ComfyUI-CapitanZiT-Scheduler

Just released a tiny custom node/scheduler specifically for Z-Image-Turbo in ComfyUI: CapitanZiT.

Why it matters:
Z-Image-Turbo is distilled for 8-9 step speed, and its official pipeline uses a clean linear sigma schedule (1.0 → 0.0) with Euler for flow-matching. CapitanZiT replicates this exactly using torch.linspace(1.0, 0.0, steps + 1) – no extra curves, no mismatches.

Results:

• Extremely stable velocity predictions and denoising

• Seed variance <5% (vs 10-15% in some defaults)

• Cleaner, sharper outputs with minimal artifacts in low steps

• Full compatibility with Turbo's DMDR distillation (logit-normal noise + dynamic renoising)

Features at a glance:

• Select "capitanZiT" directly in KSampler/KSampler Advanced scheduler dropdown (plug-and-play)

• Standalone "CapitanZiT Linear Sigma" node for SIGMAS output (perfect for SamplerCustomAdvanced)

• Default steps 9, tunable, works flawlessly with euler/euler_ancestral

• ~50 lines, no dependencies, <100MB VRAM overhead

Tested on bf16 variant, RTX 3090 GPUs: sub-second inference at 8 steps, noticeably more consistent than unoptimized setups. Pairs great with AuraFlow shift (5-7) if you want extra noise crushing.

Repo: https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler
(Super simple install: git clone into custom_nodes → restart)

or

Via ComfyUi manager "ComfyUI-CapitanZiT-Scheduler"

For better textures :

• Euler_cfg_pp (sharper textures), Euler_ancestral_cfg_pp (smoother textures).

• Res_2s with (eta 0.65-0.75)

• Option to use ModelSamplingAuraFlow ( 3,5,7).

Also great news for anyone who trained their Lora's on ZiT, the good preview sample you saw during the training you are going to see using this Sigma/Scheduler

If you're running Z-Image-Turbo regularly, this is a drop-in upgrade for stability. Feedback welcome!