SamplerLCM

Quick gut-check before you install anything: if all you want is fast LCM-style generation, ComfyUI's plain KSampler already has lcm sitting right there in its sampler_name dropdown - no custom node required. What this node from laksjdjf's LCMSampler-ComfyUI pack gives you that the built-in dropdown option doesn't is control over the noise LCM re-injects at each step: three knobs the stock option just doesn't expose. The pack itself dates back to when LCM (Latent Consistency Models) was brand new - the trick that collapses 20-50 denoising steps down to 4-8 - and was built specifically to drive an LCM-ified LoRA for the SSD-1B-anime checkpoint.

What it actually does, per its own description: it's an LCM sampler with tunable per-step noise. LCM-style sampling doesn't just walk down a denoising schedule the way Euler or DPM++ do - it jumps to a prediction of the final image almost immediately, then re-injects a calibrated dose of noise before the next jump, repeating for a handful of steps. That re-noising step is exactly what this node's three inputs control: how much noise goes back in, and how that amount shifts over the course of your steps.

The inputs that matter

• s_noise (default 1.0, 0-64) - the noise multiplier applied at your first step. 1.0 matches what the model was trained on; the author's own tooltip is explicit about that.
• s_noise_end (default 1.0, 0-64) - the same multiplier, but for your last step. Leave it equal to s_noise for a flat, constant amount of noise across all steps; set it different and you get a ramp from start to end.
• noise_clip_std (default 0, 0-10) - a safety clamp: caps per-step noise to plus/minus N standard deviations. 0 means no clamp. Handy if a high s_noise is throwing outlier artifacts you want capped rather than tuned away entirely.

It outputs a single SAMPLER - which is the thing to notice before you wire anything up.

Where it actually plugs in

This is not a KSampler replacement you can drop straight into a normal workflow. It's a component for ComfyUI's custom sampling chain - the pack's own README is blunt about this: don't use the regular KSampler, use the custom-sampler-compatible setup and pull this node in as the sampler. In practice that means a SamplerCustom (or SamplerCustomAdvanced) node taking your model, a noise source, a sigmas/scheduler, and this node's SAMPLER output. If you've only ever used the plain KSampler, this is the one adjustment that trips people up first - the node works fine, it's just not going where a beginner instinctively reaches.

The two lives of this node

The pack's original use case is narrow: it was built specifically to run an LCM-ified LoRA for the SSD-1B-anime checkpoint. The README is upfront that vanilla/official LCM checkpoints aren't a guaranteed fit - at the time of writing it says the current version can't even load those weights. If you're chasing "real" LCM speed with a mainstream distilled checkpoint, budget for some fiddling, and keep the distillation rule in mind: CFG has to come down to roughly 1-2 and step count to 4-8, because the model's own training already baked in the guidance a normal CFG value would apply a second time.

The more common way people actually run into a node called SamplerLCM today is different: on flow-matching (Flux-family) workflows circulating in 2026, tuners pair it with ModelSamplingFlux and use s_noise_end purely as a per-step noise/diversity knob - nothing to do with LCM's original "compress the steps" idea at all. If you find a SamplerLCM node already sitting in a workflow you downloaded, paired with ModelSamplingFlux rather than an LCM checkpoint, that's what's going on.

Common issues

Oversaturated, artifacted, or "burnt" output - classic sign of running a distilled/LCM-trained checkpoint at normal CFG and step counts. Drop CFG to 1-2 and steps to single digits before touching anything else.

Nothing happens, or it errors when wired into a plain KSampler - it can't be; this node produces a SAMPLER object, not an image. It needs to go into a custom-sampling chain.

Weights won't load - if you're trying vanilla/official LCM weights rather than the SSD-1B-anime LoRA this pack was built around, that's a known limitation per the author, not something you're doing wrong.

Weird noise or blotchy artifacts at high s_noise - push noise_clip_std up from 0 to clamp the outliers instead of fighting it purely through s_noise/s_noise_end.