CyberZiT Scheduler

If you've run Z-Image Turbo or an Anima-based checkpoint, you've probably hit the wall this node exists for: these are distilled flow-matching models that want 8–12 steps at CFG 1, and they're famously hostile to the schedulers everyone carries over from SDXL. The community's own words: Karras and Exponential are "universal failures" on Z-Image Turbo because it "really dislikes aggressive redistribution of denoising effort." Generic schedulers assume a long denoising run; a turbo model has almost none to spend. CyberZiT Scheduler is a tiny, opinionated fix for exactly that - preset sigma curves tuned for Z-Image Turbo and Anima-Turbo, from the cyberdeliaAI org whose CyberRealistic line is the one building checkpoints on both.

What it actually does

A scheduler in ComfyUI just decides when each of your steps happens along the noise-to-image path - the sigma trajectory. CyberZiT replaces the guesswork with a deterministic curve: exactly steps + 1 sigma values, descending from 1.0 to 0.0, every time. Same seed, same curve, trivially comparable.

The three presets are just different amounts of flow shift:

• fast - shift 1.0, which is a straight line from 1.0 to 0.0. The most direct path, and what a distilled model mostly needs.
• balanced - shift 1.75, a mild bend that keeps a bit more structure without a hard final jump.
• quality - shift 3.0, which parks more of your steps in the high and middle sigma regions.

You get it two ways. There's the standalone node under sampling > custom_sampling > schedulers, which outputs SIGMAS that you feed into a SamplerCustomAdvanced. But it also silently registers four entries - cyberzit_fast, cyberzit, cyberzit_balanced, cyberzit_quality - into the ordinary KSampler scheduler dropdown. So if you're on the standard Z-Image Turbo workflow, you may not need the node at all: just pick cyberzit_quality where you used to pick simple or beta. (cyberzit is just an alias of cyberzit_balanced.)

The inputs that matter

Three knobs, and honestly you'll touch two:

• steps (default 9). The tooltip says it plainly: 8–9 for Z-Image Turbo, 8–12 for Anima. Push past ~12 on a distilled model and you're spending steps on nothing.
• preset (default balanced). The README's starting point for both models is euler sampler, CFG 1.0, cyberzit_quality. Drop to balanced when quality looks overcooked or you want a milder schedule.
• detail (default 0, range −1 to 1). The one genuinely new idea here, and it's deliberately restrained - it applies a small power-curve tweak before the shift, scaling the exponent between about 0.75 and 1.33. Negative favors high-sigma structure; positive pushes more path into low-sigma refinement. It only exists on the standalone node - the KSampler menu entries always run at neutral 0.

Output is a single SIGMAS, wired to the sigmas input of SamplerCustomAdvanced alongside your model, positive/negative conditioning, and sampler. That's it. The node doesn't touch CFG, denoise, the sampler, or the model.

Installing it

No drama here. ComfyUI Manager: search CyberZiT Scheduler, install, restart. Or:

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

then restart ComfyUI. There is no requirements.txt to run, no pip dependencies, no model downloads, and no network calls - the schedule math is pure Python with zero PyTorch coupling. It works on Python 3.10+, which the changelog cares about enough to have patched.

Where people get burned

The most common failure is also the most confusing: the CyberZiT entries don't show up in your KSampler dropdown. That means your ComfyUI build predates the scheduler registration API the pack leans on. Update ComfyUI and restart - and know that the standalone CyberZiT Scheduler node works regardless, so nothing is bricked in the meantime.

Second: don't read the quality preset as "give me more steps." It reallocates the steps you already have. And don't crank CFG thinking it'll sharpen things - these are guidance-distilled models, and the whole point of the calibrated curve is CFG 1.0.

One honest caveat: as of this writing there's no community chatter about the pack at all - no Reddit threads, no "this fixed my Z-Image workflow" posts to lean on. The good news is it's cheap to A/B: generate one seed with your old scheduler, one with cyberzit_quality, and let your own eyes settle it. If you're chasing the gallery look from the CyberRealistic Z-Image or Anima pages, a model-matched schedule like this is exactly the variable that explains why your output and their samples never quite agreed.