MG-SpaIR restores images without training data, beating Deep Image Prior

MG-SpaIR (Multi-grade Sparse-guided Implicit Representation) is a novel training-data-free image restoration framework introduced by Jianmin Liao, Lei Huang, and colleagues. Unlike conventional deep learning approaches that require large paired datasets, MG-SpaIR reconstructs a clean image from a single degraded observation—even when the corruption is a mixture of blur, downsampling, noise, and missing pixels. The method builds on implicit neural representations (INRs) but adds a multi-grade coarse-to-fine residual hierarchy that progressively refines reconstructions across resolution levels, mitigating spectral limitations and improving fidelity. To further suppress INR-induced artifacts (spurious high-frequency patterns), the authors introduce explicit sparse proximal regularization (e.g., ℓ₀-type) applied directly in the high-resolution image domain. The optimization is solved via a multi-grade proximal alternating scheme with proven convergence guarantees.

Experimental results on mixed-degradation benchmarks show MG-SpaIR consistently outperforms strong baselines like Deep Image Prior, while remaining stable, interpretable, and data-efficient. This approach is particularly valuable for scenarios where training data is scarce or unavailable, such as medical imaging, satellite imagery, or historical photo restoration. By eliminating the need for training data while rivaling learning-based methods, MG-SpaIR represents a significant step toward practical, deployable image restoration systems that require no prior dataset curation.