⚗ Alchemica · Text (scene / detail / negative) [1.0, use Alchemica Text Encode (scene / detail / negative)]

Here's the thing about running Krea 2 Turbo: at CFG 1 your negative prompt box is decorative. ComfyUI doesn't merely down-weight the unconditional pass at exactly 1.0, it skips computing it - math.isclose(cond_scale, 1.0), and the slot stays a zero tensor the arithmetic cancels. So a node that hands you a negative conditioning is only half a story, and this one is honest about its half: the negative it emits is for NAG, and nowhere else.

What it is

A text encoder with the pack's prompt philosophy baked into its wiring. You type three fields, and it returns four outputs:

• positive_scena - the scene text, encoded. Subject, pose, framing, setting, light, colours, style. This is what decides the composition, because Krea 2 Turbo locks composition and palette in the first step.
• positive_dettaglio - the detail text, encoded. Rendering only: skin, fabric, materials, optics, grain.
• negative - encoded from the negativo field. Feeds ⚗ Attention guidance (NAG). At CFG 1 it has no effect anywhere else, and if you leave it empty the node encodes a single space rather than failing.
• testo_dettaglio - a plain STRING of whatever was actually encoded for the detail slot, so you can read back what the model saw.

unione, the field that matters most

A two-item dropdown and no third option, and it changes what the detail conditioning is:

• aggiungi alla scena (add to scene, the default) - the detail conditioning is the scene text plus the detail text, concatenated. This is the recommended setting and the reasoning is mechanical: the detail prompt on its own is a list of surface qualities with no subject in it, and below the hand-off sigma the model has to know what it's rendering. Without the scene in there, the model forgets the subject at exactly the point where it's committing to texture.
• solo dettaglio - the detail text replaces the scene text entirely for the second phase.

You'd use the second one rarely, if ever, and mostly to test something.

Why split the prompt at all

Because of a measurement. Krea 2 Turbo decides composition and palette in the first step - 98.5% of the composition is locked at sigma ≈ 0.87, and sharpness lives between sigma 0.57 and 0.13. That means a prompt handed to the model late can change texture but cannot change the picture, and it means the useful thing is to give the model different text at different depths. The author's own two-phase test: keeping only the scene prompt produced an image 0.996–0.997 similar to the two-phase result, whereas one merged prompt with the same words scored 0.35–0.71 - a different picture. The split is not cosmetic.

Wiring

positive_scena → the positive input of AlchemicaKreaPro (or the advanced AlchemicaSampler). positive_dettaglio → AlchemicaKreaPro's positive_dettaglio. negative → AlchemicaAttenzione's negativo input.

The important bit: AlchemicaKrea (the all-in-one) takes only one positive. It's AlchemicaKreaPro that has the second socket. If you're not using NAG, the negative output here is dead weight - use your sampler's negative socket instead, which the advanced samplers only consult when negative_mode is text and CFG is above 1.

Installing

Ships with the pack:

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

Restart the server. Nothing to download, no dependencies - the pack's pyproject.toml has an empty dependencies list, and the only thing it needs beyond ComfyUI is torch.

Traps

Writing negations into the positive text. On a CFG-1 model, "no watermark, no text" in the positive box is not neutral - the author's enhancer explicitly strips negations from the positive for this reason. Say what you want present.

Treating the negative as a normal negative. At CFG 1 it does nothing at all unless NAG is patched in and given this conditioning. If you want classic negative behaviour you need a model running real CFG (Krea 2 Raw is that model - the pack's raw_base preset runs it at CFG 3.5 with negative_mode: text), or NAG.

Expecting the detail conditioning to change colours. It won't. Colours and global style are settled in the first step, above the hand-off sigma, from the scene prompt. Put them there.