Research map / Image generation / Image restoration
Filterless Snapshot Hyperspectral Imaging using Guided Patch Diffusion
Dean Hazineh, Luca Sacchi, Davide Cassara, Federico Capasso, Todd Zickler
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PSF-consistency guidance during sampling (ablation) | patch diffusion (64px) · ARAD1K filterless | PSNR: 32.32 -> 34.6 | Improved |
| add guidance with 32px patches (ablation) | patch diffusion (32px) · ARAD1K filterless | PSNR: 29.27 -> 33.27 | Improved |
| reduce patch size from 64 to 32 (equal to PSF width), with guidance (ablation) | patch diffusion · ARAD1K filterless | PSNR: 34.7 -> 33.3 | Hurt |
| use overlapping patches (stride 32) with guidance (ablation) | patch diffusion (64px) · ARAD1K filterless | PSNR: 34.7 -> 34.8 | No clear effect |
| use overlapping patches without guidance (ablation) | patch diffusion (64px) · ARAD1K filterless | PSNR: 32.2 -> 33.0 | Improved |
| compute per-patch scale by least squares during sampling instead of no rescaling (ablation) | patch diffusion (64px, guided) · ARAD1K filterless | PSNR: 31.8 -> 34.7 | Improved |
| use patch diffusion instead of full-field UNet regression (ablation) | UNet backbone · ARAD1K filterless | PSNR: 29.12 -> 34.6 | Improved |
| increase guidance iterations from 20 to 60 (ablation) | Ours · ARAD1K filterless | PSNR: 32.32 (none) -> 35.22 (60 iters) | Improved |
| use strongly dispersive PSF designs R2/R3 instead of R1 (ablation) | Ours · ARAD1K lens comparison | PSNR: 35.0 -> 34.0 (R2), 34.1 (R3) | Hurt |
| use chromatic-aberration PSF R1 instead of all-in-focus lens (ablation) | Ours · ARAD1K lens comparison | PSNR: 29.8 -> 35.0 | Improved |
| patch diffusion vs best full-field baseline (SST) | Ours · ARAD1K filterless | PSNR: 31.8 -> 34.6 | Improved |
| apply model trained on ARAD1K to ICVL/Harvard without finetuning | Ours · ICVL, Harvard | PSNR: 33.48 (ICVL), 32.37 (Harvard) | Mixed |
Closest papers by meaning
- Compact Low-Cost Hyperspectral Imaging via Angular-to-Spectral Diversity Conversion (ECCV 2026 · similarity 0.90)
- Unmixing Diffusion for Self-Supervised Hyperspectral Image Denoising (CVPR 2024 · similarity 0.87)
- DiffSCI: Zero-Shot Snapshot Compressive Imaging via Iterative Spectral Diffusion Model (CVPR 2024 · similarity 0.87)
- Spectral Compressive Imaging via Chromaticity-Intensity Decomposition (NeurIPS 2025 · similarity 0.86)
- Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification (CVPR 2024 · similarity 0.86)
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.85)
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.
