Wan Motion Scale (Experimental)

The single most annoying thing about Wan 2.2, after the 81-frame ceiling, is that "the car drives across the screen" comes out at whatever pace the model felt like. WanMotionScale is the node from the comfyUI-LongLook pack that finally gives you a motion-speed dial: scale_t above 1 makes the motion faster, below 1 makes it slower, and the whole thing is one node between your loaded model and the sampler.

The mechanism is elegant because it's basically a cheat. Wan positions frames in time using RoPE (rotary position embeddings), and this node just scales the temporal positions before the model ever sees them. Crank scale_t to 1.5 and the model thinks each frame is further apart in time than it actually is - so it invents more motion between them. It doesn't add any new attention math or touch the weights; it sets ComfyUI's built-in rope_options on a cloned model and hands it back. That's why the output is a MODEL you wire straight into your sampler, and why there's an enabled toggle sitting right on top - flip it off for an A/B comparison, or to confirm the patch is actually doing something.

The inputs are blessedly short. model in, enabled on by default, then scale_t (default 1.5), plus optional scale_y and scale_x for the spatial axes. The author's recommended values are the useful part:

• 1.5 - the sweet spot for acceleration, feels like roughly 2x motion speed and stays stable
• 1.0–1.5 - safe speedup range
• 0.75–1.0 - slowdown, works reliably
• negative - reversed time positions. Fun to try once, but it contradicts what the model trained on, so results are inconsistent. Don't build a workflow on it.

The headline trick is pairing it with frame interpolation. Generate at scale_t=1.5, run the result through RIFE, and you effectively double the clip's length while keeping the same motion coverage - faster movement plus smooth filler frames instead of a short clip that just... is short. The README's whole pitch for this node is exactly that RIFE workflow.

scale_x and scale_y are more experimental. For t2v they can nudge the effective aspect ratio of the generation; for i2v the author flat-out warns they produce "wild spatial effects." If you're not chasing that, leave them at 1.0.

Installation is the pack's one shared story: ComfyUI Manager, search "comfyUI-LongLook," install, restart. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/shootthesound/comfyUI-LongLook.git

There's no requirements.txt - no new Python deps, no model downloads, nothing to configure. It patches whatever Wan 2.2 checkpoint you already load, i2v or t2v.

Gotchas are few. If you set everything to 1.0 and leave enabled on, the node detects the no-op and returns your model unpatched, which is correct but can confuse you into thinking it's broken - change something and watch the console log confirm the patch. And remember the patch is baked into the model object at load time, so if you change scale_t mid-session, re-run from the node (or reload the workflow) to pick it up.