New FGSVQA model assesses short-form video quality with frequency cues

Short-form videos (e.g., TikTok, Instagram Reels) present unique quality assessment challenges due to complex generation pipelines, rapid content changes, and mixed distortions. To tackle this, researchers from the University of Bristol introduce FGSVQA (Frequency-Guided Short-form Video Quality Assessment), an end-to-end framework that leverages a dense visual encoder built on CLIP (Contrastive Language-Image Pretraining). The key innovation is the incorporation of compression priors derived from the frequency domain, which helps isolate artifacts from structural content. The architecture explicitly decomposes features into artifact, structure, and original visual branches, then adaptively fuses them over time via a learned gating module. This design allows the model to focus on perceptually relevant distortions while maintaining efficiency.

Experimental results demonstrate strong performance on short-form video datasets, achieving rank and linear correlation coefficients (SRCC: 0.736, PLCC: 0.787) while keeping inference runtime low. The method outperforms existing general video quality metrics on this challenging domain. By explicitly modeling frequency-domain priors, FGSVQA can better handle the rapid scene cuts, varying resolutions, and compression artifacts typical of user-generated short clips. The code and additional results are publicly available on GitHub, enabling reproducibility and further research. This work represents a practical step toward automated quality control for the exploding short-form video landscape.