Acceleration Model Pair

The most annoying part of running Wan 2.2 with speed LoRAs is not the LoRAs - it's the bookkeeping. You've got a base high-noise model and a base low-noise model, plus accelerated versions of each, plus the fact that the CFG should be ~3.5 on the unaccelerated stages and ~1 on the accelerated ones. Flip your acceleration mode and suddenly half the wires in the graph point at the wrong models with the wrong CFG. Acceleration Model Pair is the node that makes that bookkeeping automatic.

What it does

It takes the four models, two CFG values, and the acceleration_state output from Sampling Plan (Wan 2.2), then routes:

• base_high - high-noise model without acceleration
• accelerated_high - high-noise model with its acceleration LoRA
• base_low / accelerated_low - the same pair for the low-noise expert
• base_cfg (default 3.5) / accelerated_cfg (default 1) - the two CFG values
• state - the bundled SAMPLING_ACCELERATION_STATE from the plan

and outputs the selected pair plus matching CFGs: model_high, model_low, cfg_high, cfg_low. The routing follows the plan's mode - None picks both base models and both base CFGs; High only picks the accelerated high model with accelerated CFG while the low stage stays base.

The lazy part is the point

All four model inputs are marked lazy, and the node's check_lazy_status only asks ComfyUI to evaluate the branches it actually needs. That means with None, neither acceleration LoRA branch executes at all - the LoRA loaders for them aren't even run. This is the pack's recommended replacement for manually bypassing acceleration groups: leave the groups enabled, and the lazy routing handles everything. No more syncing "which group is muted" with "which mode is the plan in." The CFG travels with the models too, so a base expert can never accidentally get accelerated CFG (or the reverse) - that mismatch is a classic source of blown-out Wan renders.

Recommended wiring:

base high ───────────────┐
base high → accel LoRA ──┤
                         ├─ Acceleration Model Pair → selected high
base low ────────────────┤
base low → accel LoRA ───┘                         → selected low

Install and notes

Same pack install - ComfyUI Manager (search Sampling Planner) or:

cd ComfyUI/custom_nodes
git clone https://github.com/boobkake22/ComfyUI-SamplingPlanner

then restart. It lives under sampling / Sampling Planner / Helpers. No model files or extra dependencies come with it; you provide all four model branches yourself.

Two honest caveats. First, base_cfg and accelerated_cfg default to 3.5 and 1.0 - sensible for the LightX2V-style recipe, but check them against whatever your actual acceleration LoRA's model card says. Second, this node routes models and CFG, full stop. It doesn't validate that your "accelerated" LoRA pair is compatible (rank-64 pair at strength 1.0 is the documented baseline - don't send a full distilled checkpoint through a LoRA loader). What it buys you is that you can flip acceleration modes with one dropdown and never think about the wires again. For a Wan 2.2 workflow with speed LoRAs, that's the difference between a graph you trust and a graph you babysit.