⚗ Alchemica KSampler Flex (any sampler + scheduler)

The all-in-one Alchemica KSampler gives you the pack's sampler and schedule and no choices. Flex gives you the same machinery with the front door open: pick any solver ComfyUI has (or any other custom pack has added), pair it with the calibrated schedule measured on your checkpoint, or with any of ComfyUI's schedulers.

It's the sampler in the pack's recommended starter workflow, so this isn't the exotic one. It's the default one.

The three-way split

The interesting bit is that sampler_name decides what the other widgets even mean.

alchemica_euler, alchemica_euler_2m, alchemica_heun - the pack's own loop, where detail, variety and restart_steps do something. detail adds micro-relief to skin and fabric (0.12 validated; above 0.2 coloured speckles come back). variety is ancestral noise during composition, and it has a documented trap: with a strong LoRA, above ~0.3 you can get wet-skin droplet artifacts - drop to 0.15–0.2. restart_steps re-noises and redoes the tail of the trajectory with extra steps, from the profile's restart sigma.

Any other solver - euler_flow, res_multistep, dpmpp_2m, whatever you have installed. It runs as ComfyUI runs it, and detail, variety and restart_steps are ignored (the interface hides them, which is a genuinely nice touch).

Schedulers - alchemica_calibrated is the step spacing measured on your checkpoint; without a profile it falls back to alchemica_dual. alchemica_dual puts extra steps in two lobes (composition at high sigma, detail at low). flux_shift and linear_flow are fixed curves, and then it's ComfyUI's own list: simple, beta, bong_tangent, all of it.

The profile widget feeds alchemica_calibrated and every -1 value, with any sampler or scheduler. So you can run euler_flow + beta and still have the pack fill in the measured hand-off sigma for you.

Widgets worth understanding

• steps 8, cfg 1 - the Krea 2 Turbo numbers. Above CFG 1 a connected negative gets used.
• two_stage on, stage1_steps 4, redraw -1. Composition at half resolution, refinement at full.
• contraction 0.85, shift 0 (the model's own mu; Krea 2's is 1.15). shift higher = more steps at high noise, lower = more at the end.
• handoff_k 0 = auto - a resolution correction the other samplers don't have. The hand-off sigma in your profile was measured at full resolution, but stage 1 ran at half, where that sigma means less. Auto converts it (0.87 → 0.770 for a 2× factor), so stage 2 gets a composition as settled as the profile measured. 1 = off, matching the plain Alchemica KSampler; above 1 = your own explicit factor.
• calibration - v2 is the default here only: step law matched to the solver order (cube root of the third derivative for second-order solvers), smoothing in log-SNR, and half-resolution curves for stage 1. It needs a profile calibrated after 2026-10-10; older profiles get the smoothing only.
• lowfreq_anchor 0 - marked EXPERIMENTAL, off by default, and worth leaving off unless you're chasing the problem it solves. At each stage-2 step it replaces the low frequencies of the model's x0 prediction with those of the upscaled stage-1 image (an idea borrowed from HiFlow, NeurIPS 2025), fading out as sigma falls. It protects composition while stage 2 adds detail. Full strength = 1.
• degrid 2 - note this defaults to 2 here, unlike the other samplers' 0. With a vae connected, the final decoded image gets the 2-pixel VAE grid removed before it's shown and returned. The latent is never touched.

Outputs: latent, sigma_plot, report, image.

Sanity check worth running once

With every improvement switched off, Flex comes out pixel-identical to the all-in-one KSampler in balanced mode - the authors measured a difference of 0. So when a Flex run looks worse than your old workflow, it's a widget you changed, not the node. That's unusually nice, and it makes debugging a matter of turning things off one at a time.

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. Solver and scheduler lists are read live, so anything another pack registers shows up without a restart of its own.

Common issues

detail and variety do nothing. You're on a non-Alchemica sampler. Switch to one of the alchemica_* three, or leave them alone - they're ignored by design, not broken.

The negative is ignored. cfg is at 1. That's the distilled model behaving as designed; use NAG instead of CFG if you need a negative.

Everything is a bit slow with lowfreq_anchor on. It costs model work per stage-2 step. Turn it off to confirm it's what you're paying for.