⚡| Z-Sampler Turbo v2 (Extended)

The "Extended" Z-Sampler Turbo is the original second-generation node - the one that replaced the deprecated ^g1 sampler and became the pack's flagship. It's the same three-stage core as the Simple version (2 fixed-step composition stage, details stage with the famous discontinuous sigma jump, then a refinement stage that climbs back up and re-denoises), but it hands you the knobs the Simple node hides: denoise strength, initial noise calibration, and the full turbo-creativity menu. Think of it as "Simple, plus the controls you'll actually want once the defaults stop being enough."

What the extra controls buy you

• denoise (0–1, default 1.0) - this is the big one, because it's what makes the node work for img2img and inpainting. The docs lay it out plainly: 1.0 for txt2img, 0.8–0.9 for inpainting (pair it with the pack's VAE Encode for Soft Inpainting), and 0.1–0.2 for subtle whole-image adjustments that keep the composition.
• intensity (-1 to 1) - adjusts the amplitude of the initial noise, which reads as contrast and saturation. Positive = punchier, sharper edges; negative = softer, washed-out. It only takes effect at denoise 1.0, so in an img2img pass it's a no-op. Photographic styles tend to like lower values, illustrations higher - the README's own guidance.
• intensity_bias (-1 to 1) - the companion "brightness-ish" knob, usually left at 0.0. It's prompt-dependent and can even affect apparent focus, so treat it as a fix-what-looks-wrong dial, not a real brightness slider. If intensity is cranked high, a negative bias sometimes pulls the image back into balance.
• initial_sample_size (256px / 512px / full_size) - how big a latent to use when measuring the initial noise for the intensity correction. Smaller = faster first step, less accurate. 256px is "very fast but a hack" per the docs; full_size is the correct-but-slower default.
• turbo_creativity - this is where Extended beats Simple: instead of just on/off you get off, scrambled, and refined (1/2/3-steps). The refined tiers add sampling steps after the latent scrambling to pull the image back from hallucination. Scrambling only changes pose, framing, and placement - colors and style hold.

There are also the optional positive_stg2 and positive_stg3 conditioning inputs for stage-specific prompts (the "double_trouble" example merges two styles this way), plus latent_input, model, positive, seed, and steps (3–20; 8–10 is the sweet spot, per the docs).

Why Extended and not Simple

Honestly? Start with Simple. It hides the noise-calibration internals because they're finicky and most people never need them. Reach for Extended when you want the denoise dial (inpainting, img2img) or when you want to push contrast/saturation deliberately via intensity, or when turbo-creativity's refined tiers are worth the extra steps to you. The intensity/intensity_bias pair is also genuinely fun to play with - it's the closest thing this sampler has to a "look" control, and it's the parameter most of the pack's sample galleries were built to showcase.

Installing and running

Pack install, as always - Manager → "Z-Image Power Nodes" → Install → restart, or:

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

No pip dependencies; a recent ComfyUI is required (v0.11.0+, Nodes 2.0 API). The pack won't fetch the model - you supply the Z-Image Turbo diffusion model, Qwen3-4B text encoder, and ~335 MB VAE from the README's links, into diffusion_models/, text_encoders/, and vae/.

Gotchas

Remember the intensity controls only bite at denoise 1.0 - crank them in an inpainting pass and nothing happens, which confuses people. Turbo-creativity isn't for inpainting (latent scrambling and a mask don't mix). And the usual family caveats apply: CFG is fixed at 1.0, there's no negative prompt, and LoRA/fine-tuned checkpoint behavior hasn't been extensively tested - if stacked LoRAs go mushy, this sampler is the first suspect.