⚗ Alchemica Sampler (advanced, presets + options)

Krea 2 Turbo at 8 steps and CFG 1 is already good. This sampler is for the part after "already good": step spacing measured on your checkpoint, two stages, and a report that admits what it changed.

What it is

AlchemicaSamplerAdvanced is the modular sampler in ComfyUI-AlchemicaSampler - the sibling of the pack's all-in-one KSampler, with the options broken out into a chain of little option nodes and the latent handed back to you instead of swallowed. Presets do the heavy lifting; the option nodes override individual entries; the report tells you which preset, scheduler and solver actually ran and whether a calibration profile was found.

The premise of the pack, and it's worth repeating because the author goes out of his way to temper it: this is not a miracle. Krea 2 Turbo at 8 steps is good out of the box. What you get here is the last bit of detail and a lot of control - and, in the pack's own honest A/B/C comparison, a different composition if you turn on everything. Calibrated step spacing alone came out marginally crisper with a bit more contrast and skin texture, and occasionally produced a smooth blocky patch. Plan accordingly.

The presets

preset is a list of ten. The useful ones:

• reference_euler_simple - ComfyUI's own simple scheduler, no calibration, for comparison runs.
• alchemica_turbo - the default. Detail boost 0.12, ancestral noise 0.5, contraction 0.85 on the analytic schedule. No profile needed.
• alchemica_quality - 10 steps, a second-order solver, restart steps.
• alchemica_hires_2stage - two stages, first at half resolution.
• photo_calibrated / photo_calibrated_2K - read the calibration profile of your checkpoint, including the measured detail window and hand-off sigma. These need a profile; that's what the Calibration node is for.
• auto_tarato - the interesting one. Detail boost and ancestral noise are set to −1, meaning "read them from the profile": the pack measures how much ancestral noise and detail boost this checkpoint tolerates and uses those numbers instead of ones tuned on somebody else's model. Any calibrated checkpoint can use it, merges included.
• raw_base - Krea 2 Raw at CFG 3.5 with a text negative, 28 steps.
• plunge_restart - a high-detail recipe with a plunge sigma and restarts.

profile defaults to (auto), which recognises the checkpoint by itself once you've calibrated once. No profile? It falls back to the analytic schedule and the report says so - nothing breaks.

The inputs and outputs you actually touch

model, positive, latent_image (at the final resolution - the first stage works at a fraction of it), seed, preset, profile, denoise. denoise at 1.0 is from scratch; below 1 it's img2img and it turns two-stage mode off. Optional: negative (only used with a text-negative mode and cfg above 1), refiner_model (run the second stage with a different model), opts (chain an Alchemica Options node to override preset entries), vae, preview_method (latent2rgb is instant and a bit soft; taesd is sharper but needs a TAE model in models/vae_approx; otherwise you inherit ComfyUI's setting), and degrid.

degrid deserves its own line. Krea 2's VAE - the Qwen-Image one - leaves a faint 2-pixel line grid in flat areas. The pack measured it at 2–4× the neighbouring frequency, roughly half a grey level. With a VAE connected, degrid (2 is the recommended value) strips it from the image shown in the node and returned on image. The latent is never touched.

Out: latent, sigmas, sigma_plot (a picture of the schedule - genuinely useful when a calibration profile changes the spacing), report, and image. Connect a VAE and the node decodes and shows the result itself, Efficiency-KSampler style.

Installing it

ComfyUI Manager, search "AlchemicaSampler", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler
# restart the server, not just the browser

ComfyUI 0.36.0+ and no extra Python packages. For NVFP4 checkpoints on a 50-series card, the pack is explicit that you only get full speed with a PyTorch built for CUDA 13.0 (cu130) - if NVFP4 feels slower than fp8, that's the reason, and the Calibration node reports your build. 16 GB of VRAM is enough at 4 MP.

Where it goes wrong

• "NO profile for this model". Run the calibration workflow once for that checkpoint. Fifteen to twenty minutes at 4 MP, and it's per checkpoint (and per LoRA stack - the profile records a signature).
• Two-stage is doing nothing. Check denoise is 1.0. Anything below it means img2img, and the second stage is skipped by design.
• (auto) picked a profile made without your LoRAs. It prefers a profile with the same stack; when there isn't one, the report tells you the profile was made without them.
• New widgets don't appear. Restart the server; a browser reload isn't enough. The pack prints a warning listing unrecognised inputs when it sees a stale graph.
• You expected the plain KSampler's image. Don't. Two stages, detail boost and restarts will move the composition. Keep the seed if you want the same new picture again.