ComfyUI Node
⚗ Alchemica KSampler Flex (any sampler + scheduler)
The Alchemica KSampler with a free choice of sampler and scheduler: the calibrated profile, the Alchemica curves or any sampler/scheduler installed in ComfyUI (euler_flow, res_multistep, beta, bong_tangent ...). Two stages, previews and the image in the node as in the KSampler.
⚗ Alchemica KSampler Flex (any sampler + scheduler)
- model
- positive
- latent_image
- negative
- vae
- refiner_model
- opts
- latent
- sigma_plot
- report
- image
◄seed0►
◄sampler_namealchemica_euler_2m►
◄scheduleralchemica_calibrated►
◄steps8►
◄cfg1.00►
◄profile▾►
◄denoise1.00►
◄two_stagetrue►
◄stage1_steps4►
◄redraw-1.00►
◄detail0.12►
◄variety0.60►
◄restart_steps2►
◄contraction0.85►
◄shift0.00►
◄handoff_k0.00►
◄calibration▾►
◄lowfreq_anchor0.00►
◄preview_methodlatent2rgb (fast)►
◄degrid2.0►
Category⚗ AlchemicaMente/sampling
Inputs (27)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| positive | CONDITIONING | — | |
| latent_image | LATENT | At the final resolution. In two-stage mode the first stage works at half of it | |
| seed | INT | 00–18446744073709550000 | — |
| sampler_name | COMBO | alchemica_euler_2m | alchemica_* = the Alchemica loop: 'detail', 'variety' and 'restart' work. Any other = that ComfyUI sampler as it is (those three are ignored) |
| scheduler | COMBO | alchemica_calibrated | alchemica_calibrated = the step spacing measured on YOUR checkpoint (needs a profile; without one it uses alchemica_dual). The others are fixed curves: Alchemica ones first, then every ComfyUI scheduler |
| steps | INT | 81–100 | Single stage: all the steps. Two stages: the steps of stage 2 (full resolution) |
| cfg | FLOAT | 1.000–20 | 1.0 = Krea 2 Turbo (negative skipped, no extra cost). Above 1 the 'negative' input is used, if connected |
| profile | COMBO | (auto) recognises the checkpoint and uses its calibration profile. It feeds alchemica_calibrated and every -1 value, with ANY sampler or scheduler | |
| denoise | FLOAT | 1.000–1 | 1.0 = from scratch. Below 1 = img2img (two stages off) |
| two_stage | BOOLEAN | true | Composes at half resolution, refines at full resolution. Recommended above ~1.5 MP |
| stage1_steps | INT | 41–50 | Two stages: steps at half resolution (composition) |
| redraw | FLOAT | -1.00-1–0.95 | Two stages: sigma where stage 2 takes over. -1 = from the profile. Higher = more new detail, but the composition may change |
| detail | FLOAT | 0.120–0.25 | alchemica_* only. Micro-relief of skin and fabric; above 0.2 coloured speckles come back |
| variety | FLOAT | 0.600–1.2 | alchemica_* only. Ancestral noise during composition. 0 = deterministic. With strong LoRAs lower it to 0.15-0.2 if you see wet-skin droplets |
| restart_steps | INT | 20–10 | alchemica_* only. Extra steps that re-noise and redo the end (sigma from the profile). 0 = off |
| contraction | FLOAT | 0.850.3–1.2 | Scale of the initial noise. 1 = standard ComfyUI noise; 0.85 = cleaner shadows |
| shift | FLOAT | 0.000–10 | Flow shift of the schedule. 0 = the model's own. Higher = more steps at high noise (composition), lower = more at the end (detail) |
| handoff_k | FLOAT | 0.000–4 | Two stages: resolution correction of the hand-off. 0 = auto: stage 1 stops at the half-resolution sigma equivalent to 'redraw' at full resolution (SD3 rule), so stage 2 gets a composition as settled as the profile measured. 1 = off (same as ⚗ Alchemica KSampler). >1 = explicit factor |
| calibration | COMBO | How alchemica_calibrated turns the profile into steps. v2: step law matched to the solver order (cube root of the 3rd derivative for 2nd-order solvers), smoothing in log-SNR, own curves for stage 1. Profiles calibrated before 2026-10-10 get only the smoothing: re-run the Calibration to measure the rest | |
| lowfreq_anchor | FLOAT | 0.000–1 | Two stages, alchemica_* samplers: stage 2 keeps the LOW frequencies (composition, colour) of the upscaled stage-1 image, fading out as sigma falls (HiFlow, NeurIPS 2025). Protects the composition while stage 2 adds detail. EXPERIMENTAL: 0 = off (default, same as ⚗ Alchemica KSampler), 1 = full |
| negativeopt | CONDITIONING | Used only with cfg above 1 | |
| vaeopt | VAE | Optional: the final image is decoded and SHOWN in this node, and returned on the 'image' output (like the Efficiency KSampler) | |
| preview_methodopt | COMBO | latent2rgb (fast) | Live preview while sampling. latent2rgb = instant, a bit soft; taesd = sharper (needs the TAE model in models/vae_approx, otherwise latent2rgb); 'ComfyUI setting' = whatever ComfyUI is set to (often none) |
| degridopt | FLOAT | 2.00–10 | Only with a VAE connected: removes the faint 2-pixel line grid of Krea 2's VAE from the image shown and returned here (the latent is untouched). Max correction in grey levels: 2 = recommended, 0 = off |
| refiner_modelopt | MODEL | Optional: stage 2 with another model (requires two stages) | |
| optsopt | ALCHEMICA_OPTS | For tinkerers: Alchemica option nodes override everything set here |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | — |
| sigma_plot | IMAGE | — |
| report | STRING | — |
| image | IMAGE | — |