leonardoventurini/RIFE-v4.9-mlx

RIFE v4.9 — MLX

Native MLX port of RIFE v4.9 (Real-Time Intermediate Flow Estimation) for video frame interpolation on Apple Silicon.

Converted from the Practical-RIFE PyTorch checkpoint. Runs on Metal GPU via MLX — no ONNX Runtime or CoreML compilation required.

Architecture

IFNet v4.7 code path: coarse-to-fine optical flow estimation with 4 IFBlocks, backward warp, and ensemble (TTA).

Property	Value
Parameters	5.33M
Weights format	safetensors (fp32)
Input	Two RGB frames [N, H, W, 3] + timestep float
Output	Interpolated RGB frame [N, H, W, 3]
Normalization	[0, 1]
Padding	Input must be padded to multiple of 64
Ensemble	Enabled by default (flip-and-average TTA)

Block Configuration

Block	Scale	Channels
0	8	192
1	4	128
2	2	96
3	1	64

Usage

from cortex_rife import load_model, interpolate_pair
import numpy as np

model = load_model()  # downloads and caches from HuggingFace

# BGR uint8 frames (e.g. from cv2.imread)
img0 = np.random.randint(0, 255, (720, 1280, 3), dtype=np.uint8)
img1 = np.random.randint(0, 255, (720, 1280, 3), dtype=np.uint8)

result = interpolate_pair(model, img0, img1, timestep=0.5)
# result: (720, 1280, 3) uint8 BGR

Citation

@inproceedings{huang2022rife,
  title={Real-Time Intermediate Flow Estimation for Video Frame Interpolation},
  author={Huang, Zhewei and Zhang, Tianyuan and Heng, Wen and Shi, Boxin and Zhou, Shuchang},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2022}
}

License

MIT — same as the original RIFE implementation.