⚗ Alchemica KSampler (Krea 2, all-in-one)

Here's the thing about this pack most sampler nodes don't do: the authors ran their own A/B/C test and published the result, including the part where their recipe loses. Plain ComfyUI KSampler vs calibrated schedule only vs the full two-stage recipe: branch B is "closest to the original and not always an improvement", branch C gives the most detail but a different picture every time, and if A already looks right to you, keep A.

So reach for this node when you want control and the last bit of detail, not because stock Krea 2 at 8 steps is bad. It isn't.

What it is

One node with the six knobs that matter. Everything else - step spacing, the hand-off between stages, the detail window, the restart sigma - comes from your calibration profile, or from an analytic fallback if you haven't run one. The report always tells you which of the two you got, which is the single most useful line it prints.

The knobs

• model, positive, latent_image - the latent is not created here, feed it from ⚗ Resolution (or any Empty Latent) at your final resolution.
• profile - (auto) recognises the checkpoint and picks the closest-resolution profile. Leave it on auto.
• mode - fast (~30 s), balanced (~45 s), max detail (~70 s), at 3 MP. This is the step budget; steps overrides it if you insist.
• two_stage - on by default, and effectively required above ~1.5 MP. Stage 1 composes at half resolution, stage 2 refines at full.
• redraw - how freely stage 2 may redraw. -1 = from the profile, and that's the recommendation. Crank it and you get more new detail and a possibly different composition.
• detail - micro-relief on skin and fabric. 0.12 is the validated value. Above 0.2 you get coloured speckles; 0.40 covers skin in them.
• variety - ancestral noise, i.e. texture and seed-to-seed variety. 0 = deterministic. Read the tooltip on this one: on checkpoint + strong-LoRA combinations, above ~0.3 it can produce wet-skin droplet artifacts regardless of where the noise is gated. Confirmed clean at 0.15 with a strong skin LoRA at x2. If your image looks like it's been rained on, drop variety to 0.15–0.2 first.
• denoise - 1.0 is from scratch. Below 1 turns it into img2img and disables two-stage mode.
• contraction - scale of the initial noise. Leave at 0.85. It is not the cause of the droplets (the authors tested it at 1.0), but at low variety the 0.85 default is measurably cleaner than 1.0.
• cfg - -1 means 1.0, the Turbo default, where no negative is computed and costs nothing.

Optional inputs worth knowing: negative (only used if you push cfg above 1), vae (decodes and shows the final image in the node, Efficiency-KSampler style - the node is an output node when you use it), refiner_model (stage 2 with a different checkpoint: compose with Krea 2, refine with a fine-tune), opts (the option-node chain), preview_method (latent2rgb is instant, taesd is sharper if the TAE model is in models/vae_approx), and degrid - 0 here by default, and only active with a VAE connected.

Outputs: latent, sigma_plot (an image of the schedule it actually used - save a few, they explain what the calibration bought you), report, image.

Where it fits

Load Krea 2 → LoRA Stack → Prompt → Alchemica KSampler → Save Image

That's the pack's KreaAlchemicaMente_K2AM template, and it's the right place to start. Model, text encoder and VAE travel through the chain instead of being re-wired from the loader to five nodes.

Installing

# ComfyUI Manager: search "AlchemicaSampler"
cd ComfyUI/custom_nodes && git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

ComfyUI 0.36+, no pip dependencies, restart the server. You'll also want the three model files (krea2_turbo_nvfp4.safetensors or ..._fp8_scaled, qwen3vl_4b_fp8_scaled.safetensors, qwen_image_vae.safetensors) and ideally one calibration run.

Common issues

"NO profile for this model." Expected on a fresh install. It still generates; the report says it used the analytic schedule. Run the Calibration once per checkpoint and this goes away.

Out of memory at 5 MP. The pack's own testing says 4 MP is the sweet spot and 5 MP is close-ups only. Drop to 4 MP or fast.

The picture changed when you only touched detail or variety. Those feed ancestral noise, so seeds move. If you want the same image with more bite, change the calibrated schedule branch, not the noise.

Headless image output. If you don't connect a vae, there's no image output - decode with a VAE Decode node as usual and put DeGrid after it.