Research map / Image generation / Super-resolution
Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator
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 D3SR, a one-step diffusion Real-ISR model that uses a large SDXL-based latent diffusion discriminator for adversarial training plus an edge-aware DISTS perceptual loss, instead of a multi-step teacher.
Problem
One-step diffusion distillation for real-world image super-resolution is limited by the capacity of the multi-step teacher (KD or VSD), which caps student quality and can inject unrealistic textures.
Approach
Replace the multi-step teacher with a large pretrained diffusion model (SDXL UNet mid-block features plus MLP) used as a latent-space discriminator on noised features at random timesteps, and add an edge-aware DISTS loss that also compares Sobel edge maps.
Key results
- On RealSR (x4), D3SR achieves the best NIQE (3.747), MUSIQ (68.69) and MANIQA (0.6436) among compared methods, with LPIPS 0.3150 and DISTS 0.1906.
- Inference takes 0.11 s versus 0.35 s for OSEDiff on an A100 at 512x512, with 966.3M total parameters vs 1.4B for OSEDiff.
- Replacing an SD 2.1 discriminator with SDXL improves LPIPS from 0.3166 to 0.3150 and MUSIQ from 68.08 to 68.69 on the ablation set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EA-DISTS instead of MSE-only spatial loss (ablation) | D³SR · RealSR ×4 | LPIPS: 0.3626 -> 0.3150 | Helped |
| Add edge-aware terms to DISTS (ablation) | D³SR · RealSR ×4 | MUSIQ: 67.55 -> 68.69 | Helped |
| Add a discriminator rather than training without one (ablation) | D³SR · RealSR ×4 | NIQE: 6.962 -> 3.747 | Helped |
| Use a pretrained SD 2.1 discriminator rather than a Diffusion-GAN discriminator (ablation) | D³SR · RealSR ×4 | MUSIQ: 67.51 -> 68.08 | Helped |
| Scale the discriminator from SD 2.1 to SDXL (ablation) | D³SR · RealSR ×4 | MUSIQ: 68.08 -> 68.69 | Helped |
| Train on the collected dataset rather than public datasets | D³SR · RealSR ×4 | NIQE: 3.899 -> 3.747 | Helped |
| Use D³SR instead of OSEDiff | one-step Real-ISR methods · RealSR ×4 | Inference time (s): 0.35 -> 0.11 | Helped |
Surprising findings
- Reference-based metrics like PSNR and SSIM are argued not to reflect diffusion-based Real-ISR quality well, and GAN-based methods get higher PSNR/SSIM despite lower perceptual quality.
Limitations
- PSNR and SSIM are lower than several multi-step and GAN baselines.
- The paper states the discriminator's scaling benefit is shown only for SD 2.1 vs SDXL, limited scale comparisons.
Experimental setup
Datasets: LSDIR FFHQ DIV2K-val RealSR RealSet65
Benchmarks: RealSR DIV2K-val RealSet65
Metrics: PSNR SSIM LPIPS DISTS NIQE MUSIQ MANIQA
Baselines: StableSR DiffBIR SeeSR ResShift SinSR OSEDiff AddSR Diffusion-GAN Real-ESRGAN SwinIR
Builds on: Stable Diffusion variational score distillation Diffusion-GAN DISTS Real-ESRGAN degradation pipeline
Closest papers by meaning
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 2025 · similarity 0.92)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.92)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.91)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.91)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.91)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.90)
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.