# Sensor-Adaptive Infrared Spectral Reconstruction with Plug-and-Play Diffusion Priors

arXiv: https://arxiv.org/abs/2607.05636 · HTML: https://mlsanity.com/paper/2607.05636/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Inverse problems and image restoration, one of 14 directions in Diffusion models.
- Builds on 5 of its 20 references that are in the map.
- First posted on arXiv 6 Jul 2026.

## Placed in

- Diffusion models / [Inverse problems and image restoration](https://mlsanity.com/diffusion/inverse-problems-and-image-restoration/)

## Builds on (most cited first, 5 of 5)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [Denoising Diffusion Probabilistic Models](https://mlsanity.com/paper/2006.11239/) (2020)
- [Diffusion Models Beat GANs on Image Synthesis](https://mlsanity.com/paper/2105.05233/) (2021)
- [A survey on computational spectral reconstruction methods from RGB to hyperspectral imaging](https://mlsanity.com/paper/2106.15944/) (2021)

## Abstract (arXiv)

Hyperspectral sensing enables material identification; however, state-of-the-art spectrometers are costly and bulky, which limits their use in mobile applications. We address this by proposing sparse spectrum reconstruction from narrowband photocurrents using a pseudoinverse-guided diffusion model (PiGDM). With PiGDM we use a denoising diffusion probabilistic model (DDPM) to reconstruct the spectrum, which is trained on a large public spectral dataset to learn realistic spectral priors, eliminating the need for paired sensor measurements. At inference, PiGDM alternates reverse-diffusion denoising steps with pseudoinverse projection to enforce consistency with measured photocurrents via the calibrated responsivity matrices of sensors. Consequently, our method is sensor-adaptive: when detector arrays change, we simply substitute the responsivity matrix in the pseudoinverse projection without retraining of the diffusion model. The resulting computational spectrometer achieves 1.502% average estimation error, outperforming Tikhonov, Gaussian, compressive-sensing, and multilayer perceptron (MLP) baselines, while providing calibrated uncertainty estimates via Monte Carlo sampling from different random initializations of PiGDM. Summarizing, our approach offers an accurate, compact alternative for spectral recovery on resource-constrained platforms.