Feb 3rd, 2026 Update v1.1.0v1.1.0 adds two optional components alongside the original CapitanZiT linear scheduler:<ul><li>Smooth Cosine Scheduler Alternative sigma schedule using a cosine curve (1.0 → 0.0). More gradual start/end than linear, useful for smoother transitions and photorealistic subjects. Works with low or higher step counts.</li><li>Minimal Change Flow Sampler New sampler (not a scheduler). Limits how much the latent can change per step to prevent drift and instability. Adjustable <code>max_change_per_step</code> (default 0.70). Can be used with any sigma schedule (CapitanZiT, Cosine, or others).</li></ul>Nothing about the original behavior changed — CapitanZiT linear sigma is still the same.You can now mix:<ul><li>CapitanZiT + Euler → fastest</li><li>CapitanZiT + Minimal Change Flow → speed + extra stability</li><li>Smooth Cosine + Minimal Change Flow → max smoothness</li></ul>Details, math, and usage examples are in the <a target="_blank" rel="ugc" href="https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler">GitHub</a>.<hr />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:<ul><li>Extremely stable velocity predictions and denoising</li><li>Seed variance &lt;5% (vs 10-15% in some defaults)</li><li>Cleaner, sharper outputs with minimal artifacts in low steps</li><li>Full compatibility with Turbo's DMDR distillation (logit-normal noise + dynamic renoising)</li></ul>Features at a glance:<ul><li>Select "capitanZiT" directly in KSampler/KSampler Advanced scheduler dropdown (plug-and-play)</li><li>Standalone "CapitanZiT Linear Sigma" node for SIGMAS output (perfect for SamplerCustomAdvanced)</li><li>Default steps 9, tunable, works flawlessly with euler/euler_ancestral</li><li>~50 lines, no dependencies, &lt;100MB VRAM overhead</li></ul>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: <a target="_blank" rel="ugc" href="https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler">https://github.com/capitan01R/ComfyUI-CapitanZiT-Scheduler</a> (Super simple install: git clone into custom_nodes → restart)orVia ComfyUi manager "ComfyUI-CapitanZiT-Scheduler"<h3 id="for-better-textures-:">For better textures :</h3><ul><li>Euler_cfg_pp (sharper textures), Euler_ancestral_cfg_pp (smoother textures).</li><li>Res_2s with (eta 0.65-0.75)</li><li>Option to use ModelSamplingAuraFlow ( 3,5,7).</li></ul>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/SchedulerIf you're running Z-Image-Turbo regularly, this is a drop-in upgrade for stability. Feedback welcome!

ComfyUI-CapitanZiT-Scheduler ((UPDATED))

v1.1.0Added new sampler &amp; scheduler

Files with SHA256: 0be030e8cdffee99009e073e6c710d459946895eccb86540f4e0265889a78290