Existing Transformer-based deraining methods inadequately model the physical divergence between rain artifacts and background textures, causing edge blurring and micro-texture loss from indiscriminate spectral suppression. Motivated by the critical observation that rain degradation predominantly concentrates in medium-high frequency magnitude bands while phase spectra preserve structural integrity, we propose DSCformer—a Dynamic Spectrum Coordination Transformer synergizing global frequency decoupling with local geometric refinement. To address conventional methods’ failure in separating valid high-frequency textures, we design a Frequency-guided Multi-scale Attention that suppresses rain-dominated bands via energy-weighted gating while adaptively enhancing critical components through multi-scale spectral interactions. For static parameterization limitations under varying rain densities, we develop a Frequency Dynamic Modulated Feed-Forward Network, generating spatially-channel-adaptive coefficients via lightweight networks to contextually adapt spectral degradation. Further, leveraging wavelet transforms’ directional sensitivity, we introduce a Multi-level Frequency Enhancement Module performing orientation-aware decomposition with learnable wavelet bases and direction-specific convolutions, resolving artifact-detail trade-offs via physics-informed fusion. Experiments confirm superiority over state-of-the-art methods.