Local Epistemic Uncertainty Guided Active Sampling for Plug-and-play Diffusive Image Restoration

ACM MM 2026 Oral!

Jiaqi Zhang1,*, Zheng Pang1,*, Rongrong Gao1, Qiyuan Zhang2, Yang Yang1,†

1Jiangsu University, 2Xi'an University of Technology

*Equal contribution. †Corresponding author.

Input
Degraded input for the ProDiff example
ProDiff
Input before ProDiff restoration Restoration produced by ProDiff
+ LEADer
ProDiff restoration before LEADer refinement ProDiff enhanced with LEADer

Abstract

Diffusion models have demonstrated remarkable effectiveness in image restoration tasks. However, existing Diffusion Model-based Image Restoration (DMIR) methods typically rely on fixed data constraints and uniform step sizes, overlooking the dynamic nature of the generative process. These rigid designs are vulnerable to spatially non-uniform degradations and introduce computational redundancy. We propose Local Epistemic Uncertainty Guided Active Sampling, termed LEADer. In the spatial domain, LEADer leverages pixel-wise uncertainty to dynamically modulate the prior strength within the null space, balancing detail preservation and artifact suppression. In the temporal domain, it quantifies sampling stability via the uncertainty trace to enable adaptive trajectory pruning. LEADer can be seamlessly integrated into various DMIR baselines and improves restoration quality while significantly reducing sampling time.

Method Overview

LEADer revisits plug-and-play diffusion restoration from a state-aware sampling perspective. Instead of applying the same restoration constraint everywhere and at every denoising step, it asks how reliable the diffusion prior is for each local region and how stable the current sampling state is. This uncertainty signal then controls both the spatial correction strength and the temporal sampling schedule.

Architecture of the LEADer framework.
Figure 1. Architecture of LEADer. Existing DMIR methods rely on fixed step-by-step sampling with uniform constraints, leading to computational redundancy and detail distortion in uncertain regions. To address this, LEADer quantifies local epistemic uncertainty Σt to guide reverse sampling. Spatially, UCPM utilizes Σt to adaptively modulate prior constraints for strict data consistency. Temporally, SATP computes an adaptive step size ΔtActive based on the uncertainty trace to skip redundant iterations. As a result, LEADer improves the restoration quality of various baselines while reducing sampling time.

Comparison Analysis

LEADer improves both visual quality and efficiency across restoration tasks, including super-resolution, Gaussian deblurring, motion deblurring, and compressed sensing. The comparison highlights the core benefit of uncertainty-guided restoration: baseline methods with fixed sampling rules often lose local details or introduce artifacts under non-uniform degradations, while LEADer adaptively allocates correction strength according to the current restoration difficulty.

Qualitative comparisons on five restoration tasks.
Figure 2. Qualitative comparisons on five image restoration tasks.
Quantitative results across five image restoration tasks on CelebA-HQ 1K and ImageNet 1K.
Table 1. Quantitative results for five image restoration tasks on CelebA-HQ 1K (top) and ImageNet 1K (bottom). The average change reports performance improvements and time reductions over each baseline.

Sampling Trajectory Analysis

Compared with fixed-step baselines, LEADer dynamically adjusts the sampling pace and converges to clearer local structures at lower numbers of function evaluations. The trajectory visualization shows that DDNM improves details gradually through dense and uniform updates. In contrast, LEADer identifies stable states earlier and prunes redundant steps, while preserving the important denoising updates needed for local texture recovery.

Comparison between DDNM and LEADer at different NFEs.
Figure 3. Comparison of restored states between DDNM and LEADer at different NFEs.

Ablation Analysis

As shown in Table 2, UCPM mainly improves restoration fidelity by making prior modulation spatially adaptive, while SATP reduces redundant computation by adjusting the effective number of sampling steps according to the uncertainty trace. Their combination gives the strongest trade-off between quality and speed, as the improved sampling stability from UCPM makes trajectory pruning safer. The information-loss budget controls the balance between restoration quality and acceleration: a conservative budget permits fewer skipped steps, while an excessively large budget reduces runtime at the cost of noticeable quality degradation. The corresponding sensitivity results in Table 3 and Figure 4 show that 0.01 provides a balanced default setting.

Ablation study results for UCPM and SATP.
Table 2. Ablation study of the UCPM and SATP components on three representative restoration tasks.
Hyperparameter sensitivity analysis for the information-loss budget.
Ablation study on different sampling steps.
Table 3 and Figure 4. Hyperparameter sensitivity analysis on the information-loss budget B, together with ablation results across different sampling steps on CelebA-HQ 1K.

Memory Consumption

LEADer introduces only marginal GPU memory overhead when integrated into existing DMIR methods. Across five ImageNet restoration tasks, the additional memory consumption remains between 0.07% and 0.28%, showing that the uncertainty-guided modules improve quality and efficiency without materially increasing deployment cost.

GPU memory consumption comparison across five restoration tasks.
Table 4. Memory consumption comparison across five image restoration tasks.

BibTeX

@article{LEADer,
  title={Local Epistemic Uncertainty Guided Active Sampling for Plug-and-play Diffusive Image Restoration},
  author={Jiaqi Zhang and
          Zheng Pang and
          Rongrong Gao and
          Qiyuan Zhang and
          Yang Yang},
  journal={arXiv preprint arXiv:2608.06981},
  year={2026}
}