Improving Sampling Efficiency in Diffusion Models: Degradation-Aware Adaptive Diffusion Sampling and Uncertainty Modeling for Single-Image Dehazing
Single-image dehazing is ill-posed because scene radiance and atmospheric effects are mixed in one observation. Rather than treating diffusion as a general replacement for deterministic dehazing, we examine a narrower issue: how far its reverse process can be shortened once a physics-guided estimate has already recovered much of the low-frequency scene structure. The network then predicts only the remaining signed residual. Transmission-derived pseudo-supervision supplies the global and spatial degradation cues, while a separate difficulty score enters only the reverse-time scheduling decision. An initial attempt to predict the image-dependent NFE budget directly collapsed toward nearly constant small values on validation data. We therefore keep K small and adapt the positions of those evaluations within a power-law schedule family. Experiments use 730 paired images with a fixed 567/73/90 train-validation-test split. On the 90-image internal test split, with K = 2 reverse denoiser function evaluations (NFEs), residual-x0 diffusion reaches 18.10 dB peak signal-to-noise ratio (PSNR), 0.740 structural similarity index (SSIM), and 0.300 learned perceptual image patch similarity (LPIPS). Relative to the K = 32 residual-diffusion reference, this removes 93.75% of denoiser evaluations; the original measurements report an end-to-end time reduction of about 87.8%. A deterministic residual CNN is still both faster and more accurate in absolute terms. At K = 8, degradation- and uncertainty-aware scheduling improves the uniform schedule by about 0.40 dB PSNR and 0.004 SSIM at the same reverse-step budget, although the Holm-adjusted p-value of approximately 0.056 does not meet the 0.05 significance threshold. These results support a sampling-efficiency claim within the evaluated diffusion setting, not a claim that diffusion is the fastest or strongest dehazing model overall.
Authors
- Ding Wang
- Guiguang Ding
Institutions
- Johns Hopkins University (US)
- Tsinghua University (CN)
Publication Details
- Journal
- Mathematics
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3390/math14224221
- Primary Topic
- Image Enhancement Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00