Recent SISR methods suffer from unstable feature propagation and large model sizes. While consistent information flow stabilizes activation dynamics, current approaches struggle to recover high-frequency details crucial for perceptual quality.
We propose SimFlowSR, comprising two complementary branches: CEB stabilizes dynamics through dense-residual connections, while GAB recovers fine structures via parameter-free geometric transformations (dihedral group D₄). This achieves globally consistent yet detail-rich reconstruction with minimal overhead.
When integrated into SwinIR and MambaIR, SimFlowSR improves quality with up to 37% fewer parameters, demonstrating superior performance across standard benchmarks.
Despite DRCT's competitive performance through consistent information flow, it exhibits limited high-frequency capture. Training dynamics on DF2K show DRCT achieves stable entropy but suboptimal frequency ratio—revealing the gap: consistent flow ensures stability, but lacks mechanisms for high-frequency recovery.
SimFlowSR achieves both highest frequency ratio and stable entropy, validating that CEB + GAB enables stable yet detail-rich reconstruction under low complexity.
Natural images exhibit geometric self-similarity—structural patterns recur under rotation, flipping, scaling. This fractal property manifests across diverse scenes (brick walls, leaves, snowflakes), enabling high-frequency recovery without heavy computation.
GAB explicitly models self-similarity via parameter-free transformations. Unlike methods searching patches without geometric modeling (a), GAB applies dihedral group D₄ transformations (identity, 90°/180°/270° rotations, flips) to construct candidates (b). Initially dissimilar patches align after transformation, revealing latent patterns. Top candidates are selected (cosine similarity), inverse-aligned (T⁻¹), and aggregated (adaptive attention)—no learnable parameters required.
SimFlowSR employs a modular dual-branch architecture seamlessly integrated into backbones (SwinIR, MambaIR, RWKVIR). CEB maintains consistent flow via dense-residual connections; GAB aggregates self-similar features via D₄ transformations. This plug-and-play design enhances quality across architectures while reducing complexity by up to 37%.
SimFlowSR achieves the most stable feature propagation. Conventional methods (SwinIR-RSTB, HAT-RHAG) show dramatic activation fluctuations across depths, indicating unstable information flow. DRCT (RDG) improves stability through dense connections, but SimFlowSR further compresses the dynamic range with tighter activation clustering—validating that CEB + GAB maintains stable representations while preserving fine details.
LAM showing superior spatial aggregation with higher Diffusion Index.
ERF demonstrating substantially broader spatial coverage.
SimFlowSR captures long-range correspondences effectively. LAM visualization shows significantly higher Diffusion Index, while ERF demonstrates broader spatial coverage across all backbones—confirming GAB's multi-scale D₄ transformations consistently enhance spatial modeling capability.
SimFlowSR: Self-similarity Aggregation over Consistent Information Flow for Single Image Super-Resolution
@article{lee2025simflowsr,
title={SimFlowSR: Self-similarity Aggregation over Consistent Information Flow for Single Image Super-Resolution},
author={Lee, Chia-Ming and Hsu, Chih-Chung},
journal={arXiv preprint arXiv:XXXX.XXXXX},
year={2025}
}