GMFSS

GMFlow based anime video frame interpolation

---

2023-04-03: We now propose GMFSS_Fortuna as a factual basis for training in GMFSS. Please use it. This item will not be updated!

2022-11-23: GMFSS_union —— The next generation of GMFSS with better performance.

2022-09-18: GMFupSS —— A faster GMFSS

---

Requirement

• Python 3.10.7 or other versions
• opencv-python
• pytorch
• numpy
• cupy
• pandas

Usage

• Run the following command and view the result in the 'output' folder

python test.py

Benchmarks on anime test set

• Switch to project GMFSS_Fortuna
• Download ATD-12K and unzip it, rename atd_test2k to dataset
• Put the dataset to the root of eval.py
• Run eval.py and see result at 'loss.csv'

PSNR: 29.22 SSIM: 0.932 LPIPS: 0.048

<img src="https://user-images.githubusercontent.com/68835291/190122330-1f3e0418-5e19-4383-a215-09f944cf5f85.gif" width="60%">

Also, have a look on VideoFlow’s First VFI application, VFSS!

Reference

Optical Flow: GMFlow

Video Interpolation: SoftSplat-Full SoftSplat AnimeInterp EISAI ECCV2022-RIFE

Acknowledgment

This project is sponsored by SVFI Development Team