Research map / Image generation / Super-resolution
One-Step Effective Diffusion Network 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.
Proposes OSEDiff, a one-step Real-ISR method that feeds the LQ latent directly into LoRA-finetuned Stable Diffusion 2.1 and regularizes outputs with latent-space variational score distillation.
Problem
Real-world image super-resolution with pre-trained text-to-image diffusion models requires many sampling steps and random noise initialization, which is slow and introduces output uncertainty.
Approach
Use the LQ image latent directly as the diffusion starting point with a single denoising step (no noise), finetune the SD encoder and UNet with LoRA, and apply VSD in latent space with a trainable LoRA fake-score regularizer to align outputs with natural image distribution.
Key results
- OSEDiff-s1 achieves best LPIPS (0.2921), DISTS (0.2128) and FID (123.49) on RealSR among compared methods, with a single diffusion step.
- OSEDiff runs in 0.11 s per 512x512 image on an A100, about 105x faster than StableSR (200 steps) and 39x faster than SeeSR.
- OSEDiff has only 8.5M trainable parameters (LoRA layers), versus 749.9M for SeeSR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OSEDiff-s1 instead of SeeSR-s50 | Real-ISR · DrealSR | LPIPS: 0.3189 -> 0.2968 | Helped |
| Use OSEDiff-s1 instead of ResShift-s15 | Real-ISR · DrealSR | PSNR: 28.46 -> 27.92 | Hurt |
| Use OSEDiff instead of StableSR | Real-ISR inference · 512×512 input on A100 | inference time (s): 11.50 -> 0.11 | Helped |
| Remove VSD loss (ablation) | OSEDiff · RealSR | MUSIQ: 69.09 -> 65.23 | Hurt |
| Replace latent-space VSD with GAN loss (ablation) | OSEDiff · RealSR | LPIPS: 0.2921 -> 0.2760 | Mixed |
| Replace latent-space VSD with image-domain VSD (ablation) | OSEDiff · RealSR | LPIPS: 0.2921 -> 0.2759 | Mixed |
| Remove text prompts (ablation) | OSEDiff · DrealSR | LPIPS: 0.2968 -> 0.2896 | Helped |
| Remove text prompts (ablation) | OSEDiff · DrealSR | CLIPIQA: 0.6963 -> 0.6599 | Hurt |
| Replace DAPE prompts with LLaVA-v1.5 prompts (ablation) | OSEDiff · DrealSR | LPIPS: 0.2968 -> 0.3109 | Hurt |
| Increase VAE encoder LoRA rank from 4 to 8 (ablation) | OSEDiff · RealSR | PSNR: 25.15 -> 24.86 | Hurt |
| Set VAE encoder LoRA rank to 2 (ablation) | OSEDiff · RealSR | PSNR | Mixed |
| Increase UNet LoRA rank from 4 to 8 (ablation) | OSEDiff · RealSR | PSNR: 25.15 -> 24.87 | Hurt |
Surprising findings
- Removing text prompts improves full-reference metrics (PSNR, LPIPS, FID) while worsening no-reference metrics such as MUSIQ, MANIQA and CLIPIQA.
- Finetuning the VAE encoder is needed to remove degradation, but also finetuning the VAE decoder degrades perceptual quality (CLIPIQA 0.5778 vs 0.6693), so the decoder is kept frozen.
Limitations
- Limited in detail generation capability.
- Like other SD-based methods, struggles to reconstruct fine-scale structures such as small scene texts.
- Full-reference fidelity (PSNR/SSIM) trails GAN-based and ResShift-based methods.
Experimental setup
Datasets: LSDIR FFHQ DIV2K-Val RealSR DRealSR RealLR200
Benchmarks: DIV2K-Val DrealSR RealSR
Metrics: PSNR SSIM LPIPS DISTS FID NIQE MUSIQ MANIQA CLIPIQA inference time
Baselines: StableSR DiffBIR SeeSR PASD ResShift SinSR BSRGAN Real-ESRGAN LDL FeMaSR
Builds on: Stable Diffusion variational score distillation DMD LoRA SeeSR
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.91)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.91)
- Adversarial Diffusion Compression for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.91)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.91)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.91)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · 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.