CyberKrea Sampler

Krea 2 Turbo is a distilled, CFG-1 model, and the community's consensus setup isn't "grab a KSampler and go" - it's a specific recipe: an euler-family sampler, the mu 1.15 flux-shift schedule, no real CFG. Want more than the bare minimum and you're stacking tricks - a restart pass for detail, initial noise tuned to real photographs, ancestral noise that doesn't smear shadows into blotches. CyberKrea Sampler bakes all of that into one node behind three presets. Nothing here calls Krea's API; it's all local math on the open weights.

It ships in the small cyberdeliaAI/ComfyUI-CyberKrea-Sampler pack (MIT, zero extra Python deps), and its sampling engine is derived from ComfyUI-KreaPhoton. Think of it as the "lite" edition: it keeps the calibrated parts and deliberately drops the experimental ones.

How it works

The node runs one uninterrupted custom sampling loop on a single patched model - that's a deliberate design choice. Because there's no clean-model split, your LoRAs and NegPiP patches stay active for the whole trajectory instead of getting reset partway and producing bf16/ancestral artifacts. The loop itself layers four mechanisms:

• Plunge + restart. The main pass locks the composition with a direct end step at a high-sigma floor, then a restart_frac of your step budget is reallocated (not added) to jump back up to sigma_r, re-noise, and descend again. Composition is settled; the restart only works on texture and detail.
• Detail sigma-nudge. A detail amount adjusts the effective sigma per step to emphasize fine detail; crank it too far and things look harsh.
• Gated-eta ancestral noise. eta0 adds stochastic ancestral noise in the mid-sigma band, but it's gated off below sigma_gate (default 0.10) - terminal ancestral injection is the classic dark-blotch driver.
• Manifold contraction. Stock unit noise is roughly 2.1× wider than the measured photo manifold, so contraction scales the initial noise toward it (1.0 = stock, 0.70 = preset). This is the "why does everything look synthetic" knob.

The schedule is analytic, built with Krea 2 Turbo's flux time-shift (alpha = e^1.15). Negative handling is worth calling out: leave negative disconnected and the node runs at CFG 1.0 with an honest zeroed unconditional - the distilled-model way. Connect one and it switches to sigma-window guidance, where CFG scale ramps up only in a mid-sigma window rather than everywhere. This pack was written around NegPiP, which packs the negative into the positive conditioning - so if NegPiP is handling negatives, keep negative empty.

The three presets

Picking a preset fills every control below it; you can then override any single value without losing the rest.

| Preset | Steps | Detail | Sampler | |---|---|---|---| | fast | 8 | 0.50 | euler | | balanced | 12 | 0.60 | euler | | quality | 16 | 0.70 | euler_2m |

balanced is the default and is the sensible starting point. Fast is for iterating, quality for the final render - euler_2m uses a two-step history (Adams-Bashforth 2) which is why it's reserved for the highest step count.

The inputs that matter

• model - the final Krea 2 model, connected after your LoRA and/or NegPiP chain.
• positive - your conditioning.
• latent_image - a 16-channel Krea 2 latent, e.g. from the pack's CyberKrea Empty Latent.
• seed - controls initial noise and every seeded noise stream for reproducibility.
• negative (optional) - leave disconnected when NegPiP already handles negatives.
• vae (optional) - only decodes a thumbnail on the node for live preview; it does not change the LATENT output.
• preview_method - auto, latent2rgb, taesd, or none. TAESD needs lighttaew2_1 in models/vae_approx and falls back to latent2rgb if it's missing.

Output is a single latent (LATENT) - decode it with a VAE node and preview, like any sampler.

Installing it

In ComfyUI Manager, search ComfyUI-CyberKrea-Sampler. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/ComfyUI-CyberKrea-Sampler

Restart ComfyUI; the node appears under the CyberKrea category. No model files are bundled - you still need the Krea 2 checkpoint, the Qwen3-VL text encoder (~8 GB) and the Qwen-Image VAE, which people routinely miss.

Traps worth knowing

• The author-name collision. This "Cyberdelia" is the ComfyUI node author over at cyberdelia.nl, not the CyberRealistic checkpoint trainer who shares the handle - different person, same hacker-chic name. Don't go looking for CyberRealistic-style settings here.
• Don't connect a plain negative and expect CFG. You get windowed guidance, not full classifier-free guidance. If you want real negatives, use NegPiP and leave the input empty.
• The raw/experimental preset and the clean-model split are gone on purpose. clean_model, variety and composition blending were deliberately omitted from this lite port - don't hunt for them.
• It's brand new and essentially untested in the wild. The repo is a few days old with basically zero stars, so there's no community troubleshooting pile to fall back on. The presets were calibrated internally (the presets module documents validation grids per value), but expect to verify your own settings.

Start at balanced, keep the defaults, and only touch detail and sigma_r once you've seen a few outputs.