ComfyUI-AlchemicaSampler

<p align="center"> <img src="assets/alchemicamente.svg" width="140" alt="⚗ AlchemicaMente"> </p> <h1 align="center">⚗ AlchemicaSampler</h1> <p align="center"> <b>A sampler that measures your Krea 2 checkpoint, then samples the way the model actually works.</b><br> Calibrated schedules · two-phase prompts · composition previews · a prompt enhancer built on Krea 2's own text encoder · NAG, a negative prompt that works at cfg 1 </p> <p align="center"> ⚗ Idea and direction: <b>AlchemicaMente</b> · code: <b>Claude (Anthropic)</b> · license: <b>GPL-3.0</b> · <a href="CREDITS.md">credits & thanks</a> · <a href="README_IT.md">manuale in italiano</a> </p>

---

⚗ What it is

Krea 2 Turbo is a distilled model: 8–14 steps, cfg 1. Nobody publishes the step schedule it "likes", and at cfg 1 the negative prompt does nothing. AlchemicaSampler takes the question to the model itself:

1. It measures your checkpoint on your GPU. A dense sampling run records, sigma by sigma, when the composition locks, where fine detail is born and how fast the prediction changes. From that it builds a step schedule and a profile that is matched to the checkpoint automatically.
2. It samples with those measurements: two stages (compose at half resolution, refine at full resolution), a detail boost inside the measured window, gated ancestral noise, restart.
3. It separates what the model decides early from what it decides late: a scene prompt for the composition, a detail prompt for the rendering, and a negative prompt (NAG) that acts only after the composition is locked, so it removes what you don't want without changing the picture.

No dependencies beyond ComfyUI itself (torch, PIL, numpy).

⚗ Highlights (measured on an RTX 5070 Ti, Krea 2 Turbo NVFP4, 4 MP)

| Feature | What we measured | |---|---| | NAG from the calibrated hand-off sigma | unwanted water drops removed, image 99.3 % identical to the original, about +40 % time on the refinement stage | | Two-phase prompt | composition 0.996–0.997 similar to the scene-only image; a single prompt with the same words gives 0.35–0.71 (a different picture) | | Composition previews | preview → final similarity 0.87–0.90 (different seeds: 0.45); preview stage 1 is identical to the final one | | AlchemicaKrea Pro with no detail prompt | pixel-identical to AlchemicaKrea (max difference 0) | | NAG forward rewrite | 0.0 difference from ComfyUI's Krea 2 forward when NAG is off (offline test on a miniature Krea 2 DiT) |

All numbers come from same-seed tests where only one factor changes, with 100 % crops of the same area. The full lab notes are in README_IT.md.

⚗ Requirements

• ComfyUI 0.36.0 or newer (Krea 2 support and CLIP.generate, used by the prompt enhancer).
• Krea 2 Turbo (or Raw) weights, e.g. krea2_turbo_nvfp4.safetensors or krea2_turbo_fp8_scaled.safetensors, in models/diffusion_models.
• Text encoder Qwen3-VL-4B for Krea 2 (qwen3vl_4b_fp8_scaled.safetensors) in models/text_encoders.
• VAE qwen_image_vae.safetensors in models/vae.
• NVFP4 on RTX 50-series (Blackwell) runs at full speed only with PyTorch built for CUDA 13.0 (cu130); the calibration node reports your torch/CUDA build and warns you if it is not right.
• VRAM: 16 GB is enough for 4 MP with NVFP4 (tiled VAE decode is used in the workflows).

⚗ Installation

ComfyUI-Manager: search for AlchemicaSampler (once it is published in the registry).

git:

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

zip: extract the release so that you get ComfyUI/custom_nodes/ComfyUI-AlchemicaSampler/__init__.py.

Then restart the ComfyUI server (reloading the browser page is not enough). The nodes appear under Add Node → ⚗ AlchemicaMente → sampling, and the example workflows under Workflow → Browse templates → ComfyUI-AlchemicaSampler.

⚗ Quick start

1. Calibrate once per checkpoint. Open KREA2_TARATURA (3 MP) or KREA2_TARATURA_4MP, pick your checkpoint in the first node, press Queue. Four prompts are measured and averaged into one profile (≈15–20 min at 4 MP). Press R in the browser to refresh the menus.
2. Generate. Open KREA2_COMPLETO, write an idea in the enhancer (any language, a few words are fine), press Queue. The first run takes longer because the text encoder writes the prompts; after that they stay cached until you change the idea.
3. Remove what you don't want. Edit the negative text or leave the enhancer's one; NAG 3 from -1 is already set.

No calibration yet? Everything still works: the nodes fall back to an analytic schedule and default values, and the report says so.

