Research map / Image generation / Fast diffusion sampling
Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise
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.
Resfusion introduces a residual term into the diffusion forward process and predicts a weighted residual noise (resnoise) so that image restoration starts from the noisy degraded image and needs only five sampling steps.
Problem
Diffusion-based image restoration starts reverse sampling from Gaussian white noise even though degraded inputs already carry low-frequency content, which requires many sampling steps; prior residual-diffusion methods (RDDM, ResShift) need complex noise schedules and inconsistent reverse processes.
Approach
Incorporate the residual R = x̂0 − x0 into the DDPM forward process so that the noisy degraded image is a valid intermediate state; define a weighted residual noise (resnoise) as the single prediction target, and use a smooth equivalence transformation to find an acceleration step T' (where the coefficient of x0 is about zero) from which sampling starts directly with the degraded image, allowing s
Key results
- On ISTD at 256x256, Resfusion reaches 31.81 PSNR / 0.965 SSIM overall with 7.7M params, versus 30.91 for RDDM (SM-Res-N) with 15.5M.
- On LOL (RGB, original resolution), Resfusion attains 24.63 PSNR / 0.860 SSIM / 0.107 LPIPS with five sampling steps.
- On CIFAR-10 FID, Resfusion (e.g., 6.31 at 100 steps) beats DDPM at equal steps (9.79) but stays behind DDIM (6.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use resnoise prediction instead of x₀ prediction (ablation) | Resfusion · ISTD original | PSNR: 29.67 -> 30.09 | Helped |
| use resnoise prediction instead of R prediction (ablation) | Resfusion · ISTD original | PSNR: 29.75 -> 30.09 | Helped |
| use resnoise prediction instead of x₀ prediction (ablation) | Resfusion · LOL original | PSNR: 23.10 -> 24.63 | Helped |
| use resnoise prediction instead of R prediction (ablation) | Resfusion · LOL original | PSNR: 22.87 -> 24.63 | Helped |
| use resnoise prediction instead of x₀ prediction (ablation) | Resfusion · Raindrop original | PSNR: 32.51 -> 32.61 | Helped |
| use resnoise prediction instead of R prediction (ablation) | Resfusion · Raindrop original | PSNR: 32.57 -> 32.61 | Helped |
| predict only the noise component, omitting the weighted residual component (ablation) | Resfusion resnoise-diffusion · LOL | image restoration quality | Hurt |
| predict only the weighted residual component, omitting the noise component (ablation) | Resfusion resnoise-diffusion · LOL | image restoration quality | Hurt |
| omit truncated schedule (ablation) | Resfusion · image restoration | residual shadows | Hurt |
| use Resfusion instead of DDPM at the same sampling steps | CIFAR10 32×32 · CIFAR10 32×32, 10 sampling steps | FID: 43.11 -> 28.81 | Helped |
| use Resfusion instead of DDPM at the same sampling steps | CIFAR10 32×32 · CIFAR10 32×32, 20 sampling steps | FID: 24.88 -> 15.46 | Helped |
| use Resfusion instead of DDPM at the same sampling steps | CIFAR10 32×32 · CIFAR10 32×32, 50 sampling steps | FID: 14.02 -> 7.96 | Helped |
Surprising findings
- Predicting the plain noise term epsilon is not an equivalent loss in this setting and leads to training failure due to the truncated schedule.
- Resfusion's FID on CIFAR-10 remains above deterministic DDIM at equal step counts, since its stochastic sampling mirrors DDPM.
Limitations
- Performance may trail task-specific state-of-the-art methods that use specialized architectures or physics priors.
- Feature fusion is a simple channel concatenation, and the diffusion operates in pixel space rather than latent space.
- The truncated schedule is needed for small T to control the offset between x_T' and the degraded-image estimate.
Experimental setup
Datasets: ISTD LOL Raindrop LOL-v2-real CIFAR-10 CelebA-HQ AFHQ-v2
Benchmarks: ISTD LOL Raindrop LOL-v2-real CIFAR-10
Metrics: PSNR SSIM MAE LPIPS FID Params MACs inference time
Baselines: RDDM (SM-Res-N) RDDM (SM-Res) DMTN DHAN UnfoldingNet LLFormer Restormer WeatherDiff RainDropDiff DDPM DDIM
Builds on: DDPM RDDM ResShift DDIM Shadow Diffusion
Closest papers by meaning
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.92)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.90)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.89)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.89)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.89)
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.89)
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.