Image acquisition with a camera involves several degradations due to the optical system, sensor, or low-level processing steps. We address blind deblurring in professional photography: we aim to invert unknown isotropic blur without knowledge of the degradation kernel. For such inverse problems,where some high-frequency information is lost, it is challenging to use generative models to produce details that are both photo-realistic and faithful to the input. We propose SuperSharpen, a diffusion-based blind deblurring method offering explicit control over restoration strength through a blur measure. We compare two conditioning strategies: a ControlNet-style adapter on a frozen backbone, and full finetuning of the diffusion prior. Our experiments show that finetuning achieves better fidelity with fewer hallucinated details. We validate our approach on synthetic and real-world blur, demonstrating improved perceptual quality and controllable restoration strength.
Figures & tables
Figure 1: We present the training setups for (a) ControlNet-based conditioning (SupS-CN) and (b) finetuning-based conditioning (SupS-FT). Both models rely on a conditional score-based latent denoiser ϵθ that takes as input the encoded Low Quality image and the corresponding blur measure in addition to the encoded Ground Truth to which we add noise zt . In (a) only the ControlNet adapters are trained ( ) while the diffusion backbone is frozen ( ). In (b) the entire diffusion denoiser is finetuned ( ). Noting that in both training setups, the VAE encoder E is frozen ( ). In (c), we illustrate the inference setup only for SupS-FT which generates the high-quality image by iterative use of ϵθ conditioned by the degraded image and a user-defined blur measure to control the restoration strength. Notice that at training t is sampled randomly while at inference, t is decreased from T to 0 .
Figure 2: Blur measure computation setup. We estimate the effective blur kernel of the degradation pipeline by computing its impulse response from multiple impulses located at different spatial locations. We then fit an isotropic Gaussian kernel to the average impulse response and define our blur measure as the standard deviation of the fitted Gaussian.
Figure 3: Qualitative comparison on images with synthetic blur. The sharpest results are obtained with DiffBIR, SupS-CN and SupS-FT, the latter producing less unplausible structures.
Figure 4: Qualitative comparison on real blurry images. Compared to the other methods, our SupS models attain a better compromise in terms of photo-realism and fidelity to the input.
Full-reference
No-reference
Method
PSNR ↑
SSIM ↑
LPIPS ↓
MUSIQ ↑
MANIQA ↑
CLIP-IQA ↑
HQ
–
1.00
0.00
70.4433
0.5534
0.6762
DiffBIR
18.3051
0.5284
0.2278
71.6563
0.5986
0.7107
PASD
21.9344
0.6518
0.2138
68.9798
0.4850
0.5904
SeeSR
20.1731
0.5808
0.2463
70.4701
0.5529
0.6541
ResShift
19.7569
0.4976
0.4138
48.0120
0.2960
0.4199
Table 1: Quantitative results on synthetically degraded images from Urban100 dataset. Best and second best results are highlighted in bold and underlined , respectively.
Figure 5: Qualitative results on real images with different blur measure (BM) values. We see that higher BM values lead to stronger correction and sharper images.