What RIFE Model to use?

Animation

• >=rife-v4.7 in general.
• rife-v4.9, rife-v4.12, rife-v4.13 seem to have the best quality.
rife-v2.3, if speed is not a concern, and 2X is all you need, this model was the previous go-to for animation.

Speed

• rife-v4.6, decent for animation, roughly same quality as 2.3 in that aspect./li>
• rife-v4.15-lite outperforms 4.6 in almost every test, while being slightly faster (for ncnn) Would recommend the usage of that model.

General

• If there are many fast moving objects, rife-v4.13 has the best object detection in my testing. This also applies to text on screen, like credits or subtitles.
• Again, if speed is a concern, rife-v4.6 will work fine for general interpolation.

Still don't know?

• You can use my comparison script
• Usage:

  git clone https://github.com/TNTwise/Rife-NCNN-Model-Comparisons.git && cd Rife-NCNN-Model-Comparisons
  

• Replace 0.png and 2.png with 2 images you would like to test on.

  python3 render_all.py
  

Time benchmarks of every model:
ncnn:
rife 1.2 (rife): 0:00:44
rife 1.5 (rife-hd): 0:01:03
rife 1.6 (rife-uhd): 0:01:03
rife 1.8 (rife-anime): 0:01:03
rife 2.0 (rife-v2): 0:00:39
rife 2.3 (rife-v2.3): 0:00:38
rife 2.4 (rife-v2.4): 0:00:39
rife 3.1 (rife-v3.1): 0:00:43
rife 4.0 (rife-v4.0): 0:00:23
rife 4.1 (rife-v4.1): 0:00:24
rife 4.2 (rife-v4.2): 0:00:23
rife 4.3 (rife-v4.3): 0:00:23
rife 4.4 (rife-v4.4): 0:00:24
rife 4.5 (rife-v4.5): 0:00:25
rife 4.6 (rife-v4.6): 0:00:25
rife 4.7 (rife-v4.7): 0:00:30
rife 4.8 (rife-v4.8): 0:00:29
rife 4.9 (rife-v4.9): 0:00:30
rife 4.10 (rife-v4.10): 0:00:41
rife 4.11 (rife-v4.11): 0:00:39
rife 4.12 (rife-v4.12): 0:00:39
rife 4.13 (rife-v4.13): 0:00:39
rife 4.13-lite (rife-v4.13-lite): 0:00:29
rife 4.14 (rife-v4.14): 0:00:39
rife 4.14-lite (rife-v4.14-lite): 0:00:47
rife 4.15 (rife-v4.15): 0:00:39
rife 4.15-lite (rife-v4.15-lite): 0:00:24
rife 4.16-lite (rife-v4.16-lite): 0:00:24

cuda:
rife 4.6 (cuda) 0:00:21
rife 4.15 (cuda) 0:00:37
rife4.16-lite (cuda) 0:00:24