Spectral Diffusion Prior (SiDP)

• SiDP is a learned diffusion prior that operates in a compact hyperspectral feature space to capture subtle high-frequency spectral details.
• It utilizes an HSI Feature Extractor to convert 2D measurements into low-dimensional feature vectors, enabling efficient diffusion-based denoising with a short horizon (T=4).
• SDP is injected into existing reconstruction networks via a plug-and-play Spectral Prior Injector Module, yielding improved PSNR and SSIM performance over baseline models.

Searching arXiv for the specified paper and closely related naming variants. SiDP, in the context of hyperspectral imaging, refers to what the underlying paper explicitly names the Spectral Diffusion Prior (SDP): a diffusion-model-based learned prior over compact hyperspectral spectral features, designed to be plugged into existing hyperspectral image (HSI) reconstruction networks in order to improve recovery of fine, high-frequency spectral details from degraded 2D measurements. Rather than generating full HSIs directly, the method learns a prior in a low-dimensional feature space and injects it into reconstruction backbones such as MST and BiSRNet through the Spectral Prior Injector Module (SPIM) (Yu et al., 18 Jul 2025).

1. Conceptual definition and reconstruction setting

HSI reconstruction seeks to recover a 3D hyperspectral image from a 2D coded measurement. In the formulation used here, the target HSI is $$\mathbf{G}\in\mathbb{R}^{H\times W\times N_\lambda}$$, and a measurement-derived initial 3D cube $\mathbf{H}$ is obtained by back-shifting the CASSI observation $\mathbf{Y}$ as

$$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$

where $d$ is the dispersion step size, $\lambda_n$ is the wavelength of the $n$-th channel, and $\lambda_c$ is a center wavelength. This $\mathbf{H}$ is an approximate, misaligned 3D cube used as the reconstruction starting point.

The central design choice is that the prior is not defined directly on the full HSI tensor. Instead, an HSI Feature Extractor (HFE) maps hyperspectral data to a compact global feature representation. During training, HFE takes the concatenation of $\mathbf{H}$ and the ground-truth HSI $\mathbf{H}$0, producing a global feature vector $\mathbf{H}$1. During deployment, HFE is applied to the measurement-derived cube alone, yielding $\mathbf{H}$2. The Spectral Diffusion Prior is therefore a learned prior over these compact spectral features, intended to encode global spectral information, spectral correlations, and high-frequency detail information in a form suitable for conditioning downstream reconstruction networks (Yu et al., 18 Jul 2025).

2. Diffusion-based prior learning in feature space

The prior is learned with a denoising diffusion probabilistic model applied to HSI feature vectors rather than images. Let $\mathbf{H}$3 denote the clean feature vector extracted by HFE. The forward diffusion process adds Gaussian noise according to

$\mathbf{H}$4

which can be rewritten in closed form as

$\mathbf{H}$5

with $\mathbf{H}$6 and $\mathbf{H}$7.

The reverse process denoises from $\mathbf{H}$8. A neural network $\mathbf{H}$9 is trained to predict the injected noise using the standard DDPM objective. For random $\mathbf{Y}$0, sample $\mathbf{Y}$1, define

$\mathbf{Y}$2

and optimize

$\mathbf{Y}$3

Two aspects are distinctive. First, the denoiser $\mathbf{Y}$4 is implemented as a simple MLP, because the diffusion process operates only in a low-dimensional feature space. Second, the diffusion horizon is deliberately short: the authors set $\mathbf{Y}$5. The paper attributes this to the fact that the diffusion process occurs only in a low-dimensional eigenspace, making a heavy image-space UNet unnecessary. This design suggests that the model is intended to capture the distribution of detailed spectral patterns with low computational overhead while remaining explicitly prior-like rather than fully generative at the image level (Yu et al., 18 Jul 2025).

3. Injection into reconstruction networks through SPIM

After learning the spectral diffusion model, the resulting prior is injected into a baseline HSI backbone through the Spectral Prior Injector Module (SPIM). Conceptually, the spectral diffusion model takes the measurement feature $\mathbf{Y}$6 together with its learned feature-space distribution and outputs the SDP vector, which encodes plausible spectral detail information consistent with the current measurement.

SPIM acts as a conditioning and modulation block on intermediate features. Let $\mathbf{Y}$7 denote an intermediate feature map in MST or BiSRNet, and let $\mathbf{Y}$8 denote the spectral prior vector. SPIM applies two learnable linear projections $\mathbf{Y}$9 and $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$0 to the prior and updates the feature map as

$$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$1

Here, $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$2 provides multiplicative channel-wise modulation, $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$3 provides additive channel-wise bias, and the residual term $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$4 preserves the original feature stream.

This module is explicitly plug-and-play. The paper inserts SPIM into MST and BiSRNet without redesigning those architectures. The role of SDP is therefore not to replace measurement-driven reconstruction, but to modulate internal representations so that the baseline network is guided toward sharper, more spectrally faithful reconstructions (Yu et al., 18 Jul 2025).

