ZeroIDIR: Zero-Reference Illumination Restoration

• ZeroIDIR is a zero-reference image restoration framework that addresses illumination degradation by combining adaptive exposure correction with diffusion-based detail reconstruction.
• The framework employs a two-stage pipeline where the Adaptive Gamma Correction Module adjusts spatially varying illumination and the Perturbed Consistency Diffusion Model refines textures and suppresses noise.
• Empirical evaluations demonstrate that ZeroIDIR outperforms unsupervised baselines across diverse benchmarks, highlighting its effective decomposition strategy and regularization techniques.

Searching arXiv for ZeroIDIR and closely related restoration work to ground the article. ZeroIDIR is a zero-reference image restoration framework for illumination degradation that targets three related tasks—low-light image enhancement (LLIE), backlit image enhancement (BIE), and multiple exposure correction (MEC)—while being trained solely on low-quality degraded images. It is defined by an explicit decomposition of restoration into adaptive illumination correction and diffusion-based detail reconstruction, with the central claim that diffusion models are effective for reconstruction and denoising but exhibit bias in low-frequency generation, especially exposure correction. ZeroIDIR therefore treats illumination correction as a distinct subproblem and uses a perturbed consistency diffusion model only after exposure has been corrected into a more reliable intermediate representation (Jiang et al., 12 May 2026).

1. Problem setting and conceptual basis

ZeroIDIR addresses illumination degradation image restoration (IDIR) under challenging capture conditions in which images may be under-exposed, over-exposed, or affected by mixed exposure issues. The degradation manifests as poor visibility, color distortion, detail loss, and noise amplification. In this setting, the method is explicitly zero-reference: it uses no paired normal-light references and is trained only on low-quality images. The paper positions this against traditional hand-crafted approaches such as histogram equalization or Retinex-based methods, which often fail across diverse scenes, and against supervised deep models, which require paired data and may overfit to training distributions (Jiang et al., 12 May 2026).

The framework is motivated by a particular diagnosis of diffusion-based restoration. Diffusion models are described as strong for high-frequency reconstruction, but not necessarily for accurate illumination correction. This leads to the central design principle of ZeroIDIR: exposure correction should be handled explicitly, while diffusion should focus on refinement, detail reconstruction, and noise suppression. This suggests that the method is not merely a diffusion model adapted to low-light enhancement, but a deliberately decoupled restoration system in which low-frequency illumination statistics and high-frequency content reconstruction are assigned to different modules (Jiang et al., 12 May 2026).

2. Two-stage architecture and restoration pipeline

The full pipeline begins with a degraded input image $I_d$, which is first decomposed by Retinex into reflectance and illumination components: $I_d \rightarrow R_d, L_d$ where $R_d$ denotes reflectance or content information and $L_d$ denotes the illumination map. This decomposition provides the structural basis for the first stage, the Adaptive Gamma Correction Module (AGCM), which predicts spatially varying gamma maps and spatial weight maps, outputs corrected illumination $L'_d$, and reconstructs an illumination-corrected image

$I'_d = L'_d \odot R_d.$

The second stage is the Perturbed Consistency Diffusion Model (PCDM). Rather than operating directly on the original degraded image, PCDM treats the illumination-corrected image $I'_d$ as an intermediate perturbed state $x_{t^\ast}$, diffuses further from that state, and reconstructs the final restored image $\hat I_d$. The architectural significance of this arrangement is that AGCM handles exposure correction, while PCDM handles refinement, detail reconstruction, and noise suppression. A plausible implication is that the decomposition reduces the burden on the diffusion model to infer both global exposure statistics and local texture recovery from the same degraded input (Jiang et al., 12 May 2026).

3. Adaptive Gamma Correction Module

AGCM is motivated by the observation that a single global gamma value is too rigid for real degraded images because exposure problems are often spatially varying. Some regions may be under-exposed, others already correctly exposed, and others over-exposed. AGCM therefore operates on the illumination component $L_d$ rather than directly on the RGB image, using illumination-aware structural features derived from concatenated $I_d \rightarrow R_d, L_d$0, together with channel attention and convolutional blocks (Jiang et al., 12 May 2026).

The module predicts two gamma maps,

$I_d \rightarrow R_d, L_d$1

where $I_d \rightarrow R_d, L_d$2 is intended for under-exposed regions and $I_d \rightarrow R_d, L_d$3 for over-exposed regions, and two spatial weight maps,

$I_d \rightarrow R_d, L_d$4

The corrected illumination is then defined as

$I_d \rightarrow R_d, L_d$5

and the corrected image is reconstructed by

$I_d \rightarrow R_d, L_d$6

To regularize the corrected illumination toward natural-scene exposure statistics, ZeroIDIR introduces a histogram-guided illumination correction loss: $I_d \rightarrow R_d, L_d$7 where $I_d \rightarrow R_d, L_d$8 computes the histogram and $I_d \rightarrow R_d, L_d$9 is an empirical histogram prior aggregated from a large corpus of normal-light images. The paper states that this loss prevents arbitrary or unstable exposure shifts; without it, gamma prediction becomes unstable and results may be over- or under-corrected. AGCM is also trained with an exposure control loss,

$R_d$0

with $R_d$1, and an edge-aware total variation loss,

$R_d$2

which keeps gamma maps smooth while preserving edges (Jiang et al., 12 May 2026).

