UNICE: Universal Image Contrast Enhancer

• The paper introduces UNICE as a two-stage, diffusion-based framework that addresses multiple exposure-related tasks without relying on human-annotated ground truths.
• It employs HDR raw imagery to synthesize multi-exposure sequences and uses learned weight maps for effective image fusion, achieving superior quality metrics.
• Evaluation demonstrates UNICE’s strong cross-task and cross-dataset generalization, often surpassing manually created ground truths on no-reference image quality metrics.

UNiversal Image Contrast Enhancer (UNICE) is a two-stage, diffusion-prior-based image enhancement framework designed to learn a single model for multiple contrast-related tasks, including low-light image enhancement, exposure correction, backlit image enhancement, and LDR-to-HDR-related enhancement. It is trained entirely without human-annotated ground-truth images. Instead, it uses HDR raw imagery to synthesize multi-exposure sequences (MES) and pseudo sRGB ground-truths via multi-exposure fusion (MEF), then learns first to generate an MES from a single sRGB input and subsequently to fuse that MES into an enhanced image. In the reported evaluation, the resulting model exhibits stronger cross-task and cross-dataset generalization than task-specific methods and, on several no-reference image quality metrics, often exceeds manually created ground truths (Cui et al., 23 Jul 2025).

1. Problem setting and conceptual basis

UNICE is motivated by a recurrent limitation in contemporary image contrast enhancement research: most methods are built for narrowly defined tasks such as under-/over-exposure correction, low-light enhancement, backlit enhancement, or LDR-to-HDR transformation. According to the reported study, such models often depend on modestly sized paired datasets or manually retouched targets, and they generalize poorly not only across tasks but also across datasets within the same task. Typical failure modes outside the training domain include color casts, banding, haloing, and an inability to balance local and global exposure.

The central premise of UNICE is that these tasks share a common latent requirement: exposure and contrast adjustment under limited dynamic range. The framework therefore treats contrast enhancement as a problem that can be substantially simplified if one has access to multiple exposures of the same scene. In that setting, enhancement becomes a fusion problem rather than an extrapolation problem. This framing is explicitly linked to HDR data, because HDR raw imagery can provide high signal fidelity across both shadows and highlights, support principled MEF-based synthesis of high-quality supervisory targets, and enable large-scale supervision without human labeling (Cui et al., 23 Jul 2025).

The paper formalizes image formation by the classical abstraction

$I^i = ISP(H^i), \qquad H^i = S^i \Delta t,$

where $S^i$ is the sensor response and $\Delta t$ is the exposure time. In practice, the emulated ISP is implemented with Adobe Camera Raw SDK and uses DNG metadata for demosaicing, white balance, color correction, tone mapping, and camera response or gamma mapping. This choice is significant because it allows the framework to remain in the sRGB domain while still deriving supervision from HDR raw content.

A plausible implication is that UNICE’s notion of “universal” is not universal image restoration in the broad sense, but universality across a family of exposure- and contrast-dominated enhancement problems. The reported limitations on severe motion blur and heavy noise support that narrower interpretation.

2. HDR-derived supervision and pseudo ground-truth construction

The training corpus is built from HDR raw images aggregated from public datasets: AODRaw, FiveK, HDRP, PASCAL-RAW, PPR10K, RAISE, RAW-NOD. The study reports 48,361 HDR raw images before filtering. These are rendered into 338,527 sRGB images using the emulated ISP, and after quality filtering, 46,928 HDR raws remain, producing 328,496 input–pseudo-ground-truth pairs (Cui et al., 23 Jul 2025).

For each raw image, the framework renders a discrete exposure bracket with

$EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$

These rendered frames form a multi-exposure sequence for the scene. Pseudo ground truths are then created by fusing differently exposed images. Because no single MEF method is consistently best across all scenes, UNICE uses an ensemble comprising FMMEF, GradientMEF, MDO, Mertens, and PerceptualMEF.

