Research map / Image generation / Super-resolution
Efficient Conditional Diffusion Model with Probability Flow Sampling for 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 ECDP, a continuous-time SDE-based conditional diffusion model for image super-resolution that uses probability flow ODE sampling, a hybrid epsilon/x0 denoiser parametrization, and a VGG feature-space image quality loss computed through the ODE sampler.
Problem
Diffusion-based image super-resolution is slow due to iterative stochastic sampling and suffers from quality and consistency issues such as color shifting.
Approach
Formulate conditional diffusion with a forward SDE whose mean is the bicubic-upscaled LR image and whose variance is a per-dataset constant, sample via probability flow ODE, interpolate between epsilon- and x0-parametrizations over noise scales, and add a feature-space image quality loss backpropagated through the ODE via the adjoint method.
Key results
- On DIV2K 4x, ECDP reaches LPIPS 0.108, PSNR 28.03 and SSIM 0.79, the best LPIPS among the compared methods.
- Inference takes 2.0s per 256x256 image versus 2.4s for SRDiff, 6.2s for IR-SDE and 83.1s for SR3.
- On CelebA 8x, ECDP achieves LPIPS 0.097, the best LPIPS among diffusion-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ECDP instead of SRDiff | DIV2K 4× super-resolution · DIV2K 4× | LPIPS: 0.136 -> 0.108 | Helped |
| use ECDP instead of SR3 | ImageNet super-resolution · ImageNet 64×64→256×256 | LPIPS: 0.191 -> 0.110 | Helped |
| use ECDP instead of SRDiff | CelebA 8× super-resolution · CelebA 8× | LPIPS: 0.106 -> 0.097 | Helped |
| use probability flow sampling | ECDP image generation · 256×256 HR image generation | inference time: 2.4s -> 2.0s | Helped |
| set hybrid interpolation parameter c=0.5 instead of c=1.0 (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.109 | Hurt |
| set hybrid interpolation parameter c=1.5 instead of c=1.0 (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.107 | Helped |
| replace hybrid parametrization with ε-parametrization (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.128 | Hurt |
| replace hybrid parametrization with x₀-parametrization (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.112 | Hurt |
| remove image quality loss (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.120 | Hurt |
| compute quality loss without probability flow (ablation) | ECDP · DIV2K 4× | LPIPS: 0.108 -> 0.112 | Hurt |
| add image quality loss (ablation) | SRDiff · DIV2K 4× | LPIPS: 0.136 -> 0.126 | Helped |
| add image quality loss (ablation) | SR3 · DIV2K 4× | LPIPS: 0.175 -> 0.167 | Helped |
Limitations
- Image quality loss requires generating images during training, which is costly.
- The variance and mean are approximated with bicubic upscaling and a constant variance rather than true conditional statistics.
Experimental setup
Datasets: DIV2K Flickr2K ImageNet CelebA
Benchmarks: DIV2K ImageNet CelebA
Metrics: LPIPS PSNR SSIM
Baselines: RRDB ESRGAN SRFlow HCFlow SR3 SRDiff IR-SDE GDP DiffPIR
Builds on: VP-SDE score matching RRDB SR3 SRDiff
Closest papers by meaning
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.91)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.91)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.91)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.91)
- One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation (ICML 2025 · similarity 0.90)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · 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.