Reproduce “R2H-Diff: Guided Spectral Diffusion Model for RGB-to-Hyperspectral Reconstruction” on OpenResearch

About this paper

Songyu Ding, Ronggiang Zhao, Mingchun Sun, Jie Liu

R2H-Diff introduces a guided spectral diffusion model for reconstructing high-fidelity hyperspectral images from standard RGB inputs by reformulating the task as a conditional iterative refinement. The approach delivers state-of-the-art or competitive performance across multiple benchmarks, while significantly reducing model complexity, achieving up to 64% fewer parameters and 47% fewer FLOPs than previous methods on the NTIRE2022 dataset.

Key insights

• Reformulating RGB-to-HSI as an RGB-guided conditional iterative diffusion process effectively addresses the ill-posed nature of the problem and allows for better modeling of reconstruction uncertainty.
• Specialized architectural components, such as Hyperspectral-Adaptive Transposed Attention (HATA), are crucial for efficiently modeling high-dimensional channel-wise spectral dependencies and local spatial cues without excessive computational cost.
• A combined loss function incorporating both reconstruction and gradient consistency loss is essential for preserving high-frequency structural details and preventing over-smoothing in the reconstructed hyperspectral images.

Summary by alphaXiv.