches up
calibrated_exponentFLOAT0.500–1.5alchemica_tarato: 0.5 = theoretical optimum for euler. Higher = more steps where the model changes most
calibrated_mixFLOAT0.150–1alchemica_tarato: share of steps spread uniformly
shift_overrideFLOAT0.000–5Shift mu (0 = the model's own, 1.15 for Krea 2)
flux_qFLOAT1.000.3–3flux_shift: >1 more steps at the top, <1 at the bottom
plunge_sigmaFLOAT0.000–0.950 = off. Otherwise structure stops here and the last step jumps to 0
calibrated_v2BOOLEANfalsealchemica_tarato v2: step law matched to the solver order (cube root of the 3rd derivative for euler_2m / heun), smoothing in log-SNR, and stage 1 of two stages uses the curves measured at half resolution. Needs a profile calibrated after 2026-10-10 for the new curves (otherwise only the smoothing applies)
optsoptALCHEMICA_OPTSChain from another Alchemica options node

Outputs (1)

NameTypeDescription
optsALCHEMICA_OPTS—