调度器名称

LamScheduler (调度器名称, "scheduler name") is the smallest node in this pack and it does exactly one thing: let you pick a scheduler from a dropdown and pass the selection through as a generic output. It's a scheduler name as a reusable value, so you can wire it into any sampler or sampling node that accepts a scheduler input - including samplers buried in custom nodes that don't expose the dropdown in a convenient place.

Is it life-changing? No. But scheduler choice quietly matters more than most people think: the scheduler controls how noise is removed across the sampling steps, and the difference between karras and sgm_uniform on the same model and steps is often a visible quality shift. A node like this exists so you can set the scheduler once, share it across a big graph, and change it in one spot instead of hunting through five sampler nodes.

The choices

The dropdown offers nine: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal. If you're coming from SD1.5-era defaults you'll mostly live in karras (sharp, well-behaved with most models) and ddim_uniform (the classic); the newer model families (Flux, SD3, etc.) tend to want simple or sgm_uniform. It's a good reminder that this pack's target audience is SD1.5/XL-style workflows, not the newest architectures.

Inputs and outputs

• scheduler - the dropdown. That's the whole input list.
• scheduler - the selected value, typed as * so it plugs into anything accepting a scheduler.

Install

Part of the ComfyUI_Lam pack:

cd ComfyUI/custom_nodes
git clone https://github.com/yanlang0123/ComfyUI_Lam

or via ComfyUI Manager, then the README's install + 修改文件 steps. No models, no deps beyond the pack itself.

Gotchas

The * output type is permissive, which cuts both ways: it will happily plug into inputs that aren't scheduler slots, and then your sampler will reject the value at runtime. Double-check what you're wiring it into. Also, this node just passes a name through - it doesn't validate that your model actually supports the scheduler you picked, so kl_optimal into a model that never trained with it will silently underperform rather than error. When in doubt, karras or simple are the safe starts.