Research map / Image generation / Image restoration
DiffSCI: Zero-Shot Snapshot Compressive Imaging via Iterative Spectral Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Integrates an RGB-pretrained diffusion model as a generative denoiser in a plug-and-play HQS framework for zero-shot multispectral snapshot compressive imaging reconstruction, with spectral-correlation input, wavelength matching, and a residual-accumulating accelerated data step.
Problem
Snapshot compressive imaging (CASSI) reconstruction of multispectral images is ill-posed and existing methods struggle in dark regions and sharp edges; supervised end-to-end models generalize poorly to real data, and diffusion models are trained on RGB rather than MSI.
Approach
Plug a pretrained RGB diffusion model into the PnP-HQS framework as the denoiser, adapting it to MSIs via adjacent-band spectral correlation input, wavelength matching of RGB-like band triplets, and a Nesterov-style accumulated-residual update for the data subproblem.
Key results
- On 10 KAIST simulation scenes, DiffSCI averages 35.28 PSNR / 0.916 SSIM, above zero-shot baselines PnP-CASSI (28.55) and DIP-HSI (29.75) and HLRTF (33.69).
- DiffSCI's average 35.28 dB is below supervised transformers MST++ (35.72) and CST-L+ (35.71) but exceeds several supervised CNN and unfolding methods such as TSA-Net (31.35) and HDNet (34.66).
- Residual accumulation acceleration yields about 5-6 dB average PSNR improvement at fixed sampling steps (Fig. 4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DiffSCI instead of CST-L+ | KAIST simulation reconstruction · KAIST simulation (10-scene Avg) | PSNR: 35.71 -> 35.28 | Hurt |
| apply DiffSCI instead of CST-L+ | KAIST simulation reconstruction · KAIST simulation (10-scene Avg) | SSIM: 0.960 -> 0.916 | Hurt |
| add residual-accumulation acceleration (ablation) | DiffSCI · KAIST Scene 5, 100 sampling steps | PSNR: 128.03 -> 130.15 | Helped |
| add residual-accumulation acceleration (ablation) | DiffSCI · KAIST Scene 5, 200 sampling steps | PSNR: 250.83 -> 251.98 | Helped |
| add residual-accumulation acceleration (ablation) | DiffSCI · KAIST Scene 5, 300 sampling steps | PSNR: 395.69 -> 372.77 | Hurt |
| add residual-accumulation acceleration (ablation) | DiffSCI · KAIST Scene 5, 500 sampling steps | PSNR: 899.20 -> 721.50 | Hurt |
| vary the number of diffusion sampling steps (ablation) | DiffSCI · KAIST Scene 5 | PSNR: 31.04; 128.03; 250.83; 395.69; 899.20 (unaccelerated); 30.36; 130.15; 251.98; 372.77; 721.50 (accelerated) | Mixed |
Surprising findings
- Reported average PSNR (35.28) of DiffSCI is slightly below some supervised transformer methods on simulation data, contrary to the claim of surpassing supervised transformers in the abstract.
- Accelerated data-step gives large PSNR gains at 20 sampling steps (31.04 vs 30.36 in the table, with the accelerated version reported as better), and time per step grows with steps.
Limitations
- Sampling with diffusion is time-consuming, scaling with number of bands and sampling steps.
- Hyperparameters (lambda, zeta, sc, tstart) tuned empirically on test scenes.
- Simulation results do not surpass the best supervised transformers on average PSNR.
Experimental setup
Datasets: KAIST (10 simulation scenes, 256x256, 28 bands) real CASSI measurements (5 scenes, 660x660, 28 bands)
Benchmarks: KAIST simulation scenes (S1-S10) real CASSI dataset
Metrics: PSNR SSIM
Baselines: TwIST GAP-TV DeSCI λ-Net TSA-Net HDNet MST-L MST++ CST-L+ DGSMP ADMM-Net GAP-Net
Builds on: Plug-and-Play (PnP) framework half-quadratic splitting DDIM PnP-CASSI DIP-HSI
Closest papers by meaning
- Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 2025 · similarity 0.87)
- Diffusion State-Guided Projected Gradient for Inverse Problems (ICLR 2025 · similarity 0.87)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.87)
- Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising (CVPR 2024 · similarity 0.87)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.86)
- Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.