⚗ Example workflows

| Workflow | Use it for | |---|---| | KREA2_COMPLETO | everything together: loader → prompt enhancer → text → attention guidance (NAG) → AlchemicaKrea Pro | | KREA2_SEMPLICE | the essential: LoRA stack + AlchemicaKrea, one node to sample | | KREA2_PRO | two-phase prompt, previews and refinement variants, prompts written by hand | | KREA2_TARATURA / _4MP | calibration of your checkpoint at 3 MP / 4 MP | | KREA2_ALCHEMICA | the advanced sampler with all its option nodes | | KREA2_ALCHEMICA_CONFRONTO | same-seed comparison: ComfyUI simple vs Alchemica curves vs calibrated | | KREA2_IBRIDO | compose with one model, refine with another, each with its own LoRA stack | | KREA2_RISOLUZIONE | resolution ladder 2.5 / 3 / 4 / 5 MP with the same seed |

The workflows use only core ComfyUI nodes plus AlchemicaSampler.

⚗ The nodes

Node parameters keep their original Italian names (so saved workflows keep working); every parameter has an English tooltip. The glossary at the end of this section translates them.

⚗ Alchemica · Load Krea 2

Model + text encoder + VAE in one node. The text encoder is always loaded as type krea2: with any other type Krea 2 fails with a conditioning-size error.

⚗ Alchemica · Prompt enhancer (text encoder)

Krea 2's own text encoder (Qwen3-VL-4B) writes the prompt, so no Ollama or external server is needed, and because it is a vision-language model it can start from a reference image (empty idea + image = it describes the image). It returns three texts:

• SCENE: subject, pose, framing, setting, light, colours and style. Colours and style must be here: our measurements show Krea 2 Turbo decides them in the very first step, together with the composition.
• DETAIL: rendering only (skin, fabric, material, optics). It takes over below the hand-off sigma, where it can change texture but not the picture.
• NEGATIVE: things to avoid, for NAG.

Built-in rules: no words that paint water drops on skin, no "imperfections" (they turn into bumps), no negations in the positive text (at cfg 1 a negation brings in the very thing it negates), consistent optics. A filter removes any sentence that breaks a rule, and the report lists what was removed. creativita 0 = deterministic; 0.6–0.9 with different seeds = variations. stile presets: foto realistica (natural photo), ritratto (portrait), still life / laboratorio, paesaggio (landscape), architettura, cinematografico (cinematic), illustrazione, libero (free).

⚗ Alchemica · Text (scene / detail / negative)

Encodes the three texts. The detail conditioning is scene + detail by default (aggiungi alla scena), so the model never forgets the subject. The negative is only used by NAG: at cfg 1 it has no effect anywhere else.

⚗ Alchemica · Attention guidance (NAG / SEG / PAG)

Patches the model; connect its output to any Alchemica sampler.

• NAG — Normalized Attention Guidance. In every block of Krea 2's single-stream DiT the image tokens attend twice, once with the positive text and once with the negative text; the result is extrapolated away from the negative, its norm is limited (τ) and blended (α). The negative text gets its own small stream, so the cost is one extra attention per block, not a second model pass.
  • nag_scala 3 (validated), nag_tau 2.5, nag_alpha 0.25.
  • nag_sigma_da −1 = start at the calibrated hand-off sigma (~0.87): removes the negative without changing the composition. 1.0 = from the first step: stronger, but the picture changes.
• PAG — Perturbed-Attention Guidance: a second prediction with identity attention in blocks 8–19. With scale 1 from −1 the picture stays the same and the rendering becomes crisper and more contrasty (and loses a few freckles): a style choice rather than an absolute improvement.
• SEG — Smoothed Energy Guidance: blurred image queries. Keep it at 0.3 or below; at 1 it produces halos.
• PAG and SEG cost one extra model call per step inside their sigma window.

⚗ AlchemicaKrea Pro (phases + previews)

The daily sampler. Everything AlchemicaKrea does, plus:

• Two-phase prompt: connect positive (scene) and positive_dettaglio (detail). The detail takes over below cambio_fase (−1 = the calibrated hand-off). sfumatura > 0 cross-fades the two (extra cost inside the band).
• Composition previews: anteprime = N produces N low-resolution images with seeds seed, seed+1, …; stage 1 is exactly the one of the final image and stage 2 is redone small with the same noise field. Keep the seed on fixed, pick a number, set anteprime = 0 and scelta = that number, queue again.
• Refinement variant: variante_rifinitura changes only the second-stage noise: same pose and framing, different face, lettering, jewellery and small objects.
• Profile by resolution: (auto) chooses, among the profiles of this checkpoint, the one calibrated at the closest resolution.

⚗ AlchemicaKrea (all-in-one)

One node with the knobs that matter: modo (veloce / equilibrato / massimo dettaglio), due_stadi, ridisegno, dettaglio, varieta; everything else comes from the calibration profile. Optional modello_rifinitura refines stage 2 with a different model.

⚗ Alchemica · Calibration

Runs a dense euler trajectory (40–60 steps) and records the model's prediction at every sigma. It measures:

• the cost ‖dD/dσ‖/σ → step density n(σ) ∝ √(‖dD/dσ‖/σ), the distribution that minimises euler's total error with a fixed number of steps;
• composition lock-in (low-frequency similarity to the final image) → hand-off sigma;
• birth of detail structure and sharpness (high-frequency similarity and energy) → detail window and restart sigma;
• optionally, the NVFP4 error against a reference model (e.g. FP8) on the very same states.

