Research map / Image generation / Image restoration
Reconstruct Anything Model a lightweight foundation model for computational imaging
Matthieu Terris, Samuel Hurault, Maxime Song, Julián Tachella
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 proximal estimation input block (ablation) | DRUNet backbone · Training tasks (10^4 iterations) | PSNR (dB): 25.83 -> 26.64 | Improved |
| add conditioning of inner feature maps on y (ablation) | DRUNet + prox · Training tasks (10^4 iterations) | PSNR (dB): 26.64 -> 26.71 | Improved |
| add Krylov subspace blocks applying A^T A to features (ablation) | DRUNet + prox + embed y · Training tasks (10^4 iterations) | PSNR (dB): 26.71 -> 27.61 | Improved |
| train on all three tasks instead of inpainting only (evaluated on inpainting) (ablation) | RAM-style model · Validation, inpainting | PSNR (dB): 30.84 -> 30.73 | No clear effect |
| train on all three tasks instead of deblurring only (evaluated on deblurring) (ablation) | RAM-style model · Validation, deblurring | PSNR (dB): 25.62 -> 25.58 | No clear effect |
| add proposed operator conditioning to DRUNet backbone (ablation) | DRUNet · CBSD68 motion blur sigma=0.05 | PSNR (dB): 22.46 -> 26.47 | Improved |
| swap DRUNet backbone for PromptIR backbone (ablation) | unconditioned backbone · CBSD68 motion blur sigma=0.05 | PSNR (dB): 22.46 -> 22.42 | No clear effect |
| use RAM instead of DRUNet for self-supervised finetuning on one image | self-supervised finetuning · Sentinel-2 compressed sensing, N=1 | PSNR (dB): 19.44 -> 30.40 | Improved |
| increase finetuning set from N=1 to N=100 images | RAM self-supervised · Sentinel-2 compressed sensing | PSNR (dB): 30.40 -> 33.57 | Improved |
| self-supervised instead of supervised finetuning (N=1) | RAM · Sentinel-2 compressed sensing | PSNR (dB): 30.77 -> 30.40 | No clear effect |
| tie weights across unrolled iterations instead of untied | uDPIR · fastMRI multi-coil x8 | PSNR (dB): 33.71 -> 36.06 | Improved |
| use RAM zero-shot instead of UNet on multi-coil MRI | UNet baseline · fastMRI multi-coil x8 | PSNR (dB): 28.50 -> 35.62 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Restormer — Restormer: Efficient Transformer for High-Resolution Image Restoration (CVPR 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.87)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
- Self-diffusion for Solving Inverse Problems (NeurIPS 2025 · similarity 0.86)
- Fast samplers for Inverse Problems in Iterative Refinement models (NeurIPS 2024 · similarity 0.86)
- A Diffusion Model with State Estimation for Degradation-Blind Inverse Imaging (AAAI 2024 · similarity 0.86)
- NPN: Non-Linear Projections of the Null-Space for Imaging Inverse Problems (NeurIPS 2025 · 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.
