K2 Advanced Sampler

If you've tried to run Krea 2 Turbo in ComfyUI you already know the annoying part: stock KSampler gives you a generic grid, and Turbo is a distilled flow-matching model that wants its own sigma schedule, its own shift, and about eight steps. K2 Advanced Sampler is the all-in-one fix from the ComfyUI-Krea2-FlowLab pack - one node that builds the model-native grid, runs a custom flow solver over it, and spits out a latent you can decode. Same shape as a normal KSampler, zero scheduler hunting.

It's an experimental inference-time solver, not anything Krea ships. The author (a solo dev, and this pack is brand-new enough that nobody's really beaten on it in public yet) is upfront that it's a math experiment layered on top of ComfyUI's sampling contract. Treat it as something to A/B against your current setup, not a guaranteed upgrade.

How it works

The node internally does what the pack's other two nodes do separately: it builds the Krea-2 sigma grid from the model's own model_sampling configuration (the native schedule reproduces the shift baked into Turbo, 1.15), then hands the latent to a custom sampler that makes one model evaluation per step. That solver is the whole pitch: a variable-step Adams-Bashforth-2 correction on top of Euler, gated by a curvature check (if the velocity field turns sharply between steps, it falls back toward plain Euler), with global and local trust-region limits on how much the multistep correction is allowed to push. There's an eta knob too, but it only injects noise in the middle of the trajectory - controlled by eta_start/eta_end - so the start and end stay deterministic.

The inputs that matter

For Krea 2 Turbo, the author's recommended first test is simple:

• steps 8, cfg 1.0, denoise 1.0 - Turbo is guidance-distilled, so cfg stays at 1
• preset balanced - the default; euler is your debug baseline (it disables the correction and gives you plain one-call Euler on the native grid)
• schedule_profile native - structure pushes more steps early, detail late, balanced is a mild midpoint compromise
• eta 0.0 first, then try 0.15–0.35 for the stochastic middle window

The four solver widgets (correction_strength, curvature_threshold, trust_ratio, local_trust) are ignored unless you set preset to custom - each preset carries its own values, so don't twist them and wonder why nothing changes. The optional full_profile dropdown is the headline feature: leave it on 00 Manual widgets to use the visible controls, or pick one of the 21 named profiles (tiered SAFE / NORMAL / CREATIVE / EXTREME) and it overrides preset, schedule and stochastic settings together. 01 MY BEST - balanced native eta 0.35 is the author's saved winner. The EXTREME ones are deliberate stress tests; expect artifacts.

Outputs

• output (LATENT) - the finished latent, straight into a VAE Decode
• denoised (LATENT) - the last full denoise estimate; handy for comparing, especially on partial denoise
• options (OPTIONS) - a debug object showing exactly which sampler/scheduler/profile actually ran, useful when a full_profile is overriding you

Installing it

No pip install and no model downloads - the node uses only PyTorch plus modules ComfyUI already ships, and it assumes you already have the Krea 2 files loaded through ComfyUI's native loaders (UNETLoader + CLIPLoader type: krea2 + the Qwen-Image VAE). You still need all three of those.

cd ComfyUI/custom_nodes
git clone https://github.com/ekkonwork/ComfyUI-Krea2-FlowLab.git

Restart ComfyUI; you should see [K2 FlowLab] loaded in the console. From the ComfyUI root you can run python custom_nodes/ComfyUI-Krea2-FlowLab/verify_install.py to confirm all three nodes registered.

Common issues

• Nodes don't appear - folder must be exactly ComfyUI/custom_nodes/ComfyUI-Krea2-FlowLab, then a full restart. Check the console for an IMPORT FAILED line.
• Result looks worse than your old sampler - that's expected and normal. Turbo is distilled for a short trajectory, so stronger correction isn't automatically better. Go native + conservative + eta 0 and keep the original workflow for A/B.
• No speedup - correct. It's still one DiT call per step; the correction happens tensor-side. The pitch is quality at 8 steps, not fewer calls.