⚗ Alchemica Sampler Select (SAMPLER)

Sometimes you want the pack's solver without the pack's house. This is that: the Alchemica sampling loop, wrapped as a SAMPLER object and handed to ComfyUI's own SamplerCustom.

What it is

AlchemicaSamplerSelect is the SAMPLER half of AlchemicaSampler's custom-sampling pair - the other half being ⚗ Alchemica Scheduler, which produces the SIGMAS. Together they let you sample with the Alchemica recipe (solver, detail boost, ancestral noise, restart steps) inside ComfyUI's generic sampling nodes, where you control the guider, the sigmas and the noising yourself.

Why bother? Because SamplerCustom / SamplerCustomAdvanced is where every interesting experiment lives: a custom guider, hand-built sigmas, a two-stage schedule you assembled by hand, or a latent that someone else's node produced. If you need one of those and still want the calibrated detail behaviour, this node is the door.

What you don't get is the pack's pipeline wrapper: no two-stage orchestration, no decode, no sigma plot, no report. It's a solver, nothing more.

How it works

Picking a preset here selects the solver-side recipe out of the same ten presets the rest of the pack shares - the step scheduler, detail boost amount and window, ancestral noise (eta), the gated variety window, contraction and restart steps. profile decides where the auto values come from: (auto) recognises your checkpoint and reads its calibration profile, (none) means no profile at all. With a profile in play, any preset entry set to −1 means "take it from the measurement".

noise_seed is the node's own seed, used for the ancestral noise the loop injects when the preset asks for it. It's not the seed that chooses your image - that's the noising step in SamplerCustom - so you can hold one fixed and vary the other, or freeze both for reproducible runs.

opts is the optional override socket: chain an Alchemica Options node into it and it replaces the matching entries of the preset, one group at a time. That's how you take alchemica_turbo and change only the detail group without hunting for an option node per parameter.

The practical rule for wiring it: use the same preset name on this node and on ⚗ Alchemica Scheduler next to it, unless you have a specific reason not to. The scheduler decides the step spacing; the sampler decides how each step is solved and how much noise it tolerates. A preset encodes both halves, so matching them keeps the pairing the author actually tested. Then sigmas from the scheduler goes into SamplerCustom's sigmas input, sampler from this node goes into its sampler input, and the guider and latent come from wherever you like.

Inputs and outputs

• preset (enum, ten), profile (enum, (auto) / (none)), noise_seed (INT, with the usual after-generate control).
• Optional opts - the chainable override.
• Output: sampler (SAMPLER). One socket, straight into SamplerCustom or SamplerCustomAdvanced.

Installing it

ComfyUI Manager, search "AlchemicaSampler", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler
# restart the ComfyUI server

ComfyUI 0.36.0 or newer, no extra Python packages. The SamplerCustom nodes it feeds are core ComfyUI, so there's nothing else to install. You'll want the Krea 2 model, Qwen3-VL text encoder and Qwen-Image VAE in place to have anything to sample.

Where it goes wrong

• Weird, washed-out or noisy results. Mismatched pairs. If the sigmas came from a photo_calibrated preset and the sampler is running alchemica_turbo, you've asked for steps spaced for one recipe and solved for another.
• (auto) found no profile. The calibration profile is per checkpoint and created on your machine - it isn't shipped with the pack. Run the calibration workflow once, or set (none) and accept the analytic schedule.
• The detail or variety knobs do nothing. They only bite on presets that enable them; several are plain Euler with no extras, deliberately.
• Nothing shows a preview. This node builds a sampler object, not a pipeline. Previews, plots and reports belong to ⚗ Alchemica Sampler (advanced).