Nodes/ComfyUI-ZImagePowerNodes/❌[DEPRECATED] Z-Sampler Turbo v1 (Advanced)
ComfyUI Node

❌[DEPRECATED] Z-Sampler Turbo v1 (Advanced)

The old ^g1 sampler's advanced twin — where noise calibration came from

By martin-rizzo·Created 8 months ago·Updated 7 days ago· 357
❌[DEPRECATED] Z-Sampler Turbo v1 (Advanced)
  • model
  • positive
  • latent_input
  • divider
  • latent_output
seed1
steps8
denoise1.00
initial_noise_calibration0.00
noise_bias_estimationexperimental
noise_bias_sample_sizeimage_size
noise_bias_scale0.12
noise_overdose0.33

This is the advanced version of the original (first-generation, ^g1) Z-Sampler Turbo, and the whole thing is deprecated now. Don't let the ❌ scare you off reading about it, though - it's the node where the pack's signature idea, initial noise calibration (INC), got its real controls. The current second-generation samplers absorbed this concept and rebranded it as "intensity." If you're on a fresh workflow, use the ^G2 nodes. If you're reverse-engineering an old workflow or you're curious what that intensity knob is actually doing, this is the Rosetta Stone.

The mechanism: why Z-Image needs noise calibration

Z-Image Turbo expects its starting noise to carry a small per-channel bias. Ignore that and you get flatter, grayer images that never quite hit pure black or white. The first-gen sampler therefore does a calibration pass: it takes one step of pure-noise denoising, measures the resulting latent's bias, then corrects for it. On top of that it over-amplifies the input noise slightly. The Advanced node exposes every dial of that process:

  • initial_noise_calibration (0–1) - how much of the calibration to apply. This is the contrast/saturation control in disguise: higher = more punch, all the way to near-pure-black/white output. 100% suits clean illustrations; photos usually prefer less, and the pack docs suggest starting photos at ~50%.
  • noise_bias_estimation - experimental (fast, denoises a minimally-noisy latent) vs accurate (denoises a fully-noisy latent). Experimental is the default for a reason.
  • noise_bias_sample_size - image_size / 1024px / 512px / 256px. Smaller = faster first step, less accurate bias. The "lowres_bias" hack from the plain ^g1 node, made explicit.
  • noise_bias_scale - 0 to 1; how much of the calculated bias actually gets applied (default 0.12).
  • noise_overdose - the over-amplitude of the initial noise (default 0.33; negative values reduce amplitude).

Plus the family standard: model, positive (no negative - CFG 1.0), latent_input, seed, steps (4–9), and denoise. Output is one LATENT.

Why it's deprecated

The ^g1 sampler was capped at 9 steps, and the second generation (^G2) replaced the whole noise-calibration pile with the two intuitive intensity / intensity_bias sliders plus a wider step range and better stability. The Advanced node survives mostly inside old saved workflows; the pack keeps it importable so your graphs don't explode on update. Same caveat as all the deprecated nodes: it still runs, but it's no longer a supported surface and could be removed.

Installing it

It ships with ComfyUI-ZImagePowerNodes: 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. You need the Z-Image Turbo model files (the pack doesn't download them) and a recent ComfyUI (v0.11.0+, Nodes 2.0 API).

The honest advice

If a saved workflow forces you into this node, your main job is deciding whether to migrate it to ^G2 - which is almost always worth it. And if you're here trying to understand intensity on the new nodes: intensity ≈ the old initial_noise_calibration + noise_overdose combined, and intensity_bias ≈ a manual noise_bias_scale. Same physics, better ergonomics.

Category⚡Z-Image/__deprecated

Inputs (12)

NameTypeDefaultDescription
modelMODELThe model used for generating the latent images.
positiveCONDITIONINGThe conditioning used to guide the generation process toward the desired content.
latent_inputLATENTThe initial latent image to be modified; typically an 'Empty Latent' for text-to-image or an encoded image for img2img.
seedINT11–18446744073709550000The seed used for the random noise generator, ensuring the same result is produced with the same value.
stepsINT84–9The number of iterations to be performed during the sampling process.
denoiseFLOAT1.000–1The amount of denoising applied, lower values will maintain the structure of the initial image allowing for image to image sampling.
dividerZIPN_SEPARATOR
initial_noise_calibrationFLOAT0.000–1The amount of adjustment applied to the initial noise (0 means no adjustment). This typically enhances image contrast and saturation, higher values increase these effects more significantly.
noise_bias_estimationCOMBOexperimentalMethod used to estimate the bias in each channel of the initial noise. `experimental`: Calculate the bias by denoising a latent image with minimal noise. `accurate`: Calculate the bias by denoising a fully noisy latent image.
noise_bias_sample_sizeCOMBOimage_sizeThe size of the latent image used to calculate the bias. The smaller the image size, the faster the calculation of the first step.
noise_bias_scaleFLOAT0.120–1The level of adjustament from the calculated noise bias to apply before the first denoising step. (0.0 means no bias adjustment; 1.0 means using the calculated bias).
noise_overdoseFLOAT0.33-1–1The amount of overamplitude in the initial noise generation. (negative values will reduce the amplitude).

Outputs (1)

NameTypeDescription
latent_outputLATENTThe resulting denoised latent image, ready to be decoded by a VAE or passed to another sampler.