Nodes/ComfyUI Extra Samplers/SamplerLCMCustom
ComfyUI Node

SamplerLCMCustom

The LCM sampler, but with your choice of noise

By Clybius·Created 3 years ago·Updated 2 years ago· 104
SamplerLCMCustom
    • SAMPLER
    noise_sampler_type

    SamplerLCMCustom is the ComfyUI Extra Samplers take on the LCM sampler: same few-step consistency-model math, but with one extra knob - you get to pick the noise sampler. From the pack by Clybius, it outputs a SAMPLER and shows up in the sampler dropdown after install. It is a thin node, and it should be: its only input is noise_sampler_type.

    Here's the honest setup, because this is where people go wrong with everything LCM-shaped. LCM (Latent Consistency Models) is a distillation technique - the model has been trained to jump to the answer in a handful of steps instead of walking down the whole noise schedule. That carries two hard rules from the distillation playbook: it demands a dedicated sampler (this one), and it demands very low CFG, around 1–2, because the guidance is already baked into the weights. The knowledge base's distillation essay says it plainly: run normal defaults on a distilled checkpoint and you get oversaturated artifacts. So the "Custom" in the name is about noise, not about CFG or steps - those rules are non-negotiable.

    What the node actually adds: stock LCM in ComfyUI always uses default Gaussian noise sampling. This variant lets you choose among the pack's noise samplers - gaussian, uniform, brownian, pyramid, perlin, laplacian, highres-pyramid - for the per-step noise injection. In source terms it's a few lines wrapping ComfyUI's own sample_lcm, swapping in the pack's get_noise_sampler helper. The practical effect is a subtle change in output texture and variety at your (low) step count, which is a genuinely reasonable thing to want when 4–8 steps is all you get and every bit of variety is precious.

    The inputs that matter:

    • noise_sampler_type - the only input. Gaussian is the safe default; pyramid and perlin are the interesting experiments.

    Output: one SAMPLER, wired into a custom sampler node that takes sigmas, conditioning, and a low CFG.

    Worth saying plainly: if you don't have an LCM-distilled model or an LCM LoRA, this node has nothing to do - the stock lcm sampler in ComfyUI is the same thing minus the noise choice. You'd only install and reach for this once you're running LCM weights. And when you do, remember the rules: few steps, CFG 1–2, and this sampler.

    Install is the shared pack routine - ComfyUI Manager → "ComfyUI Extra Samplers", or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Clybius/ComfyUI-Extra-Samplers
    

    then restart. No model downloads; dependency kornia.

    Categorysampling/custom_sampling/samplers

    Inputs (1)

    NameTypeDefaultDescription
    noise_sampler_typeCOMBO7 options: gaussian, uniform, brownian, highres-pyramid, pyramid, perlin, +1

    Outputs (1)

    NameTypeDescription
    SAMPLERSAMPLER