⚗ Alchemica Prompt (enhance + encode + NAG) [old, use Alchemica Prompt (idea / list / Ollama + NAG + latent)]

This node has a slightly awkward place in the AlchemicaSampler lineup, and understanding it saves you twenty minutes of wondering why you can't find it in the search box.

What it is

AlchemicaPromptAllInOne is the English restyle of the pack's original all-in-one prompt block: text encoder writes the prompt (or takes yours as written), encodes scene plus detail, patches NAG onto the model. One node between the loader and the sampler.

Then the pack moved on, and this became the middle child. Its display name says so in full: "⚗ Alchemica Prompt (enhance + encode + NAG) [old, use Alchemica Prompt (idea / list / Ollama + NAG + latent)]". The replacement is AlchemicaPromptStudio, which keeps this behaviour and adds an improve prompt mode, list/string/example sources, wildcards, an optional Ollama backend, and the start latent.

So: it's hidden from search, still registered, and it opens and runs old workflows unchanged. If you're building something new, skip to PromptStudio. There is no reason to pick this one on purpose.

What it does

Two modes. generate from idea hands your few words to Krea 2's own Qwen3-VL encoder, which writes a SCENE paragraph, a DETAIL paragraph about surfaces and skin, and a NEGATIVE list; the encoded result covers scene plus detail, and the text shows up in the report. as written skips the model entirely and encodes your scene, detail and negative_text verbatim.

style (eight photographic registers), words (scene length) and creativity (0 = deterministic, 0.6–0.9 = variations, at which point seed matters) all only apply in the generate mode. The optional image socket gives the encoder a visual reference - Qwen3-VL is a vision-language model, so with an empty idea and an image connected, it describes the image. skin decides how detail talks about skin: smooth (no body hair) strips body-hair words from the prompt and pushes them into the negative, natural (fine vellus hair) is the original 1.0 behaviour, model decides says nothing. instructions is where you put your own rules in any language; they win over the built-in ones when they conflict.

The nag_* group is the reason this node exists rather than a plain CLIP Text Encode. Krea 2 Turbo runs at CFG 1, and at CFG 1 the negative prompt is inert - there is no unconditional pass to steer. NAG applies the negative in attention space instead, and this node hands you the knobs: nag_scale (calibrated at 3), nag_tau capping the push away from the positive, nag_alpha blending back toward the original attention, nag_blocks for which DiT layers participate, and nag_sigma_start / nag_sigma_end for the window. Leave nag_sigma_start at −1 and it begins at the calibrated hand-off sigma, which is the setting that removes things without changing the composition. structure adds PAG (sharper, more contrast) or SEG (blurred queries; expect halos above 0.3).

Inputs and outputs

In: clip, model, and the text widgets. Out: model with NAG and structure applied - wire that one to the sampler, not the loader - plus positive, negative (only relevant if you raise the sampler's cfg above 1) and report. It's an output node, so the report is visible on the canvas.

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, because the prompt writing uses CLIP.generate. No extra Python packages; you do need the Krea 2 model, Qwen3-VL text encoder and Qwen-Image VAE in their model folders.

Where it goes wrong

• You can't find it in the node search. You shouldn't be able to - it's deprecated and hidden. Use ⚗ Alchemica Prompt (idea / list / Ollama + NAG + latent).
• The prompt never changes when you tweak the negative. The enhancer's answers are cached in memory against the rules, your text, the image and the creativity/seed, so changes to NAG or resolution don't rewrite the prompt. That's a feature.
• It wrote no DETAIL. The report says so explicitly when the model returned an incomplete answer, and shows the raw text.
• Negative prompt does nothing. With nag_scale at 0 the model passes through untouched, which is exactly the default-configuration behaviour of a distilled model.