msmsajjadi/FRVSR

Employs a recurrent neural network architecture that processes video frames sequentially, leveraging temporal information from previous frames to guide super-resolution of subsequent frames. The approach uses optical flow estimation and warp operations to align frames across time, enabling coherent multi-frame enhancement without explicit frame interpolation. Implemented in TensorFlow with training data sourced from YouTube videos, designed for 4x upsampling tasks on standard video benchmarks like Vid4.

212 stars. No commits in the last 6 months.