Unsupervised Exposure Correction (UEC)

• Unsupervised Exposure Correction (UEC) is a set of methods that restore properly exposed images from under-, over-, or mixed-exposure inputs using automatically generated pseudo-labels.
• Techniques in UEC include multi-exposure fusion, RAW-to-sRGB ISP emulation, and semantic guidance, all designed to bypass laborious human retouching while preserving fine details.
• UEC methods improve both low-level feature preservation and downstream applications such as face detection and semantic segmentation by ensuring radiometric consistency and enhanced contrast.

Searching arXiv for the specified paper and closely related unsupervised exposure correction work to ground the article in current literature. arXiv search query: "(Li et al., 8 Nov 2025) LoopExpose unsupervised exposure correction PSENet (Cui et al., 23 Jul 2025) UNICE (Wu et al., 27 Jan 2026)" Unsupervised exposure correction (UEC) denotes exposure-correction methods that avoid human-annotated well-exposed targets while still seeking to restore detail, contrast, and radiometric plausibility in under-exposed, over-exposed, or mixed-exposure imagery. In recent arXiv literature, UEC has been instantiated through pseudo label refinement from multi-exposure fusion, progressive pseudo-ground-truth synthesis, RAW-to-sRGB ISP emulation, closed-form iterative compensation, HDR-derived pseudo supervision, and semantic-prompt guidance from foundation models (Li et al., 8 Nov 2025, Nguyen et al., 2022, Cui et al., 23 Jul 2025, Ma et al., 2022, Cui et al., 23 Jul 2025, Wu et al., 27 Jan 2026). The term therefore refers both to a research area and, in one case, to a specific 2025 method titled "Unsupervised Exposure Correction" (Cui et al., 23 Jul 2025).

1. Problem setting and scope

Exposure correction is framed in the cited works as the restoration of visually usable images from inputs whose radiometry is too low, too high, or spatially imbalanced. "LoopExpose" formulates the task for an arbitrary-length sequence of under/over-exposed RGB images, whereas UNICE formulates it as mapping a single 8-bit sRGB image $I_{in}$ to an enhanced 8-bit sRGB image $I_{out}$ through pseudo multi-exposure sequence generation and fusion (Li et al., 8 Nov 2025, Cui et al., 23 Jul 2025). "PSENet" addresses "extreme-light image enhancement," explicitly targeting both under-exposure and over-exposure rather than only low-light cases (Nguyen et al., 2022). "CLIP-Guided Unsupervised Semantic-Aware Exposure Correction" adds the claim that improper exposure often causes severe loss of details, color distortion, and reduced contrast, and attributes color-shift artifacts partly to the ignorance of object-wise regional semantic information (Wu et al., 27 Jan 2026).

The principal motivation across these works is the impracticality of paired supervision. "LoopExpose" states that supervised learning has achieved significant progress but relies heavily on large-scale labeled datasets that are difficult to obtain in practical scenarios (Li et al., 8 Nov 2025). The 2025 "Unsupervised Exposure Correction" paper is more specific: expert-retouched targets are labor-intensive, introduce individual style biases, and can reduce cross-dataset generalization; large image-to-image models may also degrade low-level features such as edges and fine textures (Cui et al., 23 Jul 2025). "Practical Exposure Correction: Great Truths Are Always Simple" takes an even stronger position by removing learning entirely and recasting exposure correction as estimation of a single exposure-sensitive compensation map $\bm{\epsilon}(\mathbf y)$ in a linear add/subtract model (Ma et al., 2022).

A useful distinction within UEC is between single-image and multi-exposure settings. LoopExpose operates on exposure sequences and explicitly couples single-exposure correction (SEC) with multi-exposure fusion (MEF), while PSENet, PEC, the 2025 UEC model, UNICE, and the CLIP-guided method all begin from a single sRGB image and construct additional supervisory structure internally (Li et al., 8 Nov 2025, Nguyen et al., 2022, Ma et al., 2022, Cui et al., 23 Jul 2025, Cui et al., 23 Jul 2025, Wu et al., 27 Jan 2026). This suggests that contemporary UEC is less defined by input cardinality than by the rejection of manually retouched ground truth.

