Skip to content

Releases: KernelPanicKitten/ComfyUI-GapFiller

GapFiller v0.1.0 (early preview)

Choose a tag to compare

@KernelPanicKitten KernelPanicKitten released this 08 Aug 08:14

First public preview.

Frame interpolation conditioned on interpolation time, so output frames are rendered at their true timestamps rather than over-generated and resampled. That matters for non-power-of-two conversions such as 24 to 60 fps.

Weights: gapfiller_v1.pt (24M parameters). Downloaded automatically on first use; no manual step required.

Benchmark (279 held-out samples, identical inputs, PSNR dB):

Split Blend RIFE 4.7 GapFiller
All samples 23.39 27.74 27.92
Generated video 23.09 28.27 28.57
Extreme motion 14.93 19.73 19.98
Arbitrary timing 22.79 27.44 27.74

Margins are consistent but modest, and the evaluation set matches the training domain. See the README for full caveats.