Frame Interpolation using Google AI Research tool – GitHub code

Frame interpolation is a process to fill in-between frames that morph one image to another image. AI software can speed up the process. The limitation is that the changes between two images are subtle and predictable.

1. Create a “interpolation” directory on your local drive. Open a dos prompt, navigate to this directory. Download the code from  frame-interpolation GitHub repository by the command:
>git clone https://github.com/google-research/frame-interpolation.git
A new directory “frame-interpolation” is created under “interpolation.”
2. Under “interpolation” directory, create another directory called “pretrained_models.” Go to Google Drive to download “film_net” and “vgg.” Put them under “pretrained_models” directory.
3. You need to configure a virtual environment to run the code. If you haven’t installed Anaconda3, go to install Anaconda3.
4. Setup a conda environment with the instruction. If the instruction confuses you, you can download a ready-for-use tf_env at Gumroad (compatible with CUDA 11.8). Unzip it and put it at your anaconda3 installation under “envs” directory.
5. Prepare two images for the start and the end frames. Rename files to “one.png” and “two.png.” Put them under “frame-interpolation\photos” directory.
6. Open an Anaconda Prompt. Run command:
>conda activate tf_env
7. Still in the Anaconda prompt, go to the directory “interpolation” and run command:
>python -m frame-interpolation.eval.interpolator_cli --pattern “frame-interpolation/photos” --model_path pretrained_models/film_net/Style/saved_model --times_to_interpolate 6 --output_video
8. When it finishes, the image sequence and a new video mp4 file are saved at “frame-interpolation\photos” directory.
9. If you cannot play the mp4 file due to encoding differences, import the image sequence to After Effects or other video editing tools to render as a video.