2. Pseudo supervision and target construction

A defining property of UEC is the replacement of human-retouched references with automatically generated targets. The mechanisms differ substantially.

LoopExpose uses a classical Mertens multi-exposure fusion operator $\mathcal M$ to construct pseudo-labels. In warm-up, the initial pseudo-label is produced only from raw exposures, $Y^{(0)}=\mathcal M(I)$; during joint optimization, the corrected outputs are fed back into fusion so that

$$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$

creating the paper’s self-reinforcing loop in which better $E$ yields better $Y$ and vice versa (Li et al., 8 Nov 2025).

PSENet constructs a bank of darker and brighter reference images from a single input by invert-gamma mapping, then scores each candidate using non-reference measures of well-exposedness, local contrast, and color saturation. Its pseudo-ground-truth is obtained by pixel-wise selection:

$$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$

Because the candidate set includes the previous-epoch output $\hat Y_{prev}$, the pseudo-GT can only improve or stay equal from epoch to epoch, which motivates the term Progressive Self-Enhancement (Nguyen et al., 2022).

The 2025 UEC model dispenses with pseudo-label synthesis from the observed image itself and instead exploits an emulated ISP pipeline. Starting from RAW, it applies calibrated gain for $I_{out}$0, followed by demosaicing, white balance, color-space conversion, and gamma correction, producing exposure brackets with known scalar EV shifts. Training then uses image pairs or triplets sampled from these brackets as pseudo-paired supervision (Cui et al., 23 Jul 2025).

UNICE scales this RAW-based idea to 46,928 HDR raw images rendered into 328,496 sRGB images across $I_{out}$1. It forms pseudo sRGB ground-truths by applying several classical/fusion algorithms to MES triplets, ensembling them by NR-IQA ranking, and filtering out poor results with $I_{out}$2, leaving 328,496 input-GT pairs (Cui et al., 23 Jul 2025).

The CLIP-guided semantic-aware method replaces handcrafted quality scores and physical brackets with prompt-conditioned pseudo-GT generation. It fine-tunes three text prompts for well-, under-, and over-exposed images, compares CLIP similarities $I_{out}$3 and $I_{out}$4, and then applies gamma-brightening or gamma-darkening accordingly. It can further refine $I_{out}$5 by maximizing similarity to the well-exposed prompt in CLIP’s joint space (Wu et al., 27 Jan 2026).

These strategies can be summarized as follows.

Method	Pseudo-supervision source	Core mechanism
LoopExpose (Li et al., 8 Nov 2025)	Mertens MEF pseudo-labels	Nested feedback loop
PSENet (Nguyen et al., 2022)	Progressive pseudo-GT bank	Pixel-wise quality selection
PEC (Ma et al., 2022)	None beyond input image	Closed-form compensation
UEC (Cui et al., 23 Jul 2025)	Emulated ISP exposure brackets	Known EV-shift pseudo-pairs
UNICE (Cui et al., 23 Jul 2025)	HDR-rendered MES + MEF pseudo-GTs	Two-stage MES generation and fusion
Semantic-aware method (Wu et al., 27 Jan 2026)	CLIP-guided pseudo-GT	Prompt-conditioned gamma tuning

3. Model families and optimization regimes

LoopExpose is organized as a bi-level optimization. The upper level updates a correction model with parameters $I_{out}$6 by minimizing a supervised loss against pseudo-labels together with a luminance constraint; the lower level refines pseudo-labels with fixed rule-based fusion. Training is divided into warm-up, with decayed learning rate and static $I_{out}$7, and joint optimization, with constant learning rate and dynamic pseudo-label updates (Li et al., 8 Nov 2025). Architecturally, it uses a Luminance-Aware Network in encoder-decoder form, an Adaptive 3D-LUT Module, and attention-based fusion of luminance and color paths (Li et al., 8 Nov 2025).

