WanVideo ATI Comfy

Wan-Move isn't the only trajectory-based motion control node in this pack - ATI is the other one, and it takes a different route to the same goal. Where WanVideo Add WanMove Tracks folds trajectory data into your image embeds, ATI patches the model itself with trajectory conditioning. Same underlying idea - draw a path, the model follows it instead of guessing motion from your prompt - different point in the graph where it gets applied.

How it works

You give ATI a set of tracks (points and where they travel over the clip) plus the target resolution, and it returns a patched Wan model that steers denoising toward following those paths. The temperature and topk inputs are the tell that this works through a sampling-style mechanism internally - much like temperature and top-k control how much an LLM commits to its most likely next token versus exploring alternatives, here they control how strictly the model commits to the given trajectory versus allowing natural variation around it. Higher temperature and topk give the model more room to deviate from the literal path; lower values pin it down more tightly.

Because this patches the model rather than the embeds, it sits earlier in your graph - you apply it to the model before it reaches your sampler, alongside (or instead of) other model patches like LoRAs or caching nodes.

The inputs and output that matter

model (MODEL) is the Wan model you're patching. width and height (defaults 832×480) need to match the resolution you're actually generating at - trajectory coordinates are resolution-dependent, so a mismatch here means your paths land in the wrong place relative to the frame. tracks (STRING) is where the trajectory data goes in, as text - this is a different format from Wan-Move's TRACKS object type, so the two motion-control systems in this pack aren't interchangeable at the data level even though they solve the same problem.

temperature (default 220) and topk (default 2) are the two knobs actually worth touching: lower temperature and topk for tighter trajectory adherence, higher for looser, more natural-looking motion that still generally follows the path.

Output is a single model - the patched Wan model, ready for your sampler.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. No separate model download beyond your base Wan checkpoint - ATI patches the model you've already loaded rather than requiring its own weights.

Common issues & troubleshooting

Width/height mismatch is the most likely silent failure. If your tracks look right on paper but the motion in the output is offset or doesn't line up with where you drew it, check that width and height here actually match your generation's real output resolution - they default to Wan's common 832×480 baseline, which won't be right if you're generating at a different size.

Trajectory too rigid or too loose? That's temperature and topk, not tracks - before redrawing your paths, try nudging those two first. They're the equivalent of a strength dial for how literally the model treats your instruction.

This is one of two trajectory-control systems in the same pack, and they don't share data. If you're mixing WanVideo ATI with WanMove tracks in the same workflow expecting them to compose, they won't - each has its own track format and its own point in the graph. Pick one system per workflow rather than trying to combine them.