Answers are grounded in this paper and cite their sources.
All-in-One video restoration under adverse weather conditions (e.g., rain, snow, haze) remains challenging, as existing methods either rely on explicit optical flow alignment that becomes unreliable under heavy degradations, or employ implicit sequence modeling that lacks fine-grained temporal correspondence. We propose FLAME, a unified FLow guidance and Adaptive frequency fusion for all-in-one Mamba-Enhanced video restoration framework that synergistically combines explicit alignment with implicit modeling through three core components: (i) Dual-Path Mamba Modeling with complementary Space-Time Sequence Mamba and Hilbert-Sequence Mamba blocks for efficient spatiotemporal modeling with linear-complexity Mamba scanning; (ii) Flow-guided Lightweight Attention featuring degradation-aware flow estimation with triple quality assessment and temporal-spatial attention deformable fusion; (iii) Adaptive Frequency Residual Fusion employing learnable frequency decomposition and cross-channel attention for degradation-specific spectral enhancement. Comprehensive experiments demonstrate state-of-the-art performance: 35.50 dB PSNR on VRDS (+3.46 dB over RainMamba), 40.17 dB on LWDDS (+2.96 dB over RainMamba), 31.04 dB on W3 multiweather benchmark (+0.59 dB over AIM-VR), and 29.00 dB on RainSynAll100 rain-haze mixture (+1.33 dB over BasicVSR++), validating the effectiveness of explicit-implicit synergy for diverse weather degradation restoration.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: