☁️BizyAir BasicScheduler

What it is

This is BizyAir's cloud copy of ComfyUI's stock BasicScheduler - the node that turns "how many steps and how much denoise" into the actual sigma curve a sampler walks down. If you've ever built a custom sampling chain locally (Noise → Guider → Sampler → Scheduler → SamplerCustomAdvanced instead of one bundled KSampler), this is the scheduler half of that split, just running on BizyAir's servers against a model you loaded through one of BizyAir's own model-loader nodes.

You'd reach for it when you want more control than KSampler's dropdown gives you - swapping schedulers independently of samplers, or feeding the resulting SIGMAS into BizyAir's SamplerCustomAdvanced for a fully custom pipeline.

How it works

A scheduler and a sampler are two separate decisions that most UIs fuse into one name (that's where "DPM++ 2M Karras" comes from - DPM++ 2M is the sampler, Karras is the scheduler). The sampler decides how noise gets removed at each step; the scheduler decides how much gets removed and when, producing the sequence of noise levels (sigmas) the sampler steps through. This node only does the scheduler half: give it a model, a scheduler name, a step count, and a denoise strength, and it hands back the SIGMAS curve.

Which scheduler is right depends entirely on what kind of model you loaded upstream. DDPM-style checkpoints (SD 1.5, SDXL, and their finetunes) want a reshaped curve - Karras is the classic choice there. Flow-matching models (Flux, and most of what's shipped since) predict along a near-straight trajectory, and an aggressively reshaped schedule actively hurts them; beta, simple, or linear_quadratic are the safer picks for those. Karras and exponential are a near-universal bad match on flow-matching architectures - not just suboptimal, actively broken in community sweeps. Take the model card's recommended pair before guessing.

Inputs and outputs that matter

model takes a BIZYAIR_MODEL - the typed output of a BizyAir loader node, not a stock ComfyUI MODEL, so this only slots into an all-BizyAir chain. scheduler is a dropdown of nine options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal. steps (default 20) sets how many points the curve has. denoise (0–1, default 1) controls how much of the sigma range you actually use - drop it below 1 for img2img-style partial denoising, where you want to preserve more of the input.

The single output, SIGMAS, feeds directly into a custom sampler node - in this pack, BizyAir_SamplerCustomAdvanced.

Installing it

Search "BizyAir" in ComfyUI Manager and install, or git clone https://github.com/siliconflow/BizyAir.git into custom_nodes and restart ComfyUI. You need a BizyAir API key set up before any node in the pack, including this one, will run - the first-run prompt inside ComfyUI walks you through registering.

Common issues

Because model requires BIZYAIR_MODEL, dropping this node into a workflow that's otherwise stock ComfyUI just won't connect - the socket types don't match, and there's no silent fallback. If your SIGMAS output produces garbage (blown-out, oversaturated results downstream), the usual culprit is a scheduler mismatch with the model architecture rather than anything wrong with this node specifically - check whether you're on a distilled or flow-matching checkpoint before reaching for Karras out of habit. And since this whole chain runs remotely, a stuck or slow generation is as likely to be BizyAir's queue as it is your workflow.