The pseudo-ground-truth pipeline follows several explicit steps. For each scene, one triplet is selected with $EV<0$, $EV=0$, and $EV>0$. Five MEF algorithms are applied to obtain five candidates. Each candidate is scored with NR-IQA metrics—NIQE, BRISQUE, PI, and ARNIQA. A winner-frequency is then computed per method from a large scene sample; three randomly chosen candidates are blended with weights proportional to their normalized winner-frequencies; the five original candidates and the blended variants are re-scored; and the best result under average NR-IQA ranking is retained. Finally, scenes with $\text{ARNIQA} < 0.5$ are discarded, removing 1,433 scenes.

When the Mertens method is used inside the ensemble, the per-image weight map is specified as

$$W_i(x) = C_i(x)^{\alpha} \cdot S_i(x)^{\beta} \cdot E_i(x)^{\gamma},$$

where $C_i$ denotes local contrast, $S^i$0 saturation, and $S^i$1 well-exposedness. The well-exposedness term is modeled as

$S^i$2

typically with $S^i$3 and $S^i$4 in normalized intensity units. Final fusion uses multi-resolution Laplacian pyramid blending:

$S^i$5

This data-generation strategy is central to UNICE. It converts HDR raw collections into supervision that is task-agnostic in the sense relevant to exposure and contrast manipulation. The study’s ablation further reports that the ensemble outperforms single MEF methods on held-out scenes, with “Ensambled” obtaining NIQE 4.910, PI 3.474, BRISQUE 17.912, and ARNIQA 0.746.

3. Two-stage architecture

UNICE consists of two separately trained networks: MES-Net, denoted $S^i$6, and MEF-Net, denoted $S^i$7. MES-Net synthesizes an MES from a single sRGB input, while MEF-Net fuses the synthesized MES into the final enhanced image (Cui et al., 23 Jul 2025).

MES-Net models exposure transfer as a style-transfer problem in sRGB space. For a MES derived from one raw image, two renders $S^i$8 and $S^i$9 are randomly sampled. A scalar style code $\Delta t$0 is defined as the mean intensity of the sRGB image and used as a proxy for exposure. During training, the input is rescaled by the ratio of target and source style codes:

$\Delta t$1

The objective is an $\Delta t$2 loss,

$\Delta t$3

Architecturally, MES-Net is a one-step image-to-image diffusion model based on SD-Turbo, with LoRA adapters inserted into “zero-conv” residual paths. The pretrained SD-Turbo backbone is frozen, and only the LoRA-augmented zero-conv components are optimized. At inference time, given an input image with style code $\Delta t$4, the model uses target codes $\Delta t$5 to synthesize a fixed-length MES of three exposures.

MEF-Net performs learned fusion of the synthesized MES. Each of the $\Delta t$6 MES images, with $\Delta t$7 by default, is passed through a weight-estimation CNN $\Delta t$8, implemented as a VGG encoder without fully connected layers and producing a single-channel weight map. If the $\Delta t$9-th image is $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$0 and the corresponding weight map is $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$1, per-pixel normalized weights are defined as

$EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$2

These weights are used to construct an implicit 32-bit HDR image by pixel-wise weighted summation:

$EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$3

That aggregated representation is then refined by a second one-step SD-Turbo-with-LoRA model to produce the final sRGB output $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$4:

$EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$5

The training loss is again $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$6:

$EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$7

The two-stage factorization is important. The first stage handles exposure diversification, while the second stage learns scene-adaptive fusion and perceptual refinement. This suggests that UNICE separates the ill-posed single-image enhancement problem into a sequence generation problem followed by a fusion problem, each with its own inductive bias.

4. Training regimen and implementation profile

Training uses the 328,496 automatically generated input–pseudo-ground-truth pairs derived from the final set of 46,928 HDR raw images. MES-Net and MEF-Net are trained separately because of diffusion-model memory cost. Optimization uses Adam with learning rate $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$8, $EV \in \{-3, -2, -1, 0, +1, +2, +3\}.$9, $EV<0$0, weight decay $EV<0$1, and batch size $EV<0$2, on a single NVIDIA A100 40 GB GPU (Cui et al., 23 Jul 2025).

The LoRA configuration is reported as rank $EV<0$3 for the U-Net portion of SD-Turbo and rank $EV<0$4 for the VAE portion. Inference remains relatively expensive despite the one-step design. At resolution $EV<0$5, the model has approximately $EV<0$6M parameters, requires approximately $EV<0$7G FLOPs, and runs at approximately $EV<0$8 FPS on an NVIDIA A100 40 GB GPU.

