⚗ AlchemicaSampler for Krea 2 Turbo
A sampler that measures your checkpoint, then samples the way the model actually works.
⚗ Idea and direction: AlchemicaMente · code: Claude (Anthropic) · license: GPL-3.0
Custom nodes and full documentation: github.com/Nessuno76/ComfyUI-AlchemicaSampler
What this workflow does
Prompt enhancer built on Krea 2's own text encoder. Write an idea in any language (or connect a reference image). Qwen3-VL-4B, the very encoder Krea 2 uses, writes a SCENE, a DETAIL and a NEGATIVE prompt. No Ollama, no external server.
NAG, a negative prompt that works at cfg 1. Krea 2 Turbo is distilled, so the usual negative prompt does nothing. NAG applies the negative inside the attention of every DiT block. Starting at the calibrated hand-off sigma, it removes what you don't want without changing the picture (measured: 99.3 % identical image, unwanted water drops gone).
Two-phase prompt. The scene leads until the composition locks (σ ≈ 0.87, measured); the detail prompt leads after that. Same picture, different rendering of skin, fabric and materials.
Calibrated sampling. A calibration run measures your checkpoint on your GPU and builds the step schedule, the two-stage hand-off and the detail window from those measurements.
Composition previews. Generate 4–8 small previews, pick one, render it at 4 MP with the same composition.
Requirements
ComfyUI 0.36.0+
Custom nodes: ComfyUI-AlchemicaSampler (GitHub link above, or search "AlchemicaSampler" in ComfyUI-Manager)
krea2_turbo_nvfp4.safetensorsorkrea2_turbo_fp8_scaled.safetensors→models/diffusion_modelsqwen3vl_4b_fp8_scaled.safetensors→models/text_encodersqwen_image_vae.safetensors→models/vae16 GB VRAM for 4 MP (NVFP4). On RTX 50-series use a PyTorch build for CUDA 13.0 (cu130) for full NVFP4 speed.
How to use
Install the custom nodes and restart ComfyUI.
Run KREA2_TARATURA once with your checkpoint (≈15–20 min at 4 MP). It is optional but recommended: without it the nodes fall back to default values.
Open KREA2_COMPLETO, type your idea in the enhancer, press Queue.
To remove something from the image, write it in the negative (or keep the enhancer's). NAG 3 from −1 is preset.
To explore compositions: set
anteprime(previews) = 8 with the seed on fixed, pick a number, setanteprime= 0 andscelta(choice) = that number.
Recommended settings (validated)
Resolution: 4 MP (5 MP only for close-ups)
modo (mode): equilibrato (balanced)
dettaglio (detail boost): 0.12
varieta (variety): 0.6
NAG: scale 3 · tau 2.5 · alpha 0.25 · from −1
Structure guidance: off (PAG 0.5–1 from −1 for a crisper, more contrasty look)
Timing on an RTX 5070 Ti at 4 MP: ~63 s without NAG, ~85–95 s with NAG. The first run of the complete workflow takes longer because the text encoder writes the prompts; they stay cached until you change the idea.
Honest notes
The detail prompt changes texture, not colours or style: Krea 2 Turbo decides those in the first step.
PAG is a style (crisper, fewer freckles), not an absolute improvement. SEG above 0.3 makes halos.
NAG and the enhancer are written for Krea 2 only.
License
Workflows and nodes: GPL-3.0. Models are not included: Krea 2 is under the Krea 2 Community License (commercial use only below USD 1M annual revenue; acceptable-use and content-filtering obligations). Qwen3-VL-4B: Apache-2.0.
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.
Thank you to everyone who will use this, to everyone who will help us, and to the whole community. And if this idea brings even a little progress, we will be happy. ⚗ — AlchemicaMente
Description
ALCHEMICA_T2I: Load Krea 2 → LoRA Stack+ → Prompt → KSampler Flex → Save. Write an idea in any language and Krea 2's own text encoder writes the prompt. Framing and look presets with thumbnails sit at the bottom of the Prompt node: shot, camera angle, view, light, atmosphere and background.
ALCHEMICA_I2I: load a photo, press Queue, then drag the before / after slider. The front of the workflow is simple (photo, what to do, prompt). The groups hold the rest: a mask to change only one zone, LoRA Stack+, Black and White, and a SeedVR2 upscale group that stays bypassed until you enable it.
Different, not better. On the same seed the Alchemica recipe gives another picture than ComfyUI's KSampler, not a guaranteed improvement. The same KSampler Flex node also does standard sampling (euler + simple, two_stage off), so you keep the tools and choose the look.
Install from ComfyUI-Manager. 1.1.4 is the first 1.1 version that passed the Comfy Registry scan. The nodes only talk to ComfyUI: no online services and no Ollama.
New since 1.0
⚗ LoRA Stack+: as many LoRAs as you need, real strengths and previews. Trigger words are added to the prompt automatically, and ▲ ▼ try the previous or next LoRA.
⚗ LoRA A/B: the same photo with and without your LoRAs in one run, shown in the before / after slider.
⚗ Black and White: seven darkroom controls (auto levels, shadows, highlights, clarity, sharpness, vignette).
⚗ Load Image Folder: one photo of a folder per run, to process a whole folder.
⚗ Prompt: framing and look presets with thumbnails, opening phrases, compact view.
Fixes
Two or more people in a wide frame: the second person is no longer broken into fragments.
Orientation written in the text: "Vertical 9:16 frame" now wins over aspect, so a vertical scene is no longer squeezed into 16:9.
Framing filter: it removes only what the frame hides. In a chest-up shot, "a red jacket and jeans" now keeps the jacket.
LLM negative: a looping negative is cut where the loop starts, and the negative no longer works against text written in quotes.
Calibration: it takes an optional negative, for models used at a real cfg.
Requirements
ComfyUI 0.36+ and the custom nodes ComfyUI-AlchemicaSampler 1.1.4 (from ComfyUI-Manager or https://github.com/Nessuno76/ComfyUI-AlchemicaSampler).
Models:
krea2_turbo_nvfp4.safetensors (or fp8) → models/diffusion_models
qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders
qwen_image_vae.safetensors → models/vae
16 GB of VRAM for 4 MP. On RTX 50-series cards, use PyTorch for CUDA 13.0 (cu130) for full NVFP4 speed.
Full changelog: https://github.com/Nessuno76/ComfyUI-AlchemicaSampler/blob/main/CHANGELOG.md
Using MOSS-Image? It has its own pack, ⚗ AlchemicMoss: https://github.com/Nessuno76/ComfyUI-AlchemicMoss




