CyberZiT Scheduler

ComfyUI-CyberZiT

<p align="center"> <img src="assets/icon.png" alt="CyberZiT Scheduler icon" width="200"> </p>

CyberZiT is a lightweight preset sigma scheduler for ComfyUI. It was designed for Z-Image Turbo and is also tested with Anima-Turbo.

It adds:

• a CyberZiT Scheduler node with fast, balanced, and quality presets;
• four CyberZiT entries in the normal KSampler scheduler menu;
• an optional detail control when using the standalone scheduler node.

CyberZiT has no additional Python dependencies and does not download models or make network requests.

Installation

ComfyUI Manager

Open ComfyUI Manager, search for CyberZiT Scheduler, install it, and restart ComfyUI.

Git

Open a terminal in ComfyUI/custom_nodes and run:

git clone https://github.com/cyberdeliaAI/ComfyUI-CyberZiT.git

Restart ComfyUI after installation.

Manual installation

Download this repository as a ZIP, extract it, and place the resulting folder at:

ComfyUI/custom_nodes/ComfyUI-CyberZiT

Restart ComfyUI. No requirements.txt installation is needed.

Model compatibility

CyberZiT generates flow schedules from 1.0 to 0.0. Its intended models are Z-Image Turbo and Anima-Turbo. The KSampler entries require ComfyUI's ModelSamplingDiscreteFlow sampling class or a subclass. They stop with a clear error for other sampling types, including standard EPS-based SDXL, Pony, and Illustrious workflows, instead of returning an unsuitable schedule. A maximum sigma near 1.0 alone is not sufficient to pass this check. Other models using the accepted sampling class have not necessarily been tested with these presets.

The standalone CyberZiT Scheduler node has no model input and cannot perform this check. When connecting its SIGMAS output, use a compatible flow model; it does not make an EPS model compatible.

Preset shifts are fixed at 1.0, 1.75, and 3.0. They do not inherit or combine with the model's shift. For example, setting shift 5.0 in ModelSamplingAuraFlow and selecting cyberzit_balanced still generates a schedule with shift 1.75. CyberZiT does not modify the model's sampling object or its other settings.

Quick start

Normal KSampler nodes

Select one of these entries in the existing scheduler menu:

• cyberzit_fast
• cyberzit
• cyberzit_balanced
• cyberzit_quality

cyberzit is an alias of cyberzit_balanced. All four menu entries use the neutral detail=0.0 setting.

Useful starting points:

| Model | Steps | CFG | Sampler | Scheduler | | --- | ---: | ---: | --- | --- | | Z-Image Turbo | 8-9 | 1.0 | euler | cyberzit_quality | | Anima-Turbo | 8-12 | 1.0 | euler | cyberzit_quality |

These are starting points rather than fixed requirements. cyberzit_balanced is a useful alternative when you want a milder schedule. The quality name does not guarantee better images than another preset or ComfyUI's simple scheduler; see the comparison below.

SamplerCustomAdvanced

Add CyberZiT Scheduler from:

sampling > custom_sampling > schedulers

Connect its SIGMAS output to the sigmas input of SamplerCustomAdvanced. The node exposes:

• steps: number of sampling steps;
• preset: fast, balanced, or quality;
• detail: a small adjustment to the selected trajectory.

Presets

At neutral detail=0.0:

| Preset | Flow shift | Behaviour | | --- | ---: | --- | | fast | 1.0 | Exact linear 1.0 -> 0.0 trajectory | | balanced | 1.75 | Mild shift; a middle ground between directness and structure | | quality | 3.0 | Retains more sampling points in the high and middle sigma regions |

The detail control fine-tunes a preset without changing its identity:

• 0.0 keeps the preset unchanged;
• negative values retain more of the high-sigma structure phase;
• positive values allocate more of the path to low-sigma refinement.

detail is available only on the standalone CyberZiT Scheduler node. It does not need to be configured elsewhere. Refiner steps, CFG, denoise, and the sampler remain settings of your workflow and are not changed by CyberZiT.

All presets use the same number of sampling steps. fast does not reduce the step count, and quality does not add steps. Higher shifts keep intermediate sigmas higher and also increase the final jump to zero. At nine steps and neutral detail, that final jump is approximately 0.111 for fast, 0.179 for balanced, and 0.273 for quality. Negative detail increases it further. Image-quality tradeoffs need to be compared in your own workflow; this repository does not include a reproducible image benchmark for the presets.

Comparison with ComfyUI's simple scheduler

ComfyUI's built-in simple scheduler selects values from the model's sigma table. Its default Z-Image and Anima configurations use shift 3.0, with the same flow-shift formula as quality. With an unchanged 1,000-value flow table, eight requested steps, and detail=0.0, quality matches the simple schedule at float32 precision. At nine requested steps the largest difference is about 0.0018, due to selection of discrete table entries and floating-point rounding. Different model shifts, table sizes, denoise settings, or samplers that adjust the requested step count can change this comparison.

CyberZiT provides explicit fixed-shift presets and, through the standalone node, a detail adjustment. fast and balanced use lower shifts than the default 3.0 of these models; quality is not claimed to be a new or superior trajectory compared with simple under the matching conditions above.

Implementation references: ComfyUI's simple scheduler, flow sigma table, and model defaults.

How it works

Every CyberZiT curve:

• contains exactly steps + 1 sigma values;
• starts at 1.0 and ends at 0.0;
• is strictly descending;
• is deterministic, making presets easy to compare with the same seed.

The scheduler changes only the sigma trajectory. It does not replace the sampler and does not modify the model, prompt, CFG, denoise, or latent.

Troubleshooting

The CyberZiT entries do not appear in KSampler

Update ComfyUI and restart it. CyberZiT uses ComfyUI's scheduler registration API for the KSampler menu. If that API is unavailable in an older build, the standalone CyberZiT Scheduler node remains usable with SamplerCustomAdvanced.

KSampler reports that ModelSamplingDiscreteFlow is required

The selected model uses an unsupported sampling type. Choose a scheduler compatible with that model. Selecting CyberZiT from the global menu does not imply that it supports every model. Using the standalone node bypasses the check but does not resolve the incompatibility.

Existing workflows

The registered names are intentionally stable. Workflows using cyberzit, cyberzit_fast, cyberzit_balanced, or cyberzit_quality can continue using those exact names. Preset values, neutral detail, and generated curves for supported models are unchanged. The compatibility check only changes the KSampler path for unsupported sampling types, which now raise an error.

Development

The schedule mathematics is isolated from ComfyUI and PyTorch. Run the test suite with:

python -m unittest discover -s tests -v

License

MIT