WanLightx2vSchedulerBasicFromModel

This is the sibling of WanLightx2vSchedulerBasic that tries to be clever, and the pack's own README quietly tells you not to use it. Same idea - generate a shifted sigma schedule that matches the lightx2v training trajectory - except instead of letting you hardcode the sigma bounds, it reads them off the model you plug in.

How it works

The formula is identical to the Basic node: t_shift = shift * t / (1 + (shift - 1) * t) over steps points, output as a descending SIGMAS tensor. The difference is in the first two lines of the code, which reach into the model's model_sampling object and grab its actual sigma_min and sigma_max:

model_sampling = model.get_model_object("model_sampling")
sigma_min = float(model_sampling.sigma_min)
sigma_max = float(model_sampling.sigma_max)

Three inputs - model (any MODEL), steps (default 4), and shift (default 5.0) - and a single SIGMAS output to feed a sampler that accepts custom sigmas, like KSamplerAdvancedPartialSigmas in this pack.

The catch

The lightx2v LoRAs were distilled against a theoretical schedule that runs from 1.0 down to 0.0. Your model's internal sigma range is whatever the checkpoint ships with, and for Wan that isn't cleanly 0–1. So the from-model node produces a schedule on the model's scale, which can disagree with the trajectory the LoRA actually expects - and you're back to the burnt-out, over-sharp artifacts this whole pack exists to eliminate. The README says it plainly: the from-model version "may not match theoretical values," and recommends the Basic node with sigma_min=0.0, sigma_max=1.0 instead.

That's the whole review. If you want the fix to work, use WanLightx2vSchedulerBasic. Reach for this one only if you're deliberately experimenting outside the distilled-LoRA use case - say, comparing against the model's own sigma space for a non-lightx2v setup where the theoretical 0–1 curve doesn't apply. For the intended workflow it's the variant you'll try once and then unplug.

Install and notes

Same story as the rest of the pack - ComfyUI Manager (search "Wan2.2 Lightx2v Scheduler") or:

cd ComfyUI/custom_nodes/
git clone https://github.com/opparco/ComfyUI-WanLightx2vScheduler

then restart. No requirements.txt, no model downloads. One honest heads-up: it prints the generated sigmas to the console every run, same as the Basic node, which makes it easy to compare the two schedules side by side - worth doing once, just to see how far apart the model's sigma range and the theoretical 0–1 range actually are. If the outputs don't match, that's expected, not a bug.