⚡ Z-Image-Turbo Scheduler (Advanced)

The plain ZImageTurboScheduler gives you the linear sigma curve Z-Image-Turbo was distilled on, full stop. ZImageTurboSchedulerAdvanced is that same curve with one more dial: denoise, so you can start the sampler partway down the schedule instead of from pure noise. That's the difference between text-to-image and image-to-image, and it's the node this pack expects you to reach for the moment you're feeding it an existing image instead of a blank latent.

Where this actually gets used

Z-Image-Turbo has become a favorite second-pass tool in the community - not just for generating from scratch, but for refining images that came out of something else. The pattern that's caught on: generate on Z-Image Base (or Chroma, or Flux, or whatever) for composition and diversity, then run a low-denoise Turbo pass over the result for speed and photorealistic polish. People have also used Turbo as a universal enhancer for outputs from entirely different models. Both of those workflows need exactly what this node provides: a schedule that starts somewhere other than full noise. Without denoise control, you're stuck doing full generations only.

The inputs and outputs that matter

• steps (INT, default 9, range 1–50) - same as the base scheduler; 8–9 is the distilled sweet spot for Z-Image-Turbo, and there's little reason to push far past it.
• denoise (FLOAT, default 1.0, range 0–1, step 0.01) - the one that changes the workflow. The node's own tooltip is direct about it: "1.0 = full generation, <1.0 = img2img strength." At 1.0 you get the exact same behavior as the base ZImageTurboScheduler - a full txt2img run from pure noise. Drop it and you're telling the sampler to start from a noisier version of your existing latent rather than from scratch, so lower values change the image less. The community's Base-to-Turbo refinement pipeline reportedly lands around 0.4–0.8 for a meaningful realism boost without throwing away the original composition; go lower still (closer to the node's own doc example of 0.5, "50% noise") if you want a gentler touch-up.

The output is SIGMAS, same as everywhere else in this pack - wire it into a SamplerCustom/SamplerCustomAdvanced node's sigmas slot. You'll also need to feed the sampler a starting latent that actually has your source image encoded into it (through a VAEEncode or similar), because denoise control here only means anything if there's something partial to denoise from.

How to install it

ComfyUI Manager: search "ComfyUI-ZImageTurbo-FlowSampler," install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-ZImageTurbo-FlowSampler

Then restart. No extra dependencies or downloads - it's the same lightweight scheduler code as the rest of the pack, with one added parameter.

Common issues & troubleshooting

Setting denoise low but forgetting to actually encode a source image into the latent. The denoise value only shortens the sigma schedule - it doesn't pull an image in for you. If your sampler's starting latent is still empty noise, a low denoise value will just under-cook a blank generation rather than refine anything.

Expecting denoise semantics to behave exactly like KSampler's built-in slider. The concept is the same - 1.0 is full noise, lower values start further along - but you're now generating the schedule yourself rather than letting a wrapper node do it silently. If your results look off at a given denoise value, cross-check with SigmaViewer from this same pack to actually see what curve you're feeding the sampler instead of guessing.

Cranking denoise up expecting more fidelity to the source. It's backwards - higher denoise means more noise added, meaning more of the original image gets overwritten. If you want to preserve more of your source, go lower, not higher.