❌[DEPRECATED] Z-Sampler Turbo v1

This is the node that made the pack famous - the original Z-Sampler Turbo (^g1), the custom sampler Martin Rizzo built for Z-Image Turbo that the community latched onto in January 2026. It's now marked ❌[DEPRECATED], replaced by the second-generation ^G2 family. But if you're looking at an old workflow, or you want to understand why this pack exists at all, this is where the story starts.

What it does and why it mattered

Z-Image Turbo is a distilled 6B flow-matching model that's supposed to be great at 8–9 steps - but stock KSampler runs it with a generic schedule, and the results leave quality on the table. Rizzo found, through hundreds of tests while building his Amazing Z-Image Workflow, that the model wants its steps split into three stages: a fixed 2-step composition stage with specially chosen sigmas, a details stage, and a refinement stage that climbs back up the sigma ladder, re-adds noise, and re-denoises. A discontinuous sigma jump between the first two stages made no theoretical sense - it just consistently won. Community testing backed it up: better prompt adherence, less body distortion at high resolutions, and images that were "good enough to skip the refiner" from 7 steps on.

It also kills the ModelSamplingAuraFlow ritual. Stock Z-Image workflows tell you to set the shift to 7; the Z-Sampler bakes that correction into its own sigma schedule, so you just drop it in and go.

The inputs

• model - a Z-Image Turbo checkpoint.
• positive - your prompt conditioning. No negative input: the sampler runs at CFG 1.0, where negative prompts do nothing.
• latent_input - empty latent for txt2img, an encoded image for img2img.
• seed, steps (4–9, default 9), denoise - 1.0 for txt2img, 0.75–0.9 for inpainting, 0.1–0.2 for gentle whole-image adjustments.
• initial_noise_calibration - the combo that started the pack's obsession with noise: off / 25% / 50% / 75% / 100%. It's contrast-and-saturation in disguise; 100% suits clean illustrations, photos usually like ~50%.
• lowres_bias - the "calculate the noise bias on a 256px image instead" speed hack. Keep it off; it trades quality for a faster first step.

Output is a single LATENT for VAE decode.

Why it's deprecated

The ^G2 sampler (in Simple and Extended flavors) widened the step range to 3–20, improved stability, made denoise genuinely functional for inpainting, and folded the noise-calibration knobs into the friendlier intensity / intensity_bias pair. There's no real reason to build new workflows on ^g1 - if an old saved workflow uses it, the pack keeps it loading so nothing breaks, but the migration is a five-minute job. Note the versioning quirk: g1 is capped at 9 steps, so a workflow that "wants more quality" was the usual trigger for people to switch.

Installing it

It ships with the pack - Manager → search "Z-Image Power Nodes" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes.git

No pip dependencies; recent ComfyUI required (v0.11.0+, Nodes 2.0 API); the Z-Image Turbo model files come from the README's links, not the pack.

The one honest takeaway

If you came here because you googled "Z-Sampler Turbo," you found the right family - just use the ^G2 Simple node for anything new. This old one is worth knowing about for exactly one reason: it's the clearest proof that the pack's whole approach - stage-splitting, sigma gymnastics, noise calibration - was discovered by testing, not derived from a paper. That's why the newer versions are worth your time.