Profiles accumulate (re-running the same prompt updates its sample), carry a checkpoint fingerprint and the resolutions they were measured at. Checkpoints with exotic quantisations may only give an "architecture" fingerprint: the report warns you; choose the profile by hand in that case.

Other nodes

⚗ AlchemicaSampler (advanced) with option nodes (Schedule, Solver & Detail, Restart, Guidance, Two stages), ⚗ Alchemica · Sigmas and Sampler (SAMPLER) for SamplerCustom, ⚗ Alchemica · Resolution (Krea 2 latent from aspect + megapixels, sides multiple of 16), ⚗ Alchemica · LoRA (stack of 6), ⚗ Alchemica · Two-phase prompt, ⚗ Alchemica · Number previews.

Glossary of parameter names

| Italian | English | Italian | English | |---|---|---|---| | modo | mode (veloce = fast, equilibrato = balanced, massimo dettaglio = max detail) | profilo | profile | | due_stadi | two stages | ridisegno | redraw (hand-off sigma) | | dettaglio | detail | varieta | variety (gated ancestral noise) | | cambio_fase | phase switch sigma | sfumatura | cross-fade width | | anteprime | previews | scelta | choice | | variante_rifinitura | refinement variant | modello_rifinitura | refinement model | | idea / stile / parole | idea / style / words | creativita / immagine | creativity / image | | scena / negativo / unione | scene / negative / merge mode | struttura | structure guidance | | nag_sigma_da / _a | NAG from sigma / to sigma | seg_sfocatura | SEG blur sigma | | passi_densi | dense steps | modalita (accumula / sostituisci) | mode (accumulate / replace) | | modello_riferimento | reference model | pesi / forza_clip | weights / clip strength |

⚗ What the measurements taught us

• Krea 2 Turbo decides composition and palette in the first step. 98.5 % of the composition is locked at σ ≈ 0.87; sharpness grows between σ 0.57 and 0.13. A late prompt can change texture, not colours or style.
• Detail boost must stay low. 0.40 covers skin with coloured speckles; 0.12 raises skin micro-relief by 24 % with the same chroma noise as a clean reference.
• 4 MP is the sweet spot. Detail per face area keeps growing up to 5 MP, but at 5 MP flat skin shows a repeating mesh texture; 5 MP only for close-ups.
• The 3 MP and 4 MP profiles are almost identical (hand-off 0.87 vs 0.864): one calibration covers both.
• Noise downscaling needs care: averaging with overlapping windows correlates neighbouring noise values by 33 % and the model leaves coloured blobs; exact 2×2 blocks keep it white.

⚗ Troubleshooting

• New nodes or parameters don't show up → restart the ComfyUI server; a browser reload is not enough.
• "NESSUN profilo per questo modello" (no profile for this model) → run the calibration with this checkpoint.
• Krea 2 conditioning-size error → the text encoder was loaded with a type other than krea2; use ⚗ Load Krea 2.
• NVFP4 slower than FP8 → PyTorch is not a cu130 build; see the calibration report.
• First run of KREA2_COMPLETO is slow → the text encoder writes the prompts and the model is reloaded afterwards; later runs reuse the cached text.
• Out of memory at 5 MP → use 4 MP or modo veloce.
• Runtime reports are currently written in Italian; the glossary above covers their key words.

⚗ Extending

# e.g. at the end of alchemica/schedules.py
from .registry import register_scheduler

@register_scheduler("my_schedule")
def my_schedule(model_sampling, steps, cfg):
    import torch
    return torch.linspace(1, 0, steps + 1)

Solvers: @register_solver("name") in alchemica/solvers.py. Presets: one line in alchemica/presets.py. Plunge, restart, two stages and denoise work with any new scheduler.

⚗ License and models

Code: GPL-3.0 (LICENSE), the same license as ComfyUI, from which the NAG forward is adapted. Parts adapted from MIT-licensed projects are listed with their licenses in THIRD_PARTY_NOTICES.md.

Models are not included and are not redistributed. Download them from their official sources and follow their licenses: Krea 2 (Raw / Turbo) — Krea 2 Community License (commercial use only below USD 1M annual revenue, plus acceptable-use and content-filtering obligations); Qwen3-VL-4B — Apache-2.0; Qwen-Image VAE — see its model page. Calibration profiles are not included: the calibration node creates them on your machine.

⚗ Credits and thanks

⚗ AlchemicaMente had the idea, directed the project, ran every test and made every call. Claude (Anthropic) wrote the code, the measurements and the documentation in sessions guided by AlchemicaMente.

Thank you to ComfyUI and its contributors, Krea, the Qwen team, Kostiantyn Hrytsuk (KreaPhoton, CyberKrea), Auryg, the authors of NAG, PAG and SEG, pamparamm, pythongosssss and the whole ComfyUI and Civitai community. The full list is in CREDITS.md.

Thank you to everyone who will use AlchemicaSampler, to everyone who will help us, and to the whole community. And if this idea brings even a little progress, we will be happy. ⚗