⚡| Z-Sampler Turbo v2 (Advanced)

Same engine as the other second-generation Z-Sampler Turbo nodes, but this one hands you the hood. The Advanced variant is the one you grab when you need to split a generation across two samplers - a ControlNet pass for structure, then a clean pass for detail - or when you want to slice the denoise range and control exactly when noise enters the process. If you're not chaining samplers, you don't need it. The Simple node will make you happier.

What "advanced" actually adds

Every G2 sampler runs the same three-stage core: a 2-step composition stage with fixed sigmas, a variable details stage, and a refinement stage that goes back up the sigma ladder and re-denoises. That's Martin Rizzo's empirical discovery from the Amazing Z-Image Workflow, and it's what makes Z-Image Turbo look good at 8 steps with no refiner, at CFG 1.0, no negative prompt. The Advanced node then layers on the plumbing:

• add_noise - add the initial noise or not. Leave it on unless your input latent already carries residual noise from a previous sampler.
• start_at_step / end_at_step (0–100) - slice out a portion of the sampling process. This is how you run stage-one-only, or hand off to a second sampler at a specific point.
• force_final_denoising - default true, meaning the output latent is fully denoised. Switch it off when you want the next sampler in the chain to see residual noise and take over.
• initial_sample_size (256px / 512px / full_size) - how big a latent to use when calculating the initial-noise correction. Smaller is faster on step one, less accurate.
• intensity / intensity_bias - the contrast/saturation and "brightness-ish" knobs. They only take effect at denoise 1.0, so in a chain they mostly belong on the last sampler.
• turbo_creativity - off, scrambled, or refined (1-step) / refined (2-steps) / refined (3-steps). Scrambling between stages boosts composition variety but risks hallucinations; the refined options burn extra steps correcting them.

It also exposes the optional positive_stg2 and positive_stg3 conditioning inputs - different prompts for the details and refinement stages. The pack ships an example workflow ("double_trouble") that uses stage-specific prompts to merge two styles into one image.

The inputs a beginner actually touches

If you only change three things, make them these:

• steps - 3 to 20. Quality is genuinely good from 5, and 8–10 is the documented sweet spot; beyond that the curve flattens.
• seed - standard reproducibility.
• turbo_creativity - leave off until you hit the seed-diversity wall that Z-Image Turbo is famous for, then try refined (1-step).

Output is one LATENT (latent_output) that goes to your VAE decode, or into the next sampler in the chain.

Installing it

Standard pack install - Manager, search "Z-Image Power Nodes", install, restart. Or:

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

The pack has no pip requirements; the real "install" is supplying the Z-Image Turbo pieces the README links (diffusion model + Qwen3-4B text encoder + a 335 MB VAE, in diffusion_models/, text_encoders/, and vae/). And it needs a recent ComfyUI - the pack is built on the Nodes 2.0 API (minimum v0.11.0).

Gotchas

The most common footgun is forgetting what you turned off. Disable force_final_denoising and never re-enable it, and your "finished" image comes out with residual noise baked in. Same with add_noise: off is correct only when the incoming latent is already noisy. When in doubt, reset both to their defaults (true) and treat the chain as two independent generations joined at the latent. Also remember the sampler is CFG 1 - it was tuned mostly on the Q5_K_S GGUF checkpoint, and LoRAs haven't been extensively tested with it, so if you stack LoRAs and the output gets mushy, the sampler is the first suspect.