RIFE 4.25 (MLX)

Apple MLX port of Practical-RIFE 4.25 — real-time video frame interpolation on Apple Silicon. MIT.

First MLX/Apple-Silicon-native RIFE: torch-free inference, arbitrary-timestep interpolation, --multi Nx frame rate, --scale pyramid for 4K, audio-preserving video. Converted from the official RIFE 4.25 flownet.pkl (Google-Drive-only upstream) to fp32 safetensors.

Usage

pip install rife-mlx   # https://github.com/xocialize/rife-mlx
rife-mlx -i input.mp4 -o out.mp4 --multi 2          # 2x fps, keep audio
rife-mlx --img0 a.png --img1 b.png -t 0.5 -o mid.png

from rife_mlx.utils.weights import build_model
from rife_mlx.pipeline_mlx import interpolate_pair
model = build_model("4.25")                    # auto-downloads this repo
mid = interpolate_pair(model, frame_a, frame_b, 0.5) # HWC uint8

Details

• Architecture: IFNet (5 coarse-to-fine IFBlocks c=[192,128,96,64,32], LeakyReLU, ResConv+beta, Head encoder, ConvTranspose+PixelShuffle).
• Precision: fp32 (~23 MB) — RIFE's coarse-to-fine flow is fp16-sensitive.
• Parity vs PyTorch (CPU fp32): warp 2.2e-6 · interp 1.2e-7 · full IFNet 1.43e-3.

License

MIT (upstream Practical-RIFE, © hzwer). Weights are the official RIFE 4.25 release, converted to MLX.