⬡ Polyhedron Interpolate

Wan 2.2 ships its clips at 16fps as a convention, and 16fps is fine until you want the footage to actually play smooth. That's where frame interpolation comes in - RIFE-style flow interpolation invents the in-between frames so a 16fps clip becomes 32 or 48, or just moves more smoothly at the same rate. The Polyhedron Interpolate is a RIFE node, but it's a RIFE node with an unusual amount of honesty in the details.

The engine itself is not theirs: it's the MIT-licensed IFNet from Practical-RIFE, vendored byte-for-byte from Fannovel16's ComfyUI-Frame-Interpolation on purpose, and every correction this pack makes lives outside the engine so it can be audited. And it needed corrections, because the package it forked from has problems the source header documents in embarrassing detail. The big one: the model in the incumbent's own dropdown - rife426.pth - ships its training graph inside the checkpoint, and the upstream loader crashes on it. The Interpolate strips those training tensors by name so the model the RIFE author actually recommends for diffusion footage loads instead of exploding. It also calls out the "knobs that lie": fast_mode does nothing from arch 4.5 up, and ensemble is silently forced off on 4.26 - both are reported honestly in the console instead of pretending.

The inputs that matter

• ckpt_name - which RIFE checkpoint (rife417, rife426, rife47, rife49; prefer the newest for diffusion-model footage). Weights fetch into your usual ComfyUI model folders on first use.
• rate_mode - multiplier (2×, 4×) or target_fps (name the rate you want; the node works out the ratio and reports what you'll actually get).
• precision - float32 is the default on purpose: this network's coarse-to-fine flow is sensitive to half precision, and the weights are ~23MB, so there's no memory to save. The faster modes are a measurement, not a setting.
• static_skip - pairs whose two frames differ by less than this are duplicated instead of interpolated. Identical pixels cannot produce motion, only noise.
• cut_guard - pairs that differ by more than this are held, not blended. Interpolating across a scene cut produces a dissolve, which at the seam of a joined sequence is a certainty. This is the artefact that gives machine interpolation away fastest.

Take a frame batch OR a video. A wired video brings its own fps, so no fps wire is needed - and the tooltip warns that an fps wire from a Save behind a de-selected switch carries an ExecutionBlocker that silently disables the whole node. Don't.

The alignment gotcha

RIFE's IFNet pads to a mod-64 canvas internally, with zeros bottom-right, and crops back. A canvas that isn't a multiple of 64 therefore gets a black ring the flow estimator tries to track - measured here at ~16 dB worse at the border. The node's answer is to not "fix" it silently: it warns you to keep your canvas on the mod-64 grid, because its own experiments showed a "nicer" centered-replicate pad measures worse than the network's native zeros. Feed it 1088-wide frames, not 1075.

Outputs: frames, frame_count, fps, and interp_info (a string reporting what actually ran and which knobs were ignored). One law worth remembering: k frames per pair over N inputs yields k*(N-1)+1 outputs. 65 frames at 2× gives 129, not 130 - and now you know why.

Installing it

Same pack as everything else - ComfyUI Manager search "Polyhedron Suite", or clone https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack.git into custom_nodes and restart. No hard dependencies; the RIFE weights download into the standard ComfyUI model folders on first run.