4. Two-stage optimization and inference workflow

Training proceeds in two stages. In Stage I, the system forms the initial cube $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$5 from the 2D measurement, concatenates $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$6 and $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$7, and uses HFE to learn a compact feature representation jointly with the base reconstruction network. The base model in this stage is MST-S or BiSRNet, trained with standard MSE supervision on HSI reconstruction. The stated goal is that $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$8 capture high-frequency details and correlation with measurements while remaining independent of spatial resolution.

In Stage II, the trained HFE is fixed or reused, and the spectral diffusion model (SDM) is trained on the HFE feature vectors $$\mathbf{H}(x,y,n_\lambda) = \mathbf{Y}(x, y - d(\lambda_n - \lambda_c)),$$9 using the DDPM objective in feature space. After SDM training, the full pipeline is fine-tuned: for each measurement, compute $d$0, extract $d$1, generate the SDP feature through SDM, inject SDP into the backbone through SPIM, and optimize the full system with MSE reconstruction loss.

Inference follows the same pipeline without access to $d$2: shift the measurement to obtain $d$3, compute $d$4, generate SDP, run the SDP-enhanced backbone, and output the reconstructed HSI. Because diffusion occurs only in feature space with $d$5 and an MLP denoiser, the added runtime overhead is described as modest. The training setup further specifies that when plugins are added, the learning rate of the base models is halved to avoid instability; Stage II uses 50 epochs total, with the first 5 epochs training the diffusion model only, Adam optimizer with $d$6 and $d$7, and learning rate $d$8 for the diffusion phase (Yu et al., 18 Jul 2025).

5. Experimental evaluation

Evaluation is conducted on two representative HSI reconstruction backbones, MST and BiSRNet, using PSNR and SSIM as metrics. Training uses the CAVE dataset with 32 HSIs, 31 spectral channels, and spatial size $d$9. Testing uses the KAIST dataset with 30 HSIs, 31 channels, and image size $\lambda_n$0, with 10 scenes used for simulation tests. The wavelength range is 450–650 nm via spectral interpolation, dispersion step is $\lambda_n$1, and data augmentation uses random flip and rotation.

The reported mean improvements are summarized below.

Backbone	Baseline	With SDP
MST-S	34.26 dB / 0.935	34.72 dB / 0.943
BiSRNet	29.76 dB / 0.833	30.31 dB / 0.856

These correspond to gains of +0.46 dB PSNR and +0.008 SSIM for MST-S, and +0.55 dB PSNR and +0.023 SSIM for BiSRNet. Scene-wise gains are reported as consistently positive, often greater than $\lambda_n$2 dB, and reaching as high as +1.57 dB PSNR in one BiSRNet-SDP case. The paper further states that SDP-enhanced models reconstruct richer textures, sharper edges, and more accurate fine structural details, while competing methods such as TwIST, GAP-TV, DeSCI, $\lambda_n$3-Net, ADMMNet, TSA-Net, DGSMP, MST-S, and BiSRNet more often exhibit blur or missing detail. The resulting systems, MST-S-SDP and BiSRNet-SDP, are described as achieving state-of-the-art or better performance with relatively small parameter counts of 1.30M and 0.06M, respectively (Yu et al., 18 Jul 2025).

Ablation on SPIM clarifies how the prior should be injected. For BiSRNet, the reported results are: no injection, 29.76 dB / 0.833; only multiplication, 29.49 dB / 0.820; only addition, 30.00 dB / 0.839; multiplication plus addition, 30.31 dB / 0.856. This establishes that additive-only conditioning is helpful, multiplicative-only conditioning can be harmful, and the full residual combination is best.

6. Interpretation, misconceptions, and terminological clarification

The stated motivation is that conventional deep HSI reconstruction networks often fit low-frequency structure and coarse spectra while missing subtle, high-frequency spectral details, especially under the severe compression of CASSI. SDP addresses this by learning a prior over feature vectors that summarize detailed spectral patterns. Because the prior is used as a conditioner rather than as a direct image generator, it can bias the reconstruction toward statistically plausible spectral detail without discarding measurement consistency. This suggests that the method occupies an intermediate position between purely discriminative reconstruction and full generative synthesis.

Several misconceptions are explicitly resolved by the paper. First, the diffusion model is not applied to hyperspectral images themselves; it is applied to low-dimensional HSI feature vectors. Second, the method does not require redesign of the host reconstruction network; its stated philosophy is plug-and-play integration through SPIM. Third, the acronym “SiDP” is not used in the paper. The paper consistently uses SDP, standing for Spectral Diffusion Prior. In this context, “SiDP” is therefore best understood as an informal or alternate reference to the same mechanism rather than as a distinct method name (Yu et al., 18 Jul 2025).