⚗ AlchemicaSampler (advanced) [1.0, use Alchemica Sampler (advanced, presets + options)]

AlchemicaKrea gives you a mode switch and five knobs. AlchemicaKreaPro adds two-phase prompts and previews. The advanced AlchemicaSampler gives you the machine underneath both: a preset, a profile, and fourteen extra parameters reached through a chain of small option nodes that plug into one socket each.

What it does

It samples Krea 2 Turbo from a named preset. The preset is a flat dictionary of every setting the sampler understands - scheduler, solver, step counts, detail boost window, ancestral noise gates, restart, CFG curve, two-stage geometry - and the option nodes override only the keys in their own group. Values of −1 mean "take it from the calibration profile", which is where the profilo field comes in.

Nine presets ship with it, and they're worth reading as documentation:

• rif_euler_simple - ComfyUI's own simple scheduler and Euler, no Alchemica scheduling at all. The control group, for comparisons.
• alchemica_duale - the analytic two-lobe schedule (extra step density over composition and over final detail, nothing measured). Works with no calibration.
• alchemica_turbo - the default. Duale schedule plus detail 0.12, ancestral noise 0.5, contracted initial noise 0.85.
• alchemica_quality - 10 steps, euler_2m, stronger ancestral noise, 3 restart steps.
• alchemica_hires_2stage - the two-stage geometry as a preset: stage 1 at half side, 5 composition steps, hand off at 0.65, 6 refinement steps.
• foto_tarata / foto_tarata_2K - the calibrated ones. Scheduler alchemica_tarato, solver euler_2m, detail window and restart sigma taken from your profile. These are the ones that actually benefit from a ⚗ Calibration run, and the preset's own notes cite the measurements they encode.
• plunge_restart - a rougher recipe: flux_shift schedule, plunge_sigma 0.75, restart at 0.65, detail 0.60, dual CFG.
• raw_base - for Krea 2 Raw: 28 steps, euler_2m, CFG 3.5 both ends, and negative_mode: text, i.e. a real negative prompt because Raw runs real CFG.

The inputs

model, positive, latent_image (at the final resolution - in two-stage mode stage 1 works at a fraction of it), seed, preset, profilo ((auto) recognises the checkpoint; needed by the foto_tarata presets and by every −1 value), denoise (1.0 = from scratch; below 1 is img2img and disables two-stage mode).

Optional: negative (only used when negative_mode = text and CFG is above 1), modello_rifinitura (second stage with a different model), and opts - the chain socket. Its tooltip is blunt: it overrides the preset entries.

The option chain

Each of the five option nodes - ⚗ Schedule, ⚗ Solver & Detail, ⚗ Restart, ⚗ Guidance, ⚗ Two stages - has an optional opts input and an ALCHEMICA_OPTS output. The input is there so you can stack them:

Alchemica · Schedule ─opts?─ Alchemica · Restart ─opts?─ Alchemica · Guidance ─opts→ AlchemicaSampler · opts

Each node overrides only its own parameter group, so order between them doesn't matter - the last writer wins only within a group. If you connect ⚗ Guidance, you get its CFG curve and its negative_mode; if you connect ⚗ Two stages, you get its stage geometry. Disconnect a node and the preset's values for that group come back, which makes A/B testing genuinely easy: one wire.

Outputs

latent → VAEDecode. sigmas → anything that wants a SIGMAS (SamplerCustom, or just to inspect), and it's the single-stage sigmas, oddly, so don't feed it back into this node. grafico_sigma is the schedule plot as an IMAGE. report is a STRING, Italian, listing the resolved values including which ones came from the profile and which fell back.

Installing

Part of the pack:

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

Restart the server. ComfyUI 0.36.0+, no Python dependencies, Krea 2 + Qwen3-VL + Qwen-Image VAE on disk. The example workflow KREA2_ALCHEMICA wires this node with all its option nodes; KREA2_ALCHEMICA_CONFRONTO runs the same seed through Comfy's simple scheduler, the analytic Alchemica curve, and the calibrated one - genuinely the fastest way to decide whether the whole calibration idea is worth 20 minutes of your GPU time.

Traps

Connecting an option node and forgetting what it clobbers. opts doesn't merge politely with your preset - it replaces that group's values. If a run suddenly behaves like the default preset, check the chain.

Using foto_tarata without a profile. It resolves to fallbacks and tells you in the report, but you're paying complexity for nothing. Calibrate first, or use alchemica_duale.

Feeding it a small latent for the first stage. You give it the final size; it does the downscale.

Chasing the community's sampler rankings. Krea 2 threads are full of "er_sde + simple", "12 steps minimum", "euler_ancestral_cfg_pp + beta" - all reasonable, all different, none of them derived from your checkpoint. This node's entire premise is that you can measure the answer instead.