Dehaghi Ali Mollaahmadi ; Razavi Reza ; Moshirpour Mohammad

Abstract

In this paper, we introduce DiQP; a novel Transformer-Diffusion model forrestoring 8K video quality degraded by codec compression. To the best of ourknowledge, our model is the first to consider restoring the artifactsintroduced by various codecs (AV1, HEVC) by Denoising Diffusion withoutconsidering additional noise. This approach allows us to model the complex,non-Gaussian nature of compression artifacts, effectively learning to reversethe degradation. Our architecture combines the power of Transformers to capturelong-range dependencies with an enhanced windowed mechanism that preservesspatiotemporal context within groups of pixels across frames. To furtherenhance restoration, the model incorporates auxiliary "Look Ahead" and "LookAround" modules, providing both future and surrounding frame information to aidin reconstructing fine details and enhancing overall visual quality. Extensiveexperiments on different datasets demonstrate that our model outperformsstate-of-the-art methods, particularly for high-resolution videos such as 4Kand 8K, showcasing its effectiveness in restoring perceptually pleasing videosfrom highly compressed sources.