Fourier-Enhanced TecoGAN Improves Video Super-Resolution Quality and Temporal Stability

A newly developed framework named Fourier-Enhanced TecoGAN consistently outperforms the original TecoGAN architecture across video super-resolution benchmarks on the REDS dataset, according to research by Md Asif Hasan and Radee Jamil Khan published in Multimedia Tools and Applications. Video super-resolution is designed to reconstruct high-resolution video frames from low-resolution inputs by exploiting spatial details together with temporal dependencies across successive frames. In benchmark testing on the REDS dataset, the enhanced framework achieved higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) scores than the original baseline. The system also reduced Learned Perceptual Image Patch Similarity (LPIPS) by approximately 20 percent and decreased temporal optical flow error (tOF), while qualitative evaluations revealed sharper textures and superior temporal stability compared to prior state-of-the-art video super-resolution techniques.

The research addresses specific technical constraints identified in the foundational TecoGAN framework. Although TecoGAN introduced adversarial training utilizing spatio-temporal discriminators and an innovative Ping-Pong loss to generate perceptually realistic and temporally coherent video sequences, its generator relies heavily on transposed convolutions. This dependency causes transposed convolutions to introduce noticeable checkerboard artifacts while restricting the generator’s capacity to recover fine high-frequency details. To resolve these generator limitations, Hasan and Khan integrated Residual-in-Residual Dense Blocks (RRDBs) adopted from the ESRGAN architecture to strengthen feature extraction. Furthermore, the researchers replaced the transposed convolution upsampling operations with PixelShuffle layers to actively suppress and reduce reconstruction artifacts.

Alongside structural revisions to the generator, the framework introduces complementary loss functions aimed at improving fine-grained visual accuracy and edge definition across frames. The authors implemented a frequency-aware loss operating directly in the Fourier domain to specifically emphasize the reconstruction of high-frequency components. To protect structural boundaries and better preserve edge details throughout video sequences, the method incorporates a dedicated gradient-magnitude loss. Training stability was also targeted through discriminator-side modifications, where the authors employed a hinge-based adversarial loss and applied spectral normalization directly to the discriminator network to enhance overall stability during training.

The documented performance gains establish that architectural and loss-function enhancements can simultaneously improve objective reconstruction metrics, perceptual quality, and temporal coherence in video super-resolution pipelines. The authors confirmed that the underlying code and pretrained model weights for Fourier-Enhanced TecoGAN will be released upon publication. These experimental findings indicate that integrating frequency-aware and gradient losses with enhanced feature extraction can effectively address checkerboard artifacts and detail recovery limitations in video super-resolution, with reported outcomes evaluated on the REDS benchmark dataset against the original TecoGAN model.

Subject of Research: Video super-resolution using Fourier-enhanced generative adversarial networks

Article Title: Fourier-enhanced TecoGAN: advancing video super-resolution with spectral and gradient losses

Article References: Hasan, M. A., & Khan, R. J. (2026). Fourier-enhanced TecoGAN: advancing video super-resolution with spectral and gradient losses. Multimedia Tools and Applications, 85(11), Article 837. https://doi.org/10.1007/s11042-026-21990-4

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21990-4

Keywords: video super-resolution, TecoGAN, generative adversarial networks, Fourier domain, PixelShuffle, ESRGAN, REDS dataset, computer vision, Fourier-enhanced, advancing, video, super-resolution