PSENet uses a lightweight U-Net-style encoder-decoder that predicts a per-pixel gamma map $I_{out}$8 and reconstructs the enhanced image by

$I_{out}$9

Its encoder uses MobileNet-V2 blocks, while the decoder is symmetric with skip-connections and ends with a MobileNet-V3 layer to regress three gamma channels (Nguyen et al., 2022).

PEC is structurally distinct because it is zero-training and zero-learning. Its central adversarial generator is

$\bm{\epsilon}(\mathbf y)$0

which is axisymmetric and mid-tone emphasizing. For under-exposure, it iterates

$\bm{\epsilon}(\mathbf y)$1

within a segmented shrinkage scheme of up to $\bm{\epsilon}(\mathbf y)$2 built-in blocks and per-block iteration counts $\bm{\epsilon}(\mathbf y)$3 (Ma et al., 2022). The same machinery is used for over-exposure by changing the sign.

The 2025 paper titled "Unsupervised Exposure Correction" introduces a small learned radiometric corrector. Given an exposure-difference estimate $\bm{\epsilon}(\mathbf y)$4, its transformation block performs a pixelwise interpolation between linear scaling and a learned nonlinear curve:

$\bm{\epsilon}(\mathbf y)$5

where $\bm{\epsilon}(\mathbf y)$6 is predicted from $\bm{\epsilon}(\mathbf y)$7 by a small MLP and $\bm{\epsilon}(\mathbf y)$8 is implemented by three sequential $\bm{\epsilon}(\mathbf y)$9 convolutions with ReLU. The block is iterated three times. The full system comprises an Exposure Feature Encoder, a Difference Predictor, and an Exposure Corrector, for a total of 19,388 parameters (Cui et al., 23 Jul 2025).

UNICE adopts a substantially larger two-stage architecture. MES-Net maps a single image to a pseudo multi-exposure sequence using a pretrained SD-Turbo diffusion model with frozen original weights and inserted LoRA adapters; MEF-Net then predicts pixelwise weight maps, forms an implicit HDR representation

$\mathcal M$0

and refines it with another one-step diffusion model to produce the final enhanced output (Cui et al., 23 Jul 2025).

The CLIP-guided semantic-aware model is an encoder-decoder built from Semantics-Informed Mamba Reconstruction blocks. Each block first applies Adaptive Semantic-Aware Fusion, which injects FastSAM semantic features into the image feature space, then a Residual Spatial Mamba Group, which augments a Vision Mamba Module with spatial attention (Wu et al., 27 Jan 2026). This architecture directly addresses the claim that object-wise semantics matter for exposure correction under mixed lighting.

4. Objective functions, priors, and self-supervised constraints

The losses used in UEC are not uniform; they encode different hypotheses about what must be preserved when explicit ground truth is absent.

LoopExpose defines

$\mathcal M$1

with

$\mathcal M$2

Its distinctive self-supervised term is the Luminance Ranking Loss

$\mathcal M$3

where $\mathcal M$4 and the input exposures are sorted from darkest to brightest. This enforces preservation of relative luminance ordering without ground truth (Li et al., 8 Nov 2025).

PSENet uses a simpler training objective. The network is trained with a reconstruction MSE toward the pseudo-GT and a total variation regularizer on the gamma map:

$\mathcal M$5

Here the prior is not only smoothness in $\mathcal M$6 but also the design of pseudo-GT generation itself, because the reconstruction target is chosen from a bank optimized for well-exposedness, contrast, and saturation (Nguyen et al., 2022).

The 2025 UEC model formalizes three losses: reconstruction $\mathcal M$7, exposure-consistency $\mathcal M$8 under the "Monopoly Principle," and detail-preservation $\mathcal M$9:

$Y^{(0)}=\mathcal M(I)$0

with $Y^{(0)}=\mathcal M(I)$1 and $Y^{(0)}=\mathcal M(I)$2. The exposure-consistency term enforces monotonicity between outputs derived from reference images with different EVs, while the detail-preservation term is a total-variation-style gradient penalty (Cui et al., 23 Jul 2025).

