Interpolate Frames (FILM)

How it fits

Typical backends

Caveats

• State lifetimeThe component is STATEFUL: it holds the previous Image across calls. The very first Image has no predecessor and passes straight through, so num_intermediate only inserts frames from the second frame onward.
• Hard constraintThe model file slot MUST be a serialized TorchScript FILM checkpoint (a single `.pt`). A filename containing `fp16` runs at half precision and needs a tensor-core GPU; otherwise it runs at full precision.
• Accuracynum_intermediate frames are produced by recursive midpoint subdivision. Powers of two minus one (1, 3, 7, 15) split each interval exactly in half and give the cleanest spacing; any other count emits exactly that many frames on an uneven time grid.
• LatencyPer-input latency scales linearly with num_intermediate — each inserted frame is one FILM forward pass.
• AccuracyVery large motion between two consecutive input frames can produce ghosting; FILM is tuned for moderate inter-frame motion.
• Fallbackdevice `cuda` falls back to CPU when no CUDA backend is visible — a warning is printed but inference continues at CPU speeds.
• Hard constraintThe model checkpoint, device, and num_intermediate are captured once at startup; changing any requires a redeploy.