SamplerLCMDuoFusion

SamplerLCMDualNoise from this pack is genuinely good but has one annoying trait: its weight dial needs re-tuning every time you change the step count. The author built SamplerLCMDuoFusion in July 2024 specifically to kill that friction. Under the hood it's the same dual-denoiser algorithm - it literally calls the same sampling function - but the weight is now computed adaptively per step from the sigma schedule, so you tune it once per model instead of per step count. Two inputs instead of four, same idea: at each step it runs the model once as an Euler-style denoise and once more on a fresh-noise re-version of the previous result, then blends the two predictions. Roughly double the model calls per step, but the trade exists because you can usually cut the step count hard.

Inputs - both of them

• weight (FLOAT, default 0.8) - at 0.0 the sampler behaves like plain Euler; anything above 0.0 and you're running two denoisers and combining them. Because the effective weight is scaled per step, this is a "set it per model, forget it" knob rather than a per-run one. The author's theory, shared in the README rather than as a law: photorealism likes ~1.0, anime tends to prefer ~0.6, or something in between for your particular checkpoint and style. Treat those as starting points, not verdicts.
• reuse_noise (BOOLEAN, default false) - when true, the second denoiser reuses the same noise every step instead of drawing fresh noise. Results get a bit sharper and you may want to drop the weight slightly to compensate. It's mainly an img2img tool for adding fine detail to an existing scene.

Output is a SAMPLER for a CustomSampler node, paired with the pack's LCMScheduler (or any sgm_uniform scheduler) feeding its sigmas input.

Installing it

# ComfyUI Manager: Manager → Install Custom Nodes → search "ComfyUI-sampler-lcm-alternative" → Install → Restart

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/jojkaart/ComfyUI-sampler-lcm-alternative
# restart ComfyUI

No requirements, no model downloads - the whole pack is a single Python file on top of ComfyUI's built-in sampler machinery. GPL-3.0.

Which one should you actually use?

If you're already happy with SamplerLCMDualNoise, this is a sideways move, not an upgrade - the author says it's "very similar" and only maybe slightly better on average, which he admits is hard to verify. The real argument for DuoFusion is ergonomics: you tune weight once when you change models, not every time you nudge steps, and the smaller input list is easier to wrap your head around. For a beginner that's a meaningful difference. The arguments against are the same as DualNoise's: double compute per step, and the same README advice applies - add a RescaleCFG node so you can run CFG around 3.0 with your LCM LoRA, because that's where the extra denoising pass actually shows up in the output. At CFG 1.0 the negative prompt is basically inert, which is fine for quick iteration but a reason to reach for RescaleCFG when the result matters.