film-conditioned spectral reconstruction

This work addresses the challenge of directly transferring pretrained RGB vision models to hyperspectral image analysis, where a fundamental mismatch exists between the three-channel input assumption and the high-dimensional spectral nature of hyperspectral data. To overcome this, the authors propose a novel partially trainable tensor decomposition strategy that decouples pretrained convolutional kernels into spatial and spectral components. The original three-channel spectral part is replaced with a high-dimensional, learnable spectral component, thereby constructing new filters tailored for hyperspectral inputs. This approach uniquely integrates trainable tensor decomposition into transfer learning, preserving the powerful spatial feature extraction capabilities of the original model while effectively modeling hyperspectral characteristics. Extensive experiments demonstrate that the proposed method significantly outperforms existing transfer learning approaches across multiple hyperspectral datasets, achieving both higher accuracy and improved robustness.

This work addresses the instability of reconstruction in ill-posed inverse problems caused by noise, as well as limitations of conventional regularization methods—such as reliance on manual hyperparameter tuning—and the lack of interpretability and cross-resolution generalization in deep learning approaches. To this end, we propose the Spectral Correction Network (SC-Net), which learns a signal-to-noise-ratio-adaptive, pointwise filtering function in the spectral domain of the forward operator to reweight spectral coefficients, yielding stable and interpretable solutions. Our method uniquely integrates interpretable adaptive spectral filtering with operator learning, and we theoretically prove that it can approximate continuous inverse operators, possesses discretization invariance, and achieves minimax optimal convergence rates. Experiments on 1D integral equations demonstrate a convergence rate of $O(\delta^{0.5})$, matching theoretical optimality; the learned filter outperforms Oracle Tikhonov regularization and generalizes zero-shot from $N=256$ to $N=2048$ with reconstruction error maintained at approximately 0.23.

This study addresses the prohibitive computational cost of domain optimization based on spectral functionals of differential operators, which typically relies on expensive PDE solvers. To overcome this limitation, we propose two neural network surrogate models that integrate Fourier coefficient encoding, landscape function representations, and Gram-Schmidt orthogonalization to directly learn spectra from geometric descriptions, enabling efficient optimization of star-shaped and landscape domains. Furthermore, coefficient scaling is introduced to satisfy eigenvalue scaling laws, while output averaging ensures rotational and reflectional invariance. The proposed approach achieves 0.2% accuracy for star-shaped domains with a 1% error across the first ten eigenvalues, significantly outperforming Fourier Neural Operators (FNO). By successfully reproducing classical spectral optima, this work establishes a novel paradigm for efficient spectral shape optimization.

This study investigates the stability of spectral methods for coefficient estimation in supervised regression under additive label noise. Focusing on multi-basis sparse spectral representations, the authors derive a closed-form expression for the overlap between noisy and noise-free coefficient vectors via eigen-geometric whitening. Their analysis reveals that the degradation in spectral learning performance is governed by a single intrinsic noise scale and establishes a critical noise threshold beyond which the underlying function structure becomes irrecoverable. The theoretical framework applies to various orthogonal bases—including Fourier, Legendre, Bessel, and Haar—and is corroborated by numerical experiments demonstrating that spectral coefficient estimates become significantly unstable once the noise level exceeds this threshold, thereby impeding accurate recovery of the true function.

This work addresses the unsupervised disentangled modeling of hyperspectral data. We propose PhISM, a physics-informed deep learning framework that requires no labeled data. PhISM explicitly disentangles abundance and endmember representations in latent space by embedding spectral physical priors—namely, endmember spectral continuity and the linear mixing assumption—and parameterizes endmember spectra using differentiable continuous basis functions (e.g., splines), enabling structured and interpretable modeling. By tightly integrating domain-specific physical constraints with deep representation learning, PhISM achieves a favorable trade-off between model interpretability and generalization capability. Extensive experiments on multiple hyperspectral unmixing and classification benchmarks demonstrate that PhISM significantly outperforms existing unsupervised and weakly supervised methods, markedly reducing reliance on annotated data while delivering physically consistent and semantically meaningful representations.

This work addresses the challenge of hyperspectral image restoration, which is hindered by data scarcity, sensor-specific characteristics, and the high dimensionality of spectral information, making it difficult to learn robust priors. The authors propose a lightweight transfer framework that projects hyperspectral data into a low-dimensional subspace, leverages a frozen pre-trained RGB denoiser for noise removal, and reconstructs the hyperspectral cube through a lightweight adapter coupled with constrained linear aggregation. This approach is the first to efficiently transfer large-scale RGB image priors to hyperspectral restoration tasks, achieving plug-and-play performance with minimal training. It consistently outperforms specialized hyperspectral methods across multiple datasets in denoising, deblurring, and super-resolution, demonstrating the remarkable transferability of RGB-based priors.

This study addresses the inefficiency of frequency composition and optimization difficulties inherent in implicit neural representations by proposing a learnable spectral activation method. Specifically, this approach replaces fixed nonlinear functions with residual truncated Fourier series to decouple feature selection from spectral shaping. By separating the gradients of linear weights and activation coefficients, it directly optimizes the spectral shape, thereby enhancing the energy concentration of the principal modes within the neural tangent kernel. Integrated with differentiable programming techniques, the proposed method significantly improves both reconstruction quality and optimization efficiency across diverse tasks, including audio, image, and neural radiance field synthesis.

Existing methods for non-rigid 3D shape correspondence often struggle to balance accuracy and efficiency due to their neglect of spectral basis optimization and reliance on computationally expensive solvers. This work proposes the Advanced Functional Maps framework, which, for the first time, enables end-to-end unsupervised learning of spectral basis optimization and reveals its equivalence to spectral convolution. By introducing a learnable suppression function that jointly optimizes feature representations and the spectral basis, and integrating a heat diffusion module with an unsupervised loss, the method constructs a lightweight architecture that avoids complex solvers or auxiliary losses. The approach significantly outperforms existing techniques under challenging conditions such as non-isometric deformations and topological noise, while maintaining high computational efficiency.

This work addresses the inverse design of multilayer optical thin films, where discrete material selection and continuous thickness optimization are tightly coupled. To tackle this challenge, the authors propose PRISM—a unified decoder-only autoregressive Transformer model that jointly predicts material categories and layer thicknesses through a single backbone network. The method innovatively incorporates spectral prefix conditioning to inject target spectral information and encodes cumulative depth into rotary positional embeddings to accurately capture the physical ordering of layers. The PRISM-13M variant achieves over a 50% reduction in mean absolute error (MAE) with only one-fifth the parameters of prior approaches, while PRISM-44M attains state-of-the-art performance on the in-distribution validation set (MAE = 0.010) and demonstrates significantly faster inference than simulated annealing algorithms.

This work addresses the limitations of existing spectral compressive imaging methods, which are predominantly confined to single-frame reconstruction and struggle to recover spatial-spectral information occluded by coding masks while lacking temporal consistency across video sequences. To overcome these challenges, we introduce DynaSpec, the first dynamic hyperspectral video dataset, and propose the Propagation-Guided Spectral Video Reconstruction Transformer (PG-SVRT). By leveraging a spatio-temporal attention mechanism and a novel bridging token design, PG-SVRT effectively integrates complementary information from adjacent frames. This approach significantly enhances reconstruction quality, spectral fidelity, and temporal consistency, all while maintaining extremely low computational overhead (FLOPs), thereby establishing a new benchmark for video-level spectral compressive imaging.