The framework nevertheless exposes a simple control mechanism through the style code. A single scalar exposure preference can be used to bias the result brighter or darker by altering the target codes $EV<0$9 employed during MES synthesis. The paper also identifies a fixed $EV=0$0 MES as a strong accuracy–efficiency trade-off, and the default target codes $EV=0$1 are reported to produce well-spaced exposures in practice.

An implementation recipe is explicitly given. It consists of gathering HDR raws with DNG metadata, rendering sRGB images at the stated exposure values through an emulated ISP, constructing triplet MES examples, generating pseudo ground truths by running multiple MEF algorithms and selecting the best-scoring result under NR-IQA, training MES-Net with exposure-style transfer in the sRGB domain, training MEF-Net by fusing a synthesized three-image MES, and then performing test-time inference by computing mean intensity, generating three exposures, estimating weights, fusing, and refining.

5. Evaluation across tasks, datasets, and perceptual criteria

The evaluation spans four task families: Low-Light Image Enhancement (LLIE), Exposure Correction (EC), Backlit Image Enhancement (BIE), and LDR-to-HDR-related enhancement (L2HT). The test sets are UHD-LL and LSRW for LLIE, MSEC and SICE for EC, BAID and Backlit300 for BIE, and HDR-Eye and HDR-Real for L2HT. Full-reference metrics are PSNR, SSIM, LPIPS, and DISTS; no-reference metrics are NIQE, PI, BRISQUE, and ARNIQA (Cui et al., 23 Jul 2025).

Task	Test sets	Metrics
LLIE	UHD-LL, LSRW	PSNR, SSIM, LPIPS, DISTS, NIQE, PI, BRISQUE, ARNIQA
EC	MSEC, SICE	PSNR, SSIM, LPIPS, DISTS, NIQE, PI, BRISQUE, ARNIQA
BIE	BAID, Backlit300	PSNR, SSIM, LPIPS, DISTS, NIQE, PI, BRISQUE, ARNIQA
L2HT	HDR-Eye, HDR-Real	PSNR, SSIM, LPIPS, DISTS, NIQE, PI, BRISQUE, ARNIQA

A notable empirical result is that UNICE frequently surpasses manually created ground truths under NR-IQA. On LLIE/UHD-LL, it achieves NIQE 5.684 versus 6.738 for ground truth, PI 4.100 versus 5.013, and ARNIQA 0.773 versus 0.714. On LLIE/LSRW, the corresponding values are NIQE 5.813 versus 6.165, PI 4.137 versus 4.473, and ARNIQA 0.741 versus 0.709. On EC/SICE, UNICE reports NIQE 5.327 versus 8.050, PI 3.429 versus 5.110, BRISQUE 11.914 versus 32.065, and ARNIQA 0.695 versus 0.666. On EC/MSEC, it is comparable on NIQE and PI but reaches ARNIQA 0.807 versus 0.749 and BRISQUE 19.122 versus 19.856. On BIE/BAID, ARNIQA is 0.817 versus 0.761 and BRISQUE 20.664 versus 22.069. On L2HT/HDR-Eye, ARNIQA is 0.794 versus 0.625 and BRISQUE 14.845 versus 18.299. The paper attributes these outcomes to deficiencies in human-created targets, including darkness, local overexposure, halos, and color casts.

Cross-task generalization is a central claim. Models trained for exposure correction are evaluated on LLIE, BIE, and L2HT. On LLIE/LSRW, UNICE reports PSNR 19.399, SSIM 0.718, LPIPS 0.153, and DISTS 0.129, exceeding LCDPNet at 15.348, 0.552, 0.213, and 0.170, and CSEC at 15.899, 0.659, 0.252, and 0.202. On LLIE/UHD-LL, it reaches PSNR 22.007 and SSIM 0.907, above LCDPNet at 19.148 and 0.835 and RF at 16.383 and 0.755. On BIE/BAID, it reports PSNR 19.058, SSIM 0.876, LPIPS 0.095, and DISTS 0.106, improving on CSEC at 18.476, 0.870, 0.119, and 0.130. On L2HT/HDR-Real, UNICE obtains PSNR 16.818, SSIM 0.721, LPIPS 0.265, and DISTS 0.201, higher than LCDPNet at 14.261, 0.521, 0.359, and 0.259 and CSEC at 14.766, 0.610, 0.343, and 0.260.

