From Dual-Stream Manifold-Aligned Tokenizer to Latent Sensor-Invariant Consistency Model: Rapid Generative Super-Resolution for Cross-Sensor Remote Sensing Imagery

Diffusion models have established state-of-the-art visual realism in remote sensing Super-Resolution (SR) but suffer from high inference latency. Although accelerated approaches such as latent consistency models (LCMs) enable efficient one-step sampling in a compressed space, their effectiveness in cross-sensor scenarios is hindered by latent manifold misalignment. This limitation arises from the magnified distributional discrepancy within the compressed latent space, which undermines the stability of consistency training and hinders convergence on heterogeneous data. To address this challenge, we propose a unified framework that explicitly decouples cross-sensor SR into sensor-invariant manifold learning and single-step latent residual generation. First, to effectively bridge the domain gap, we introduce the Dual-Stream Sinkhorn-Contrastive Masked Autoencoder (DSSC-MAE). Functioning as a domain-aligned tokenizer, it employs a novel Hierarchical Alignment Strategy that integrates reconstruction objectives with global Sinkhorn optimal transport and local patch-wise contrastive constraints. This design aligns heterogeneous distributions within a unified, structure-preserving latent manifold, effectively mitigating sensor-specific discrepancies. Second, building upon this aligned latent space, we propose the Latent Sensor-Invariant Residual Consistency Model (LSIRCM). Leveraging the sensor-invariant latent representation as a structural anchor, LSIRCM reformulates SR as a controllable latent residual generation task. By incorporating cross-attention mechanisms to integrate this aligned context, it utilizes an Anchor Diffusion Transformer (AnchorDiT) backbone with a hybrid consistency training strategy to condense the iterative sampling trajectory into a deterministic single-step mapping. Extensive experiments on public and newly constructed datasets demonstrate that our approach achieves a superior efficiency-quality trade-off. It surpasses the perceptual quality of state-of-the-art diffusion models while maintaining high structural fidelity, matching the inference speed of regression-based networks. Consequently, our framework offers a robust and computationally efficient solution for cross-sensor remote sensing SR tasks. The code will be publicly available at https://github.com/afomyj/LSIRCM.