☁️BizyAir VPScheduler

Honest take up front: this node exists because ComfyUI's own core ships an equivalent, and BizyAir mirrors a lot of core functionality behind its cloud execution. It's not something the vast majority of workflows will ever touch. If your model card doesn't explicitly mention a variance-preserving SDE, you don't want this node - you want Karras, beta, or simple, matched to whether your checkpoint is DDPM-style or flow-matching.

That said, if you know why you're here, here's what it actually does.

What it is

This generates a SIGMAS tensor from the raw hyperparameters of the Variance-Preserving (VP) SDE - the noise formulation from the score-based diffusion literature that predates the Karras/EDM schedule most people actually use today. It's a scheduler in the strict sense: it decides how much noise gets removed at each step, separately from which sampler algorithm does the removing. The output plugs into a custom sampling chain (something like SamplerCustom) alongside a compatible sampler node - it's not a drop-in replacement for the scheduler dropdown on a normal KSampler.

Why does this matter at all? Because the schedule has to match the model. Modern flow-matching checkpoints (Flux-family, Wan, Z-Image and friends) train on a near-straight denoising trajectory, and aggressive reshaping - which is exactly what a hand-tuned VP schedule does - tends to actively break them rather than just underperform. This node is for the older DDPM-style corner of the ecosystem the VP-SDE formulation was actually built for, not for anything you're likely running in 2026.

Inputs and outputs that matter

• steps (default 20) - how many sigma values to generate. Standard step-count meaning, same as any scheduler.
• beta_d (default 19.9) - the drift coefficient of the VP-SDE, controlling how quickly noise ramps up toward the end of the schedule.
• beta_min (default 0.1) - the noise floor near the start of the schedule.
• eps_s (default 0.001) - a numerical-stability floor on the smallest timestep the schedule will produce.

If those three names don't mean anything to you, that's the tell that this isn't your node - they're research-paper hyperparameters, not the kind of thing you tune by feel the way you'd nudge CFG or steps.

Output is SIGMAS, a single socket, meant to feed a custom sampler's sigma input.

How to install it

• ComfyUI Manager - search "BizyAir", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, restart.
• Comfy CLI - comfy node install bizyair.

No API key gate specifically blocks this node from appearing in your menu, but the broader pack expects one configured for its cloud-execution nodes - set it up via the in-app login prompt or bizyair.siliconflow.cn once, and it covers everything in the pack.

Common issues & troubleshooting

Output looks like noise, not a coherent image, and there's no error. This is the actual danger with a scheduler node: it will happily compute sigmas regardless of whether they suit your model. A mismatched schedule doesn't throw - it just produces garbage. If you're getting noise-adjacent output and you're on a modern flow-matching checkpoint, that's very likely this node fighting the model's training distribution. Swap to a standard beta or simple scheduler and see if the problem disappears; if it does, you had the wrong tool for the job.

You're not sure if your model wants this. Check the model card, and check the release date is not a reliable proxy - an SDXL finetune released in 2026 is still DDPM-style, and even then, this specific VP parameterization is a narrower case than "DDPM-style" generally. When in doubt, reach for whatever schedule the model's own example workflow ships with instead of assembling one from raw hyperparameters.