Research map / Image generation / Super-resolution
Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs
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 DoSSR, a domain-shift diffusion SR model that starts sampling from LR images, fine-tunes SD 2.1 via LAControlNet, and derives continuous DoS-SDEs with fast customized solvers reaching strong quality in 5 steps.
Problem
Diffusion-based super-resolution either leverages pretrained diffusion priors but needs many steps starting from Gaussian noise, or starts from LR images but trains from scratch and cannot use pretrained diffusion priors.
Approach
A domain shift equation linearly interpolates between LR and HR with a shifting sequence that is compatible with the DDPM noise schedule, so pretrained SD can be fine-tuned and sampling starts at an intermediate timestep from LR. The shift is extended to continuous DoS-SDEs with an exact solution and customized first-to-third-order solvers that include a domain shift guidance term.
Key results
- DoSSR uses only 5 sampling steps (NFE 5) and reports 1.03s latency per 128x128 image on V100, versus 5.85s for DiffBIR (50 NFE) and 7.24s for SeeSR (50 NFE), a 5-7x speedup over prior SD-based methods.
- On RealSR, DoSSR achieves the best MANIQA (0.5781) and TOPIQ (0.6985) and the best CLIPIQA (0.7025), while its PSNR (24.18) trails the best GAN-based and ResShift baselines.
- DoSSR can match prior methods' performance even with a single step, reaching 0.5115 MANIQA on RealSR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DoSSR's domain-shift diffusion process and DoSG instead of the original DDPM diffusion equation (ablation) | Diffusion-based SR model · SR benchmarks in the DoSG ablation | Metrics reported in the ablation table | Helped |
| Decrease the starting point t1 from T to T/2 (ablation) | DoSSR · SR performance across t1 settings | SR performance | Helped |
| Decrease the starting point t1 further from T/2 to 3/T (ablation) | DoSSR · SR performance across t1 settings | SR performance | Hurt |
| Increase the number of inference steps (ablation) | DoSSR · RealSR | Reference-based metrics | Hurt |
| Increase the number of inference steps (ablation) | DoSSR · RealSR | Non-reference metrics | Helped |
| Use one inference step instead of more sampling steps (ablation) | DoSSR · RealSR | MANIQA: 0.5115 MANIQA; 0.6258 CLIPIQA | Mixed |
| Use a higher-order DoS-SDE solver rather than the first-order solver (ablation) | DoSSR · SR benchmarks under limited inference steps | Non-reference metrics | Helped |
| Use DoSSR rather than SeeSR | Stable Diffusion-based image SR · DIV2K-Val, RealSR, DRealSR, and Real200 | CLIPIQA: 0.6935 -> 0.7014; 0.6699 -> 0.7025; 0.6708 -> 0.6776; 0.7167 -> 0.7437 | Helped |
| Use DoSSR rather than SeeSR | Stable Diffusion-based image SR · DIV2K-Val, RealSR, DRealSR, and Real200 | MANIQA: 0.5041 -> 0.5294; 0.5406 -> 0.5781; 0.5115 -> 0.5214; 0.5588 -> 0.5794 | Helped |
| Use DoSSR rather than SeeSR | Stable Diffusion-based image SR · x4 SR, 128×128 LR images | NFE: 50 -> 5 | Helped |
| Use DoSSR rather than SeeSR | Stable Diffusion-based image SR · x4 SR, 128×128 LR images | Latency/Image: 7.24s -> 1.03s | Helped |
Surprising findings
- Reference-based fidelity metrics (PSNR/SSIM) of DoSSR are lower than several GAN-based baselines, so perceptual gains come at fidelity cost.
Limitations
- Output quality is sensitive to the random seed; some seeds produce visually unfriendly details such as inconsistent eyes.
- Reference metrics (PSNR, SSIM) are lower than GAN-based and some diffusion baselines.
Experimental setup
Datasets: DIV2K DIV8K Flickr2K OST RealSR DRealSR RealLR200 ImageNet
Benchmarks: DIV2K-Val RealSR DRealSR RealLR200
Metrics: PSNR SSIM LPIPS CLIPIQA MUSIQ MANIQA TOPIQ latency
Baselines: StableSR ResShift DiffBIR SeeSR Real-ESRGAN BSRGAN LDL DASR
Builds on: DDPM ResShift ControlNet StableSR Stable Diffusion
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.94)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.93)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.92)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.92)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.91)
- Efficient Conditional Diffusion Model with Probability Flow Sampling for Image Super-resolution (AAAI 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.