Research map / Image generation / Image restoration
PhoCoLens: Photorealistic and Consistent Reconstruction in Lensless Imaging
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.
Proposes PhoCoLens, a two-stage lensless reconstruction method: a spatially varying deconvolution network (SVDeconv) recovers range-space content for consistency, then a StableSR-style conditional diffusion model adds null-space high-frequency detail for photorealism.
Problem
Lensless imaging reconstruction must balance data consistency with photorealism, but existing methods assume a shift-invariant PSF (inaccurate at the field-of-view periphery) and generative restoration methods can alter scene content or hallucinate objects.
Approach
Decompose the reconstruction via range-null space decomposition: a spatially varying deconvolution with learnable multi-kernel PSFs (initialized from one calibrated PSF, interpolated by inverse-distance weights) recovers the range-space component, and a pre-trained Stable Diffusion with frozen weights, conditioned on that component and trained with a null-space loss, adds the null-space details.
Key results
- On PhlatCam, PhoCoLens reaches 22.07 PSNR, 0.601 SSIM, and 0.215 LPIPS, the best among compared methods.
- On DiffuserCam, PhoCoLens achieves the best SSIM (0.748) and LPIPS (0.161) while UPDN has slightly higher PSNR (24.67 vs 24.12).
- FlatNet+DiffBIR gets higher no-reference scores (e.g., MUSIQ 51.20 on DiffuserCam) but much lower full-reference fidelity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PhoCoLens instead of FlatNet-gen | lensless image reconstruction · PhlatCam | PSNR: 20.53 -> 22.07 | Helped |
| Use PhoCoLens instead of FlatNet+DiffBIR | lensless image reconstruction · PhlatCam | MUSIQ: 57.13 -> 62.20 | Helped |
| Use PhoCoLens instead of UPDN | lensless image reconstruction · DiffuserCam | SSIM: 0.747 -> 0.748 | Helped |
| Use PhoCoLens instead of UPDN | lensless image reconstruction · DiffuserCam | PSNR: 24.67 -> 24.12 | Hurt |
| Use PhoCoLens instead of FlatNet+DiffBIR | lensless image reconstruction · DiffuserCam | MUSIQ: 51.20 -> 32.84 | Hurt |
| Use spatially varying deconvolution rather than a single-kernel deconvolution (ablation) | SVDeconv · PhlatCam and DiffuserCam | reconstruction fidelity | Helped |
| Use spatially varying deconvolution rather than MultiWienerNet without weighted interpolation (ablation) | SVDeconv · PhlatCam and DiffuserCam | reconstruction fidelity | Helped |
| Condition null-space diffusion on reconstructed range-space content rather than WienerDeconv output (ablation) | null-space diffusion · PhlatCam and DiffuserCam | consistency | Helped |
| Condition null-space diffusion on reconstructed range-space content rather than FlatNet-gen output (ablation) | null-space diffusion · PhlatCam and DiffuserCam | consistency | Helped |
| Condition null-space diffusion on reconstructed range-space content rather than SVD-OC output (ablation) | null-space diffusion · PhlatCam and DiffuserCam | consistency | Helped |
| Use null-space diffusion rather than DiffBIR, StableSR, or DDNM (ablation) | range-space reconstruction · PhlatCam and DiffuserCam | consistency and visual quality | Helped |
Surprising findings
- FlatNet+DiffBIR achieves higher non-reference quality scores than PhoCoLens on DiffuserCam while losing substantial fidelity, showing the quality-consistency trade-off in generative restoration.
Limitations
- Two-stage pipeline and diffusion sampling time hinder real-time applicability.
- Diffusion may introduce high-frequency details that deviate from the original scene, especially for smooth scenes.
- Relies on a single calibrated PSF; range-space targets for training are simulated approximations rather than true transfer-matrix outputs.
Experimental setup
Datasets: PhlatCam dataset DiffuserCam dataset
Benchmarks: PhlatCam dataset DiffuserCam dataset
Metrics: PSNR SSIM LPIPS ManIQA ClipIQA MUSIQ
Baselines: WienerDeconv ADMM Le-ADMM-U MMCN UPDN FlatNet-gen DDNM+ FlatNet+DiffBIR
Builds on: StableSR DDNM DiffBIR FlatNet-gen MultiWienerNet
Closest papers by meaning
- A Physics-Informed Blur Learning Framework for Imaging Systems (CVPR 2025 · similarity 0.81)
- ReconFusion: 3D Reconstruction with Diffusion Priors (CVPR 2024 · similarity 0.81)
- Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.81)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.80)
- 3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction (NeurIPS 2025 · similarity 0.80)
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.80)
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.