⚗️🔬| Z-Sampler Turbo ^X2.1

Z-Sampler Turbo ^G2.1 is where the pack's author experiments with what comes after the second generation. It's filed under the ⚗️🔬 experimental menu, which in this pack means: it works, it's genuinely interesting, and it could change or vanish in the next release without notice. If you want a stable daily driver, the ^G2 Simple node is still your answer. If you want a peek at the sampler's future - and a couple of knobs nothing else in the pack has - this is it.

What's new in G2.1

The three-stage core (composition / details / refinement) is still there, but it's wrapped in new machinery:

• latent_scaling - instead of sampling at full resolution the whole way, this runs the stages at reduced internal scales and upscales: max_speed uses (0.25 / 0.50 / 0.75) scaling across stages for faster generation, max_quality uses (0.50 / 0.75 / 1.00) to preserve fine detail and structure. "none" is also an option. This is a real speed/quality trade, not marketing: max_speed genuinely cuts processing time at the cost of micro-detail.
• spectral_tilt - a frequency-based bias on the noise. Early stages get low-frequency noise boosted (that's what locks in composition), later stages shift to high-frequency components (that's what sharpens detail). The presets (none, stage3_H, stages12x_H, stages12x_l, stages123_H) pick which stages get the tilt and how hard. It's the kind of thing that sounds like magic and mostly shows up as "sharper but sometimes slightly off" - worth a test, not a default.
• initial_bias (-0.5 to 0.5) - the G2 intensity_bias idea, re-tuned. Positive amplifies dominant prompt features (bright scenes brighter, dark scenes darker); negative tempers them. The tooltip's warning applies: it reacts differently to every prompt, it's not a brightness slider.
• detailed_refiner (default true) - the DPM++ SDE refiner for the final stage, boosting contrast and fine-detail sharpness at the cost of time. This is the alternative_refiner from the G2 Simple node, promoted to default.
• new_scheduler (default true) - the optimized sigma preset; turning it off drops to the legacy scheduler, which can win in specific edge cases.
• turbo_creativity - same latent-scrambling feature as the rest of the family, here as a simple on/off, with the same hallucination warning.

Steps now run 2–14 (a wider band than G2's 3–20, oddly enough, but the extra low end means it can draft at 2 steps for fast iteration). The optional positive_stg2 / positive_stg3 inputs carry over, and output is the usual single latent_output for VAE decode.

How to actually use it

It slots into the same workflow as any Z-Sampler: empty latent or encoded image in, model and positive conditioning in, LATENT out to the VAE. My honest suggestion is to start with everything at default (which already includes the experimental refiner and new scheduler), then toggle latent_scaling to max_speed for iteration and back to max_quality for finals. spectral_tilt is the fun one to A/B - just do it on a fixed seed.

Installing it

Pack install as usual - 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, recent ComfyUI required (Nodes 2.0 API, v0.11.0+), and you supply the Z-Image Turbo model + Qwen3-4B encoder + VAE yourself from the README's links.

The catch

It's experimental, which is both the appeal and the trap: parameters can shift between releases, so a workflow you save today might need re-plumbing after an update. Keep your favorite ^G2 workflow as the fallback while you play with this one. And the family rules still hold - CFG is 1.0, no negative prompt, and LoRAs are not extensively tested territory.