UNICE uses only two reconstruction losses, one for MES generation and one for fusion:

$Y^{(0)}=\mathcal M(I)$3

$Y^{(0)}=\mathcal M(I)$4

The paper explicitly notes that no explicit perceptual or smoothness losses are used; all supervision is via automatically generated targets (Cui et al., 23 Jul 2025).

The CLIP-guided semantic-aware method combines pixel fidelity, chromatic fidelity, and semantic/prompt consistency:

$Y^{(0)}=\mathcal M(I)$5

where $Y^{(0)}=\mathcal M(I)$6. Its Semantic Feature Consistency is defined over FastSAM features using both feature distances and Gram distances, while Image-Prompt Alignment pushes the corrected image toward the well-exposed prompt and away from the under-/over-exposed prompts in CLIP space (Wu et al., 27 Jan 2026).

PEC, by contrast, uses no explicit learning loss at all. Its regularization is implicit in the bounded, axisymmetric form of $Y^{(0)}=\mathcal M(I)$7 and in the shrinkage behavior of repeated updates; the paper explicitly states that there is no explicit Retinex prior, no data-driven loss, and no external denoiser (Ma et al., 2022). This contrast is important: within UEC, "unsupervised" covers both learned self-supervision and fully closed-form optimization-free correction.

5. Benchmarks, quantitative behavior, and efficiency

The empirical landscape is heterogeneous because methods are evaluated on different datasets and with different metric suites. LoopExpose reports results on SeqMSEC and SeqRadio derived from MSEC and Radiometry512, using PSNR and SSIM. On MSEC SEC comparison, among unsupervised methods it achieves approximately $Y^{(0)}=\mathcal M(I)$8 versus the next-best UEC at $Y^{(0)}=\mathcal M(I)$9, and it nearly matches supervised CoTF at $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$0. On Radiometry512, it reports $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$1 versus UEC at $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$2. For MEF, Mertens scores $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$3 on MSEC and $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$4 on Radiometry512, while LoopExpose with correction plus fusion reaches $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$5 and $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$6 (Li et al., 8 Nov 2025).

PSENet is evaluated on SICE, Afifi, and LOL. On SICE, the unsupervised result is PSNR $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$7 and SSIM $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$8, compared with ZeroDCE at $$Y^{(t+1)}=\mathcal M\bigl(I_1^{obs},\ldots,I_N^{obs},E_1^{(t)},\ldots,E_N^{(t)}\bigr),$$9 and EnlightenGAN at $E$0. On Afifi, PSENet reports under-exposure $E$1, over-exposure $E$2, and full $E$3. On LOL generalization, it reaches $E$4 versus ZeroDCE at $E$5 (Nguyen et al., 2022).

PEC emphasizes no-reference quality and speed. On Exposure-Errors under-exposure, it is among the top three in full-reference metrics at approximately PSNR $E$6 dB and SSIM $E$7, while being best on LOE at approximately $E$8 and NIQE at approximately $E$9; on over-exposure it is again best in LOE at approximately $Y$0 and NIQE at approximately $Y$1. Runtime on a GeForce RTX 2080Ti is reported as $Y$2-$Y$3 s for 720p-2K images, and on an i7-8700K CPU as $Y$4-$Y$5 s per 2K image (Ma et al., 2022).

The 2025 UEC method reports on MSEC, LOL generalization, and its Radiometry Correction Dataset. On MSEC test, it achieves PSNR $Y$6 and SSIM $Y$7; on LOL when trained on MSEC, it reports $Y$8, exceeding Afifi $Y$9, ECM [E→LOL] $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$0, and ECM [R→LOL] $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$1. On RCD, the per-EV average is $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$2 versus ECM at $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$3. The same paper reports real-time 4K coverage at $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$4 ms on GPU and $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$5 ms on CPU with only 19k parameters (Cui et al., 23 Jul 2025).

