SamplerLCMParallel

SamplerLCMParallel is the mystery node of this pack: it exists in the code and shows up in your node list, but the README never mentions it, it's not in the pyproject's node list, and there's no example workflow. That should tell you everything about how to treat it - it's the newest, least-documented, most-experimental thing jojkaart shipped, and you're partly a beta tester if you use it. Its idea is at least easy to grasp: at every denoising step it creates samples_per_step differently-noised copies of the current latent, runs the model on each one, and merges the predictions into a single robust average using an iterative geometric-median-style solver. The theory is that a consistency model at 4 steps is noisy about where it lands, so averaging several independent denoise attempts should cancel some of that variance out.

Inputs

• samples_per_step (INT, default 5) - how many denoise attempts per step. More = smoother, lower-variance output, and proportionally more compute. This is the dial that actually matters, and 5 is already a lot.
• consistent_noise (BOOLEAN, default false) - false means fresh noise per attempt every step. True means each of the samples_per_step slots keeps its own noise for the whole run, which trades a bit of variance reduction for temporal consistency. Leave it false unless you see the output shimmering between generations.
• adaptive_precision (INT, default 6) - the convergence tolerance for the merging solver, as 10^(-precision). 6 is fine; lower values (like 2-3) converge looser and faster, higher values are more exact for negligible visible gain. You'll almost never touch it.

Output is a SAMPLER for a CustomSampler node, paired with LCMScheduler for the schedule.

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 dependencies beyond ComfyUI's own stack; the solver is implemented in the node itself with torch.

The honest warning

Do the math before you get excited: default settings mean five model calls per step, and you still want enough steps for LCM to do its thing - so this runs several times slower than a normal 4-step LCM pass while producing a similar image. The whole pitch is that you might not need as many steps, or that the variance reduction gets you a cleaner image at the same step count, but there's no README guidance and no community write-up to lean on, which is a big reason to treat it as a curiosity. If you've got a GPU with headroom and you're chasing the last bit of stability out of a 4-step LCM-LoRA workflow, spend ten minutes with it; if your time or VRAM is precious, use SamplerLCMAlternative or SamplerLCMDuoFusion instead - the author's own favorites, both of which are documented and far cheaper per step.