AnyBand-Diff: A Unified Remote Sensing Image Generation and Band Repair Framework with Spectral Priors

2026-05-15 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Remote Sensing & Earth Observation · Depth: Expert, extended

Summary

AnyBand-Diff is a novel spectral-prior-guided diffusion framework designed for generating and repairing remote sensing (RS) imagery while adhering to physical laws. Existing diffusion models often produce spectral distortion and radiometric inconsistency in RS data, limiting their scientific utility. AnyBand-Diff addresses this by integrating a Masked Conditional Diffusion (MCD) backbone with a dual stochastic masking strategy to recover complete spectral information from arbitrary band subsets. It also incorporates a Physics-Guided Sampling (PGS) mechanism, which uses gradients from a differentiable physical model to steer the denoising process towards physically plausible solutions. Furthermore, a Multi-Scale Physical Loss (MSPL) enforces radiometric fidelity across pixel, region, and global levels. Experiments show AnyBand-Diff outperforms state-of-the-art baselines, achieving a PSNR of 17.11 dB and SSIM of 0.73, and maintaining high fidelity even with 50% band loss, demonstrating its effectiveness in physics-aware generative methods for Earth observation.

Key takeaway

For Computer Vision Engineers developing generative AI for remote sensing, AnyBand-Diff demonstrates that integrating explicit physical priors is crucial for producing scientifically valid and radiometrically consistent imagery. You should consider incorporating differentiable physical models and multi-scale loss functions into your diffusion frameworks to overcome limitations of purely data-driven approaches, especially when dealing with incomplete or diverse sensor data. This approach ensures generated data is suitable for quantitative analysis, not just visual plausibility.

Key insights

AnyBand-Diff generates physically consistent remote sensing images by integrating spectral priors and multi-scale physical constraints into a diffusion model.

Principles

• Physical laws must guide generative models for scientific data.
• Robustness requires training with diverse, incomplete input conditions.
• Multi-scale constraints ensure hierarchical physical consistency.

Method

AnyBand-Diff uses a Masked Conditional Diffusion backbone with dual stochastic masking, Physics-Guided Sampling via differentiable physical models, and a Multi-Scale Physical Loss for pixel, region, and global consistency.

In practice

• Use differentiable physical models to guide diffusion sampling.
• Employ stochastic masking to train for arbitrary data loss.
• Apply multi-scale loss functions for hierarchical fidelity.

Topics

• Remote Sensing Image Generation
• Diffusion Models
• Spectral Reconstruction
• Physics-Guided Sampling
• Multi-Scale Physical Loss

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• D2-CDIG: Controlled Diffusion Remote Sensing Image Generation with Dual Priors of DEM and Cloud-Fog — D2-CDIG uses dual DEM and cloud-fog priors with a diffusion model for precise, natural remote sensing image generation.
• Adaptive Band Selection for Hyperspectral Classification with Spatially Disjoint Evaluation — Combining supervised spectral ranking with trainable Hard-Concrete gates improves hyperspectral band selection and classification accuracy.
• Spectral Progressive Diffusion for Efficient Image and Video Generation — Spectral Progressive Diffusion accelerates image/video generation by progressively increasing resolution based on frequency domain content.
• SatSplatDiff: Geometry-preserving generative refinement for high-fidelity satellite Gaussian Splatting — SatSplatDiff integrates monocular depth and shadow-guided generative refinement to achieve geometry-preserving, high-fidelity satellite 3D reconstruction.
• A Scale-Adaptive Framework for Joint Spatiotemporal Super-Resolution with Diffusion Models — A scale-adaptive framework reuses one architecture for spatiotemporal super-resolution across diverse scaling factors.
• DiffuSAM: Diffusion Guided Zero-Shot Object Grounding for Remote Sensing Imagery — DiffuSAM combines diffusion models with segmentation models for superior zero-shot object grounding in remote sensing.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.