Research map / Image generation / Super-resolution
SinSR: Diffusion-Based Image Super-Resolution in a Single Step
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.
Distills a deterministic version of the ResShift diffusion super-resolution teacher into a one-step student network, adding a learned-inversion consistency loss that uses ground-truth HR images.
Problem
Diffusion-based super-resolution needs many inference steps (e.g., 15 for ResShift, 100 for LDM), which hinders practical use; prior acceleration either needs an explicit degradation model or still takes many steps.
Approach
Derive a deterministic (DDIM-like non-Markovian) sampler for ResShift to obtain well-matched noise-to-HR pairs, distill this xT-to-x0 mapping into a single-step student initialized from the teacher, and add a consistency-preserving loss that learns an inversion and regularizes with ground-truth HR images.
Key results
- SinSR-1 reaches CLIPIQA 0.6887 and MUSIQ 61.582 on RealSR, outperforming its 15-step ResShift teacher (0.5958, 59.873).
- On ImageNet-Test, SinSR-1 gets LPIPS 0.221 versus 0.228 for ResShift-15 with a single step, at 0.012s batch-64 runtime vs 0.047s for ResShift-15.
- Distillation-only student without the consistency loss is already competitive, e.g., CLIPIQA 0.6822 on RealSet65 vs 0.6537 for the teacher.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill ResShift into a one-step SinSR model | ResShift · RealSR | CLIPIQA: 0.5958 -> 0.6887 | Helped |
| distill ResShift into a one-step SinSR model | ResShift · RealSet65 | CLIPIQA: 0.6537 -> 0.7150 | Helped |
| distill ResShift into a one-step SinSR model | ResShift · ImageNet-Test | PSNR: 24.90 -> 24.56 | Hurt |
| distill ResShift into a one-step SinSR model | ResShift · ImageNet-Test | LPIPS: 0.228 -> 0.221 | Helped |
| distill ResShift into a one-step SinSR model | ResShift · ImageNet-Test | Runtime (bs=1): 0.633s -> 0.058s | Helped |
| add the consistency-preserving ground-truth loss (ablation) | SinSR-1 · RealSet65 | CLIPIQA: 0.6822 -> 0.7150 | Helped |
| replace deterministic sampling with default stochastic sampling (ablation) | SinSR distillation · RealSet65 | CLIPIQA: 0.6822 -> 0.4166 | Hurt |
| replace deterministic sampling with default stochastic sampling (ablation) | SinSR distillation · RealSet65 | MUSIQ: 61.27 -> 51.53 | Hurt |
| reduce model size from 118.59M to 24.32M parameters (ablation) | ResShift · RealSet65 | CLIPIQA: 0.6537 -> 0.5365 | Hurt |
| reduce model size to 24.32M parameters (ablation) | SinSR (distill only) · RealSet65 | CLIPIQA: 0.6822 -> 0.6499 | Hurt |
| replace the proposed distillation with Rectified Flow and iterative distillation (ablation) | ResShift · ImageNet-Test | LPIPS: 0.2221 -> 0.2322 | Hurt |
| reduce ResShift inference steps from 15 to 1 (ablation) | ResShift · ImageNet-Test | LPIPS: 0.228 -> 0.383 | Hurt |
Surprising findings
- Learning the deterministic xT-to-x0 mapping is easier than denoising at multiple noise levels: a 24.3M student trained this way beats a 24.3M ResShift on RealSet65 (CLIPIQA 0.6499 vs 0.5365).
- Iterative rectified-flow distillation degraded performance relative to the single-step distillation used here.
Limitations
- PSNR and SSIM drop slightly relative to the 15-step teacher (24.56 vs 24.90 PSNR on ImageNet-Test).
- Training still requires solving the ODE teacher sampler per iteration, adding training overhead.
- Evaluated only on ImageNet-based and real-world SR benchmarks with the ResShift backbone.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet-Test RealSR RealSet65
Metrics: PSNR SSIM LPIPS CLIPIQA MUSIQ runtime
Baselines: ResShift LDM ESRGAN BSRGAN SwinIR RealESRGAN DASR RealSR-JPEG Rectified Flow
Builds on: ResShift DDIM Rectified Flow LDM
Closest papers by meaning
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.94)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.93)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.93)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.92)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.92)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.91)
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.