Research map / Image generation / Fast diffusion sampling
Bigger Isn’t Always Better: Towards a General Prior for Medical Image Reconstruction
Lukas Glaszner, Martin Zach, Thomas Pock
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 the self-attention block (d=4* vs d=4) (ablation) | score model · CORPD 4x Gaussian 1D, PC sampling | PSNR: 32.52 -> 32.81 | Hurt |
| shrink network to d=1 (smallest receptive field) (ablation) | score model, fastMRI-trained · CORPDFS (OOD) 4x Gaussian 1D, PC | PSNR: 31.08 -> 31.70 | Improved |
| shrink network to d=1 (ablation) | score model, fastMRI-trained · CORPD (ID) 4x Gaussian 1D, PC | PSNR: 32.52 -> 32.19 | Mixed |
| shrink network to d=1 (ablation) | score model, fastMRI-trained · CORPDFS (OOD) 8x Gaussian 1D, PC | PSNR: 29.52 -> 30.24 | Improved |
| train on natural images (CelebA-HQ) instead of fastMRI (ablation) | score model d=1 · CORPDFS (OOD) 4x Gaussian 1D, PC | PSNR: 31.70 -> 31.86 | Improved |
| train on natural images (CelebA-HQ) instead of fastMRI (ablation) | score model d=1 · CORPD (ID) 4x Gaussian 1D, PC | PSNR: 32.19 -> 32.61 | Improved |
| train on CelebA-HQ instead of fastMRI (ablation) | score model d=4* · Brain MRI (OOD) 4x Gaussian 1D, PC | PSNR: 32.69 -> 32.10 | Hurt |
| shrink network to d=1 with Poisson 15x mask, CelebA-trained (ablation) | score model · CORPD Poisson 15x, PC | PSNR: 30.37 -> 28.62 | Hurt |
| shrink network on CT (d=4* to d=1) (ablation) | score model, CT-trained · Thorax CT (ID) 60 views | PSNR: 36.29 -> 36.14 | No clear effect |
| use diffusion prior instead of TV (ablation) | score model d=4* · Thorax CT 60 views | PSNR: 23.21 -> 36.29 | Improved |
| reduce depth from d=4* to d=1 (inference cost) (ablation) | score model, PC sampling · reconstruction time per slice | seconds: 84.4 -> 39.23 | Improved |
| lower data-consistency weight lambda (near 0) for CelebA-trained networks (ablation) | score model PC sampling · CORPD slice | qualitative | Hurt |
Closest papers by meaning
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.90)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.88)
- DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction (NeurIPS 2024 · similarity 0.88)
- Solving Inverse Problems with Latent Diffusion Models via Hard Data Consistency (ICLR 2024 · similarity 0.88)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
- Learning Image Priors Through Patch-Based Diffusion Models for Solving Inverse Problems (NeurIPS 2024 · similarity 0.87)
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.
