An arXiv preprint reports results from DPC-Net, one image-restoration model tested across five kinds of image damage: denoising, deraining, dehazing, deblurring and low-light enhancement. In the paper’s expanded five-degradation comparison, its row averages 31.00 decibels in peak signal-to-noise ratio (PSNR) and 0.923 in structural similarity (SSIM), two image-quality measures for which higher scores are preferred. The row lists 27 million parameters.
On another all-in-one benchmark configuration, the reported average is 33.01 dB PSNR and 0.922 SSIM, again with 27 million parameters.
The model uses two kinds of clues
DPC-Net is built around two types of information: degradation-semantic priors, which describe the image’s damage, and low-level visual priors, which carry fine visual information. The paper combines both within the restoration network.
A vision-language model (VLM) supervises a module called DAN by constraining its feature distribution. Another module, DSMM, turns that guidance into degradation-semantic coupling and sends it to the decoder.
During decoding, knowledge bases supply the low-level visual priors, which are injected alongside the DSMM priors in a component called DPCR. The reported implementation uses LLaVA as the VLM, Restormer as the backbone and knowledge bases with 256 features.
The tests span denoising, deraining, dehazing, deblurring and low-light enhancement. For denoising, the paper reports noise-level values of 15, 25 and 50.
What the tables report
In single-task comparisons, DPC-Net reports 32.15 dB PSNR and 0.982 SSIM for dehazing, 37.65 dB and 0.981 for deraining, and 31.73 dB and 0.900 for denoising.
For single-task dehazing, the authors report a PSNR 0.65 dB higher than R2R.
The component ablation compares the full model with four variants. The full model scores 33.01/0.922; the variants score 32.65/0.920, 32.86/0.920, 32.66/0.919 and 32.79/0.920, leaving the full model as the highest listed entry.
The guidance comparison also puts the LLaVA option at the top of its table, at 33.01/0.922, versus 32.46/0.918 for the prompt option and 32.64/0.919 for DA-CLIP.
In a QGM comparison, the m=256 setting with QGM reports 33.01/0.922, compared with 32.75/0.920 without QGM. The with-QGM m=256 entry is the highest listed result in that table.
What remains unanswered
The document is an arXiv preprint, version 1, dated 20 August 2026.
The supplied text does not report dataset image counts or exact split sizes; it refers to an Appendix for dataset details. The scale of the evaluation therefore cannot be judged from the supplied material alone.
The reported figures are benchmark results on the named tasks and configurations. The supplied review leaves open how DPC-Net would perform on unseen datasets, real-world inputs or mixed and unknown degradations.
The supplied analysis reports no confidence intervals, repeated-run variability, random seeds or inferential tests alongside the benchmark comparisons. It also says the evaluation relies mainly on PSNR and SSIM, without a human perceptual study or separately quantified semantic-consistency assessment.
Although the tables report 27 million parameters, runtime, memory use and energy consumption are not reported in the supplied review.
Paper data and sources
Original title: DPC-Net: Dual-Prior Collaborative Network for All-in-One Image Restoration
Authors: Zhaokun He, Kangbiao Shi, Axi Niu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text