⚗ AlchemicaKrea Pro (phases + previews) [1.0, use Alchemica KSampler Pro (two-phase prompt, previews)]

Composition is decided in the first step of a Krea 2 Turbo run. So the dumbest possible workflow is the one most people use: burn a full 4 MP two-stage generation, look at it, decide the framing is wrong, change the seed, repeat. AlchemicaKrea Pro argues you should be able to see what the composition will be before you spend the rest of the trajectory on it - and it turns out you can.

What it adds over AlchemicaKrea

The model, positive, latent_image, seed, profilo, modo, due_stadi, ridisegno, dettaglio, varieta, denoise inputs are all the same node you already know. Four things are new.

positive_dettaglio - the optional second-phase conditioning, from ⚗ Two-phase prompt. Leave it disconnected and this behaves exactly like AlchemicaKrea.

cambio_fase - the sigma at which the detail prompt takes over. −1 means "the calibrated hand-off", where composition is locked (~0.87). Lower pushes the detail prompt later, so it weighs less; 0 turns the second phase off. The author is explicit about its limit: it acts on skin, fabric, patterns and small objects, not on colours or global style, because those are decided in the first step together with the composition.

sfumatura - cross-fade width between the two prompts. 0 is a hard switch and costs nothing. Above 0 it blends gradually and costs one extra model call per step inside the band.

anteprime / scelta / variante_rifinitura - the interesting trio, below.

Previews: the feature worth the install

Set anteprime to N and the node generates N low-resolution images using seeds seed, seed+1, seed+2 … The stage 1 of each preview is exactly the stage 1 of the corresponding final image, and stage 2 is redone small with the same noise field. So the preview isn't a rough sketch of approximately the right thing - it's the same trajectory, truncated in resolution.

The measured similarity between preview and final is 0.87–0.90 across seeds, and the author is honest about the rest: pose and framing match, clothing may change.

The workflow is a two-pass ritual:

1. Set the seed widget to fixed, put anteprime at 8, queue. You get 8 cheap compositions.
2. Pick the one you like - say number 5. Set anteprime back to 0 and set scelta to 5. Queue.

Because preview k used seed + k, scelta = k reconstructs it exactly. That's the whole trick, and it's the reason scelta exists as a separate field rather than being inferred.

variante_rifinitura does something different and more subtle: it changes only the second-stage noise. Same pose, same framing, but face, lettering, jewellery and small objects shift. Use it when the composition is right and the small stuff isn't - which is a much more common situation than most samplers admit.

Outputs

latent, grafico_sigma (the schedule plot), report (Italian, as usual), and seed_usato - the seed actually used, which is what you want if the preview/chosen path or the refinement variant moved things around. Wire it into a Save Image filename node and you'll never lose a good seed again.

Wiring the two-phase path

⚗ Two-phase prompt · positive_scena ────→ AlchemicaKrea Pro · positive
                    · positive_dettaglio → · positive_dettaglio
⚗ Attention guidance · model ──────────→ · model
⚗ Resolution · latent ─────────────────→ · latent_image

Installing

Part of the pack:

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

Restart the ComfyUI server (page reload won't do it). Requires ComfyUI 0.36.0+ and the Krea 2 model, Qwen3-VL encoder and Qwen-Image VAE. The example workflow KREA2_PRO is the intended way to meet this node.

Traps

Expecting previews to work with denoise below 1. They don't. img2img disables two-stage mode and previews together, for the same reason: there's no composition phase left to preview.

Preview seeds vs the scelta field. scelta = 0 is the "just finish preview 0" case, not "off". If you set previews to 0 and leave scelta at 0, you get the base seed - which is usually what you want, but it means the two fields interact and neither is a plain on/off switch.

Comparing a preview to the final pixel by pixel. 0.87–0.90 is the honest number. The composition is the same; the paint isn't.

Profile picking. (auto) chooses, among the profiles of this checkpoint, the one calibrated at the closest resolution. If you calibrated at 3 MP and generate at 5 MP, you're getting the 3 MP profile and you should know that.