UNICE reports broader cross-task and cross-dataset evaluation. For EC trained on MSEC and tested on SICE, it obtains PSNR $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$6 dB, SSIM $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$7, and NIQE $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$8, compared with LCDP at $$Q_n(x,y)=\frac{C_n(x,y)S_n(x,y)}{E_n(x,y)+\epsilon}, \qquad T(x,y)=Y_{n^*}(x,y), \quad n^*=\arg\max_n Q_n(x,y).$$9, $\hat Y_{prev}$0, and $\hat Y_{prev}$1. It also reports no-reference comparisons against manually captured GTs, including EC on SICE where NIQE decreases from $\hat Y_{prev}$2 to $\hat Y_{prev}$3, PI from $\hat Y_{prev}$4 to $\hat Y_{prev}$5, and ARNIQA increases from $\hat Y_{prev}$6 to $\hat Y_{prev}$7 (Cui et al., 23 Jul 2025). The cost of this generality is model scale: approximately 965 M parameters, approximately 3348 G FLOPs, and $\hat Y_{prev}$8 FPS at $\hat Y_{prev}$9 (Cui et al., 23 Jul 2025).

The CLIP-guided semantic-aware model reports average PSNR/SSIM of $I_{out}$00 on MSEC and $I_{out}$01 on SICE, with first- or second-place results in LPIPS, BRISQUE, and NIMA across both datasets. Its ablations show average SICE PSNR/SSIM dropping from $I_{out}$02 to $I_{out}$03 without ASF, to $I_{out}$04 without SpatialAttn, and to $I_{out}$05 without the SPC loss (Wu et al., 27 Jan 2026).

Taken together, these results show that UEC cannot be summarized by a single Pareto frontier. Some methods optimize SEC/MEF quality on controlled exposure sequences, some optimize no-reference quality and runtime, some optimize cross-task generalization, and some explicitly trade parameter count for semantic richness.

6. Downstream impact, conceptual boundaries, and open problems

A recurrent claim in this literature is that exposure correction should be evaluated beyond appearance metrics. The 2025 UEC paper directly measures edge-detection quality using LDC and reports average PSNR $I_{out}$06 dB and $I_{out}$07 versus ECM’s $I_{out}$08 dB and $I_{out}$09, stating that ECM can degrade edges below the uncorrected input whereas UEC consistently enhances them (Cui et al., 23 Jul 2025). PSENet reports improved true-positive rates for face detection on synthetically exposed FDDB after preprocessing, and PEC reports state-of-the-art mAP and mIoU among enhancers when plugged into a face detector or semantic segmentation network (Nguyen et al., 2022, Ma et al., 2022). The CLIP-guided paper includes DarkFace as a face-detection test set, reflecting the same downstream concern (Wu et al., 27 Jan 2026).

One conceptual boundary concerns the meaning of "unsupervised." In this literature, the term does not imply the absence of supervision signals. Rather, it indicates the absence of manually edited well-exposed targets. Supervision is instead supplied by pseudo-labels from MEF, pseudo-GTs from synthetic exposure banks, RAW-derived EV brackets, NR-IQA-ranked HDR renderings, or frozen priors from FastSAM and CLIP (Li et al., 8 Nov 2025, Nguyen et al., 2022, Cui et al., 23 Jul 2025, Cui et al., 23 Jul 2025, Wu et al., 27 Jan 2026). PEC is the principal exception because it is fully closed-form (Ma et al., 2022).

Another boundary concerns radiometric correction versus creative retouching. The 2025 UEC paper explicitly notes that, because it adjusts only radiometry, it cannot perform more elaborate creative color- or tone-style retouching; it also states that, in extreme under/over exposures, texture information may be physically lost (Cui et al., 23 Jul 2025). LoopExpose identifies future work in replacing fixed fusion $I_{out}$10 with a learned MEF module and in more advanced stabilization of bilevel updates (Li et al., 8 Nov 2025). UNICE pursues universality through large-scale HDR-derived pseudo supervision (Cui et al., 23 Jul 2025), while the semantic-aware approach pursues region-wise correction through object-level priors (Wu et al., 27 Jan 2026). A plausible implication is that current UEC research is bifurcating into lightweight radiometric correctors optimized for detail preservation and efficiency, and heavier universal or semantic systems optimized for broader generalization and scene awareness.

Across these directions, the core problem remains stable: to recover exposure and contrast without manual target construction, while preserving the low-level and semantic information on which subsequent vision processing depends.