Wan 2.2 Noise Scheduler

Wan 2.2 isn't one model, it's two. The architecture runs a high-noise pass that handles motion and scene composition, then a low-noise pass that refines detail. Every generation has to decide how many of your total steps each stage gets - and that split genuinely changes what comes out. Get it wrong and you get either mushy motion or plastic-looking detail. Wan 2.2 Noise Scheduler is a small node that just hands you that split, computed from a table the author built by actually testing eleven schedulers across a range of shift values and recording where the high-noise stage should hand off.

It's a niche convenience node, and it's honest about what it is: a curated starting point, not a law of nature. But if you're hand-wiring a two-pass Wan 2.2 graph instead of using a canned sampler workflow, it saves you from either guessing the split or re-deriving it from scratch.

How it works

Pick a scheduler, a total step count, and a shift, and the node looks up that combination in a hardcoded table of empirically measured high-noise steps (measured at 20 steps, then scaled proportionally to your total). Three inputs:

• scheduler - eleven options: normal, beta, beta57, bong_tangent, ddim_uniform, exponential, karras, kl_optimal, linear_quadratic, sgm_uniform, simple. Pick whichever your sampler is using.
• total_steps - your full step count, default 20.
• shift - the noise-schedule shift, default 5.0, adjustable 0.5–9.0 in 0.5 steps. This is the Wan shift you'd otherwise fiddle with by hand.

Three outputs: scheduler_name (the scheduler string, passed through so you can wire it straight into your sampler's scheduler input), high_noise_steps (INT), and low_noise_steps (INT) - the remainder.

What the numbers mean in practice: run the first high_noise_steps steps through the high-noise model (motion, composition, camera), then hand off to the low-noise model for the remaining low_noise_steps (texture, fine detail). For the default 20 steps at shift 5.0 with normal, that's 8 high and 12 low. The community shorthand for the split is exactly this: high = form and motion, low = texture.

Install

Same pack, same drill - Manager search "Nader Tagging", or:

cd ComfyUI/custom_nodes
git clone https://github.com/NMWave/ComfyUI-Nader-Tagging

No models, no dependencies beyond what ComfyUI ships.

What to keep in mind

The table is one person's empirical tuning, baked in for the common case. It was measured for 20 steps and scaled linearly to other totals, and real quality depends on your scheduler, your LoRAs (speed LoRAs are usually applied to the low-noise pass only - the whole community's compromise), and the footage itself. Treat the output as a sensible default to start from, then nudge the split based on what your clips look like.

It's also worth knowing that Wan 2.2 is the last open-weight Wan - everything after is API-only. So this node sits on top of a model that isn't getting a successor, which makes tooling like this a bit of a fossil-by-definition: useful while you're on 2.2, and 2.2 is where the local crowd stays.