Cross-dataset generalization within each task is also emphasized. For LLIE, UNICE reports PSNR 22.007 and SSIM 0.907 on UHD-LL and PSNR 19.399 and SSIM 0.718 on LSRW, exceeding task-specific baselines trained on different LLIE data. For EC, it reaches PSNR 17.509 and SSIM 0.644 on SICE, with leading NR-IQA values NIQE 5.410, PI 3.557, BRISQUE 20.244, and ARNIQA 0.692; on MSEC it reaches PSNR 19.781 and SSIM 0.877, with NIQE 5.257, PI 3.531, BRISQUE 19.122, and ARNIQA 0.807. For BIE, it tops FR and NR metrics on BAID and reports NIQE 4.627, PI 3.123, BRISQUE 16.177, and ARNIQA 0.782 on Backlit300, which lacks ground truth. For L2HT, the paper states that UNICE improves across HDR-Eye and HDR-Real in both FR and NR-IQA relative to HDR-specific and conventional 8-bit enhancers.

User-study results also favor UNICE against the strongest task-specific competitor in each category: 70% for LLIE, 76% for EC, 68% for BIE, and 83% for L2HT. Backbone ablation further shows that using SD-Turbo in both stages gives the best average performance across datasets, with PSNR 21.455, SSIM 0.905, LPIPS 0.091, and DISTS 0.112, outperforming U-Net- and small-CNN-based alternatives.

6. Position within prior work, scope, and limitations

UNICE is situated at the intersection of low-light enhancement, exposure correction, backlit enhancement, LDR-to-HDR-related processing, and MEF. The paper contrasts it with prior task-specific methods that depend on paired supervision and narrowly targeted priors, arguing that such designs limit generalization. UNICE instead uses HDR-based supervision synthesized at scale, with MES and MEF approximating the desired exposure–contrast balance without manual labels (Cui et al., 23 Jul 2025).

Relative to LDR-to-HDR approaches, the framework adopts a distinct stance. Rather than attempting to invert the ISP or reconstruct explicit HDR radiance, which is described as ill-posed and ISP-specific, UNICE learns an sRGB-domain MES and then fuses it. The stated rationale is that a sufficiently diverse sRGB MES is ISP-agnostic yet still broad enough to cover a large dynamic range once fused. In this sense, the model operationalizes signal fidelity through exposure diversity rather than radiometric reconstruction.

Relative to classical MEF, UNICE uses MEF twice: first as a data-generation mechanism, through an ensemble that produces diverse pseudo ground truths, and second as a learned fusion principle, where a neural weight estimator predicts soft per-pixel fusion maps and a diffusion prior refines the weighted aggregate. The theoretical justification given in the paper is that, with an MES, well-exposed pixels exist for almost every region, so the enhancement problem can be reduced to selection and blending of best-exposed content across the sequence.

Several practical limitations are explicitly stated. Despite one-step diffusion, the model remains computationally heavy. Extreme degradations such as severe motion blur or heavy noise are filtered during data generation, but can still challenge the learned fusion at test time. Future work is suggested in lighter backbones, adaptive MES length, joint end-to-end training under memory constraints, and explicit noise or motion modeling in MES generation.

A common misconception would be to interpret the reported superiority over manual ground truth as a general statement that human retouching is inferior. The evidence given is narrower: on the evaluated datasets and no-reference criteria, the learned outputs often score higher, and the paper associates this with artifacts or exposure imbalances in some human-created targets. Another misconception would be to read “universal” as coverage of arbitrary restoration tasks; the study supports universality specifically across exposure- and contrast-centered enhancement regimes. Within that domain, the reported results suggest that large-scale HDR-driven pseudo-supervision combined with one-step diffusion priors is a viable label-free route to a single generalized enhancement model.