4. Perturbed Consistency Diffusion Model

PCDM is built around a reinterpretation of diffusion training for the zero-reference setting. Standard diffusion training assumes access to a clean image $R_d$3, with forward process

$R_d$4

where $R_d$5. ZeroIDIR lacks clean targets, so it instead treats the AGCM output as an intermediate noisy state: $R_d$6 The forward process is then continued from that intermediate state over additional steps $R_d$7, with $R_d$8, via

$R_d$9

The denoising network $L_d$0 is trained to predict injected noise using

$L_d$1

where the condition $L_d$2 is the illumination-corrected image $L_d$3. In this formulation, the diffusion model is conditioned on an exposure-corrected representation rather than on the original degraded image. The paper explicitly argues that this focuses the diffusion network on detail reconstruction, texture recovery, and noise suppression rather than requiring it to jointly solve exposure correction from scratch (Jiang et al., 12 May 2026).

To stabilize the restoration trajectory, ZeroIDIR adds a perturbed diffusion consistency loss. The model predicts a clean image estimate $L_d$4, re-perturbs it to time $L_d$5, and constrains the result to remain consistent with the original perturbed input in VGG-16 feature space: $L_d$6 According to the paper, removing this loss causes diffusion drift and artifacts. This suggests that $L_d$7 functions as a trajectory regularizer that anchors the generated solution to the illumination-corrected intermediate state when no clean supervision is available (Jiang et al., 12 May 2026).

5. Training protocol and empirical evaluation

ZeroIDIR uses a two-stage training strategy. In Stage 1, the diffusion model is frozen and AGCM is optimized with

$L_d$8

In Stage 2, AGCM is frozen and the diffusion stage is trained with

$L_d$9

The training set contains 10k low-quality illumination degradation images, including low-light, backlit, and over-exposed images, with no paired normal-light targets (Jiang et al., 12 May 2026).

The method is evaluated on LOL, LSRW, and MIT5K for LLIE; BAID and Backlit300 for BIE; and MSEC and SICE for MEC. The reported metrics are PSNR, SSIM, and LPIPS for paired data, and NIQE and CLIPIQA for unpaired data. The paper states that ZeroIDIR outperforms most unsupervised baselines, is comparable to supervised diffusion-based methods, and often generalizes better to unseen datasets. It reports that on LSRW ZeroIDIR beats supervised methods in the table, on MIT5K it achieves very strong performance and the best reported PSNR among the listed methods, on BAID and Backlit300 it achieves the best overall performance among compared methods, and on SICE it shows particularly strong generalization relative to supervised methods whose performance is more dataset-specific (Jiang et al., 12 May 2026).

The qualitative comparisons are described as showing that earlier methods often produce incorrect exposure, blurred details, color distortion, and noise amplification, whereas ZeroIDIR produces better global and local illumination correction, sharper details, more vivid colors, and less noise. Within the framework of the paper, this empirical pattern is presented as evidence that decoupling exposure correction from diffusion-based reconstruction is effective across low-light, backlit, and multiple-exposure regimes (Jiang et al., 12 May 2026).

6. Ablations, significance, and limitations

The ablation studies attribute specific functions to each design choice. Fixed gamma values are reported to perform poorly and to lack universality; removing Retinex decomposition leads to noise amplification and color distortion; removing histogram guidance destabilizes brightness correction. On the diffusion side, replacing AGCM output with a simple gamma-corrected image degrades performance, conditioning the diffusion model on the original degraded image hurts exposure stability, and removing $L'_d$0 causes diffusion drift and artifacts. The paper therefore treats AGCM, histogram-guided regularization, conditioning on the illumination-corrected state, and perturbed consistency as mutually reinforcing rather than interchangeable components (Jiang et al., 12 May 2026).

The broader significance of ZeroIDIR lies in how it aligns model structure with the decomposition of the restoration problem. Exposure correction is handled explicitly by AGCM rather than by a diffusion model that is described as biased in low-frequency generation, while reconstruction and denoising are delegated to diffusion, where diffusion models are described as strong. This suggests that ZeroIDIR is best understood not as a generic unsupervised enhancer, but as a structured zero-reference IDIR system in which the intermediate state $L'_d$1 provides the bridge between illumination modeling and generative reconstruction (Jiang et al., 12 May 2026).

The paper does not present an extended explicit limitations section, but it implies several constraints. The framework still relies on Retinex decomposition, which may be imperfect in extremely complex scenes. The histogram prior is empirical, so a domain that differs strongly from the collected natural-scene distribution may be less well matched. The diffusion stage is computationally heavier than purely feed-forward enhancement models. Although the method is zero-reference, it is still trained on a curated set of low-quality images, so unusual degradations may remain difficult. These caveats place ZeroIDIR within the broader class of unsupervised restoration methods that improve generalization relative to supervised training while retaining dependencies on architectural priors and training-domain coverage (Jiang et al., 12 May 2026).