Research map / Image generation / Super-resolution
TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution
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 the SD3 text-to-image diffusion model into a one-step real-world super-resolution student using a Target Score Distillation loss and a Distribution-Aware Sampling Module.
Problem
Multi-step diffusion-based real-world image super-resolution is computationally expensive, and existing one-step distillation methods (e.g., OSEDiff with VSD) yield unreliable gradients and poor detail recovery.
Approach
Adds a Target Score Matching loss that aligns teacher score predictions on synthetic and HQ latents, combined with VSD into a Target Score Distillation loss, plus a sampling module that denoises along the diffusion trajectory to accumulate gradients at early timesteps for detail recovery.
Key results
- On DRealSR, TSD-SR one-step achieves LPIPS 0.2967, DISTS 0.2136, FID 134.98 and best MUSIQ 66.62 and CLIPIQA 0.7344 among one-step methods.
- Inference time of 0.1362 s per 512x512 image on A100, about 40x faster than SeeSR-50s and 120x faster than SUPIR.
- User study: 69.2% of 50 participants preferred TSD-SR outputs over other one-step methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TSD-SR instead of competing single-step methods | TSD-SR · DRealSR | DISTS: 0.2968 -> 0.2136 | Helped |
| Use TSD-SR instead of competing single-step methods | TSD-SR · RealSR | LPIPS: 0.2920 -> 0.2743 | Helped |
| Use TSD-SR instead of competing single-step methods | TSD-SR · DIV2K-Val | DISTS: 0.1975 -> 0.1821 | Helped |
| Add Target Score Matching (TSM) (ablation) | TSD-SR · DRealSR | DISTS: 0.2327 -> 0.2136 | Helped |
| Add Target Score Matching (TSM) (ablation) | TSD-SR · DRealSR | MUSIQ: 63.90 -> 66.62 | Helped |
| Add Target Score Matching (TSM) (ablation) | TSD-SR · DRealSR | MANIQA: 0.5749 -> 0.5874 | Helped |
| Add Target Score Matching (TSM) (ablation) | TSD-SR · DRealSR | CLIPIQA: 0.6958 -> 0.7344 | Helped |
| Add Distribution-Aware Sampling Module (DASM) (ablation) | TSD-SR · DRealSR | DISTS: 0.2311 -> 0.2136 | Helped |
| Add Distribution-Aware Sampling Module (DASM) (ablation) | TSD-SR · DRealSR | MUSIQ: 63.56 -> 66.62 | Helped |
| Add Distribution-Aware Sampling Module (DASM) (ablation) | TSD-SR · DRealSR | MANIQA: 0.5812 -> 0.5874 | Helped |
| Add Distribution-Aware Sampling Module (DASM) (ablation) | TSD-SR · DRealSR | CLIPIQA: 0.7123 -> 0.7344 | Helped |
Surprising findings
- PSNR and SSIM are lower for TSD-SR than for several baselines, which the authors attribute to reconstruction metrics being poorly suited to Real-ISR evaluation.
Limitations
- Large number of model parameters compared with GAN or non-diffusion approaches.
- Lower PSNR and SSIM than some baselines, reflecting a fidelity-realism trade-off.
Experimental setup
Datasets: DIV2K Flickr2K LSDIR FFHQ RealSR DRealSR DIV2K-Val
Benchmarks: DIV2K-Val RealSR DRealSR
Metrics: PSNR SSIM LPIPS DISTS FID NIQE MUSIQ MANIQA CLIPIQA
Baselines: OSEDiff AddSR SinSR StableSR DiffBIR SeeSR SUPIR PASD ResShift
Builds on: Variational Score Distillation OSEDiff ProlificDreamer Stable Diffusion 3
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.92)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.91)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.91)
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 2025 · similarity 0.91)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.91)
- RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic 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.