- The paper presents a novel training-only diffusion prior supervision strategy that enables efficient single image dehazing with zero additional inference cost.
- The method leverages a dual-path feed-forward backbone with semantic and color-aware structural encoders to achieve superior PSNR and SSIM improvements.
- Quantitative benchmarks confirm its robustness against complex haze, outperforming existing CNN and transformer models in speed and restoration fidelity.
ZID-Net: A Zero-Inference Diffusion Prior Network for Efficient and Accurate Single Image Dehazing
Introduction
Single image dehazing remains a central problem in image restoration, especially within challenging real-world environments characterized by spatially heterogeneous and dense haze. Traditional learning-based approaches confront a persistent trade-off: feed-forward CNN and transformer architectures enable efficient inference but lack the capacity to robustly model complex haze-induced degradations, whereas diffusion-based models offer powerful generative priors at the expense of severe inference latency and instability. "ZID-Net: Zero-Inference Diffusion Prior Decoupling Network for Single Image Dehazing" (2604.23709) introduces a novel resolution to this dilemma by strictly decoupling diffusion model supervision from inference, thus integrating the benefits of generative priors into a computationally efficient feed-forward backbone.
Methodology
Frequency-Spatial Decoupled Feed-Forward Backbone
The core of ZID-Net is a dual-path deterministic backbone composed of two orthogonal streams: a semantic context encoder and a color-aware structural encoder. The semantic context encoder utilizes four-stage residual blocks and stride-2 convolutions, outputting a compact bottleneck representation. To facilitate high-capacity global semantic reasoning without transformer-like computational burdens, ZID-Net incorporates Lightweight Global Context Blocks (LGCBs) at the bottleneck. LGCBs reformulate self-attention in the channel space, computing context aggregation with O(CN) complexity, linear in spatial resolution.
The color-aware structural encoder applies a Channel-Spatial Laplacian Mask (CSLM), extracting and adaptively purifying chromatic high-frequency residuals at the channel level. Unlike traditional edge detectors or grayscale Laplacians, CSLM leverages channel-specific Laplacian computations combined with dynamic channel and spatial gating to preserve critical chromatic information while suppressing haze-amplified noise.
A Dynamic Feature Arbitration Block (DFAB) adaptively fuses semantic, structural, and upsampled decoder states using channel Squeeze-and-Excitation reweighting per scale, optimizing the structural-semantic information balance at each reconstruction stage.
Zero-Inference Prior Propagation Head (ZI-PPH)
A defining contribution is the Zero-Inference Prior Propagation Head, active only during training. ZI-PPH is formulated as a conditional diffusion branch: it receives jointly the haze residual and the backbone bottleneck and predicts residual noise at variable diffusion timesteps. Through a batch-adaptive severity sampling schedule, ZI-PPH enforces degradation-aware noise prediction across multiple haze concentrations. The resultant loss directly regularizes the backbone, compelling it to encode degradation-consistent, generative cues. At test time, ZI-PPH is entirely disabled, so inference remains a single deterministic forward pass with zero additional latency.
Training Objectives
The overall objective combines (1) pixel-wise L1 reconstruction loss, (2) a VGG-19-based contrastive perceptual loss, and (3) L1 diffusion noise prediction loss from ZI-PPH. This joint supervision enables dense pixel-level fidelity, high perceptual quality, and implicit generative guidance.
Quantitative and Qualitative Results
Synthetic Benchmarks
On the RESIDE SOTS datasets, ZID-Net achieves 42.77 dB PSNR / 0.997 SSIM (Indoor) and 38.73 dB PSNR / 0.993 SSIM (Outdoor), outperforming transformers and pure CNN models by significant margins. This performance highlights that frequency-spatial decoupling combined with zero-inference generative supervision resolves feature entanglement and effectively reconstructs haze-free images under strict synthetic degradations.
Real-World Benchmarks
In real-world thin and dense haze scenarios, ZID-Net demonstrates leading generalization, with PSNR gains of 1.29 dB (Thin) and 0.97 dB (Dense) over the strongest competitors. The architecture robustly manages intrinsic color preservation and structural fidelity where classic methods either over-smooth or hallucinate artifacts, confirming the advantage of the diffusion-regularized backbone.
Remote Sensing
On the StateHaze1k benchmark, especially under thick haze, ZID-Net displays a PSNR advantage of 5.57 dB over prior state-of-the-art, indicating superior resilience to severe real-world atmospheric scattering. The backbone’s generative regularization, via the training-only diffusion prior, is notably effective in highly non-homogeneous and information-depleted conditions.
Efficiency
Despite its architectural richness, ZID-Net demonstrates superior inference speed: 5.14 ms (256x256) / 19.35 ms (512x512) per image, well within real-time constraints. This contrasts sharply with iterative diffusion-based models, which require several seconds to minutes per image due to their multi-step stochastic sampling chains.
Ablation and Component Analysis
Ablation studies confirm each core module’s necessity. Removal of the structural branch leads to a PSNR drop exceeding 2 dB on real-world datasets, substantiating the importance of explicit high-frequency modeling. LGCBs and DFAB also contribute clear numerical and perceptual improvements, validating their design rationale. Critically, eliminating diffusion-based supervision reduces PSNR by approximately 2 dB, and competing auxiliary supervision branches (e.g., atmospheric light or transmission estimation) provide only moderate benefits relative to ZI-PPH. The color Laplacian as a structural prior outperforms classic gradient or Laplacian operators across both distortion and perceptual metrics.
Implications and Future Directions
By architecting a framework where diffusion generative priors are distilled into deterministic models exclusively during training, ZID-Net establishes a new paradigm in low-level restoration. This approach addresses the fundamental limitation of iterative inference in diffusion models, making state-of-the-art restoration achievable in real-time systems such as autonomous driving, surveillance, and remote sensing.
Potential extensions include:
- Extending zero-inference prior distillation to related inverse problems (e.g., deraining, desnowing, low-light enhancement).
- Further optimizing the conditional diffusion supervision to adaptively target harder instances during training.
- Exploring hybrid training regimes where other expressive generative models provide prior regularization, potentially increasing generalizability in unseen degradations.
Conclusion
ZID-Net presents a methodologically rigorous and effective integration of generative diffusion priors into an efficient feed-forward image dehazing architecture. The decoupling of diffusion supervision from inference fundamentally resolves the tension between expressive degradation modeling and real-time applicability. Empirical results confirm ZID-Net’s consistent outperformance across synthetic, real-world, and remote sensing benchmarks, both in restoration fidelity and computational efficiency. This zero-inference regularization strategy has substantial implications for the future development of practical, high-performance low-level vision systems.
