This project has been replaced by AFI-ForwardDeduplicate

Absolutely Eliminate Anime Stutters By SpatiotemporalResampling

Install the requirement of GMFSS
Download the weights and put it to the train_logs folder
Set the video path at the "video="
Set the save path at the "save="
Set the output frame resolution at the "global_size="
Set the resampe times at the "times=" (output_fps = input_fps / 2 * times)
Choose whether to enable scene change detection at the "scene_det=" (True/False) (If you want to ensure smoothness, please set scene_det to False)
Run the following command

python3 resample.py
Turn output frame sequenece into video using ffmpeg (set fps to output_fps)

Avoiding reduced accuracy caused by occlusion of characters in traditional one shot N recognition algorithms
Avoiding inaccurate recognition of rhythm when the proportion of characters in the screen is small and the overall movement of the screen is large
Solved the problem of multiple characters with different rhythm in the same screen
The method is also applicable to one shot two/three, two/three shot interlaced (simple and efficient)
Avoiding significant degradation in exported frame quality caused by misjudgment in one shot two/one shot three recognition
The method is applicable to any animation vfi algorithm

The processing results strongly depend on the strength of the vfi algorithm
The input video should meet the frame rate of commonly used anime videos (23.976)
The method cannot handle one shot four/five or ... which not commonly seen in anime production
In an ideal state, the overall frame sequence is offset back by 0.5 frame times. In the case of imperfect vfi algorithms, the offset is uncertain (but not more than 1.0 frame times)
Reduced the quality of the original one shot one segment