⚗ AlchemicaKrea (all-in-one) [1.0, use Alchemica KSampler (Krea 2, all-in-one)]

Ask five Krea 2 users which sampler and scheduler to use and you get five answers - er_sde + simple, euler_ancestral_cfg_pp + beta, lcm + simple, dpmpp2m + sgm_uniform, "12 steps is the safe zone", "8 steps is the safe zone", and a running argument about whether anything above 1200 px per side holds together. That's not a community failing. Nobody shipped the step schedule Krea 2 Turbo actually wants, because a distilled model doesn't document one.

AlchemicaKrea is the pack's answer: a single node that takes a calibration profile measured on your checkpoint and turns it into a step schedule and a set of defaults, while leaving you the handful of knobs that visibly change the image.

What it does

Krea 2 uses a flow-matching noise schedule (sigma in [0,1], x = σ·ε + (1−σ)·x₀), which means step placement is a real degree of freedom rather than an inherited default. The pack's calibrated scheduler derives a step density from the model's own behaviour - more steps where the prediction is changing fastest - and this node applies it along with a two-stage trajectory, a bounded detail boost inside the measured sharpness window, and gated ancestral noise. The mechanism is described properly on the ⚗ Calibration page; the short version is that the profile tells the sampler where composition locks (≈sigma 0.87) and where detail is born (sigma 0.57 down to 0.13).

The knobs you'll actually turn

profilo - (auto) recognises the checkpoint and uses its profile; (nessuno) means none. No profile is not an error: everything falls back to an analytic two-lobe schedule and default values, and the report says so. It just won't be tuned to your file.

modo - the step budget. veloce (fast, ~30 s at 3 MP), equilibrato (balanced, ~45 s, the default), massimo dettaglio (~70 s). Internally these change step counts and solver: fast runs 3 composition + 6 refinement steps on plain Euler, balanced runs 4 + 8 on euler_2m with 2 restart steps, max detail runs 5 + 12 with 3 restart steps.

due_stadi - compose at half resolution, refine at full. On by default, and you want it on above ~1.5 MP. This is the same idea as a hires-fix pass, except it's built into the trajectory rather than bolted on after.

ridisegno - how freely the second stage may redraw, i.e. the hand-off sigma. −1 means "from the profile" and that's the recommended value. Raise it and you get more new detail at the risk of the composition shifting.

dettaglio - micro-relief on skin and fabric. 0.12 is the validated value; the author measured 0.40 putting coloured speckles across skin. Above 0.2 the speckles come back.

varieta - the ancestral-noise amount that makes texture and separates seeds. 0.6 is the default and is clean; 0 means deterministic.

denoise - 1.0 is txt2img. Below 1 turns this into an img2img sampler, and it switches two-stage mode off, because you can't half-compose something that's already been composed.

Optional sockets: negative (only consulted when you've set CFG above 1 through the advanced option nodes), modello_rifinitura (compose with the main model, refine with a second one - anatomy from a finetune, skin from base Krea 2; needs two-stage mode), and opts (any ⚗ options node, which overrides the preset wholesale - "for tinkerers", in the author's words).

Outputs

latent → VAEDecode. grafico_sigma is an IMAGE plot of the schedule you just ran, which is worth glancing at when a run misbehaves - you can see whether the steps clustered where you thought. report is a STRING with the resolved settings, and yes, it's in Italian; the README's glossary covers the vocabulary.

Installing

ComfyUI Manager: search AlchemicaSampler. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

Restart the server - reloading the page is not enough for this pack. No pip install (the project declares zero dependencies), but you need ComfyUI ≥ 0.36.0 and the three Krea 2 files in place. The fastest start is to open Workflow → Browse templates → ComfyUI-AlchemicaSampler → KREA2_SEMPLICE, which is this node plus a LoRA stack.

Traps

Running with no profile and blaming the node. It works, it just isn't calibrated. The report tells you which case you're in; the cheap fix is one ⚗ Calibration run per checkpoint.

Expecting the negative prompt to do anything here. Not at CFG 1. That's a job for NAG (⚗ Attention guidance), or for running Krea 2 Raw at real CFG.

Turning dettaglio up because "more detail". This is the single most common way to make an Alchemica image worse. 0.12.

Out of memory. 5 MP on a 16 GB card is the usual culprit, and modo veloce is the quickest way down.