Mask Propagate (Cutie)

Optical flow has a hard ceiling: fast motion, occlusion, big appearance changes, and the mask stops being trustworthy. Mask Propagate (Cutie) is the escape hatch - it hands your first-frame mask to Cutie, a deep video-object-segmentation model, and lets the network carry the mask forward using appearance rather than motion. It's the "when things get hard" option in the mask-propagate family, and it's entirely optional: the pack works fine without it, which is why it's a separate install.

How it works

Cutie is a pure-PyTorch video object segmentation model (the successor to XMem, if you've seen that around). You give it the frames and a mask on one reference frame; it propagates that mask frame by frame, outputting a soft per-frame mask plus a confidence score. Because it's a forward-only model, a reference frame in the middle of the shot triggers two passes - forward to the end, then a fresh pass backward to the start - so mid-shot references work, at double the compute.

Expect roughly 1–3 seconds per frame at 480p on CPU. That's the price of deep segmentation, and it's the honest reason the classical nodes exist: planar is free, grid is cheap, Cutie is accurate but slow. Weights auto-download on first use.

Inputs and outputs

• images - the frames.
• mask - your mask on the reference frame.
• reference_frame (default 0) - which frame the mask is drawn on.

Outputs: masks (soft per-frame batch) and confidence (list).

Installing it

Cutie isn't on PyPI under this name, so it installs from source:

cd ComfyUI/custom_nodes/comfyui-platetrack
pip install git+https://github.com/hkchengrex/Cutie.git

Or run pip install -r requirements-optional.txt to grab the optional extras in one go. Run the node without Cutie installed and it raises a clear ImportError telling you exactly what to install.

The honest caveat

The author's own release note: the Cutie wrapper is written against Cutie's documented API but was untested on a CPU-only Python 3.14 machine at release, because Cutie wasn't installed there and has no verified 3.14 wheels. Treat this node as best-effort until you've exercised it on your own box - install it, run it on a short clip, verify the masks look sane before trusting it on a 5,000-frame job. If it misbehaves, the classical Mask Propagate in grid mode is a perfectly good fallback.

Where people get burned

• Expecting it to be fast. It isn't. Budget a minute-plus per few seconds of 480p footage on CPU and plan accordingly.
• Feeding it a sloppy first-frame mask. Cutie locks onto whatever you hand it and faithfully propagates your sloppiness; the output is only as good as the reference mask.
• Installing the whole deep stack for a flat-screen corner-pin job. Planar propagation would have been free.

Use Cutie when occlusion or deformation defeats flow. Use the cheap nodes when it doesn't. That's the whole strategy.