CustomEnhancer: Personalized Image Enhancement

• CustomEnhancer is a framework for personalized image enhancement that conditions outputs on user-defined parameters such as brightness, chromaticity, and noise.
• It integrates diverse methodologies including attribute-level controls, global parameter prediction, and content-aware style modeling to cater to varied enhancement needs.
• The approach replaces monolithic image-to-image mappings with structured, interpretable controls, improving adaptability for low-light, color transformations, and recognition-driven tasks.

CustomEnhancer is an Editor’s term for enhancement systems whose output is conditioned by user preference, reference exemplars, controllable parameters, or downstream objectives rather than by a single fixed image-to-image mapping. In current research, this design appears in intelligible low-light enhancement controlled by brightness, chromaticity, and noise (Zheng et al., 2022), controllable image signal processing through a low-dimensional task vector (Kim et al., 2022), global parameterized color transformation learned in supervised and unsupervised settings (Chai et al., 2019), content-aware personalization from user preference galleries (Kosugi et al., 2023), and a later diffusion plug-in that explicitly adopts the name “CustomEnhancer” for personalized photo generation (Ren et al., 25 Sep 2025). The common thread is not a single architecture but the replacement of monolithic enhancement with explicit conditioning, structured control, or set-conditioned inference.

1. Conceptual scope and problem formulation

A central premise across this literature is that enhancement is frequently one-to-many rather than one-to-one. Low-light image enhancement is described as “an inherently subjective process whose targets vary with the user’s aesthetic,” and CRISP similarly argues that a single photograph of auto-enhancement is challenging to fit user preferences, which are subtle and even changeable (Zheng et al., 2022, Kim et al., 2022). Personalized image enhancement extends this argument further by rejecting the assumption that one user has only one fixed style for all images, and instead targets content-aware personalization, where landscapes, portraits, and other contents may require different preferred edits from the same user (Kosugi et al., 2023).

A second recurring premise is that customization can be achieved through very different conditioning interfaces. Some systems expose interpretable image attributes, some predict compact global transformation parameters, some infer styles from example galleries, and some optimize enhancement for a non-perceptual target such as classification accuracy. This suggests that “customization” is better understood as a conditioning strategy than as a specific model family.

Formulation	Conditioning signal	Representative papers
Interpretable low-light personalization	brightness, chromaticity, noise; unpaired reference images or explicit user-set parameters	(Zheng et al., 2022)
Global controllable retouching	task vector $\boldsymbol{t}\in\mathbb{R}^D$; 19 ISP parameters	(Kim et al., 2022)
Parameterized color translation	$\theta_x\in\mathbb{R}^{10\times 3}$ for a global quadratic RGB mapping	(Chai et al., 2019)
Content-aware personalized enhancement	preferred image set; masked style modeling	(Kosugi et al., 2023)
Zero-shot diffusion personalization enhancement	scene-rich SDXL representations, face swapping, ResInversion, training-free controls	(Ren et al., 25 Sep 2025)

The literature also separates perceptual enhancement from other objectives. “Classification Driven Dynamic Image Enhancement” explicitly argues that enhancement for machine recognition should not be equated with enhancement for human perceptual quality; the former should amplify image structures and texture patterns that improve discriminative features and class separation (Sharma et al., 2017). In compressed-image restoration, the objective is framed as aligning the enhancement domain with the raw domain while explicitly distancing it from the compression domain (Xing et al., 2024).

2. Interpretable attribute-level personalization

The clearest interpretable formulation is the intelligible unsupervised personalized enhancer, or iUP-Enhancer, for low-light images (Zheng et al., 2022). It decomposes personalization into three user-friendly attributions—brightness, chromaticity, and noise—rather than into an opaque latent style embedding. The pipeline uses Retinex decomposition with a pre-trained Decom-Net to obtain illumination and reflectance components, then applies three specialized modules: Brighten-Net for illumination control, Enhance-Net for chromaticity transfer, and Denoise-Net for controllable denoising. The final enhanced image is reconstructed as

$I_{en} = L_{en} \cdot R_{de}.$

Each attribute is tied to a separate correlation mechanism. Brightness is controlled by the histogram of the reference illumination map; chromaticity is controlled by hue and saturation statistics in HSV space on reflectance maps; noise is controlled by the correlation between estimated noise-level maps. The training objective is correspondingly decomposed into histogram, Gram, chromaticity, spatial consistency, and perceptual terms: $$L = L_{his} + w_{1} L_{gram} + w_{2} L_{chr} + w_{3} L_{spa} + w_{4} L_{per},$$ with

$$w_{1} = 0.0001,\quad w_{2} = 0.01,\quad w_{3} = 0.03,\quad w_{4} = 0.001.$$

The practical significance of this design is that personalization can be driven in several modes without paired scene-matched supervision. A single reference image may provide the desired brightness histogram, hue-saturation statistics, and noise profile; multiple references may be averaged; and cross-attribute references allow brightness from one image, chromaticity from another, and noise level from a third. The same framework also supports direct parameter control via $\gamma$ for brightness, $[C_h,C_s]$ for chromaticity, and $C_n$ for noise. In the authors’ formulation, “unsupervised” means that low-light inputs and normal-light references are unpaired, not that no data are used.

This attribute-level transparency corrects a common misconception about personalization. It does not make the enhancer a fully general editor. The method explicitly supports brightness, chromaticity, and noise, but does not explicitly support arbitrary aesthetic changes such as local relighting, selective region edits, sharpness or texture style, or semantic edits. The paper therefore presents a structured customizable enhancer rather than a universal aesthetic manipulation system. Its reported performance is competitive on LOL, FiveK, and Huawei, and its user study with 50 human subjects records 392 votes, substantially above the compared baselines.

3. Parametric control through global transformations and ISP pipelines

A different CustomEnhancer line replaces direct image synthesis with prediction of compact global parameters. In “Supervised and Unsupervised Learning of Parameterized Color Enhancement,” the generator predicts a matrix

$\theta_x \in \mathbb{R}^{10\times 3}$

from a $256\times256$ proxy image, and then applies a global quadratic color mapping to every pixel of the full-resolution input (Chai et al., 2019). The basis vector is

$\theta_x\in\mathbb{R}^{10\times 3}$0

and the residual form of the transform is

$\theta_x\in\mathbb{R}^{10\times 3}$1

This parameterization preserves spatial structure while allowing channel gains, second-order tone shaping, cross-channel interactions, and bias terms. In the unpaired setting, the same global transform family is trained in a two-way GAN with cycle consistency.

CRISP, or ContRollable Image Signal Processing, moves even closer to an explicit enhancement interface by decoding a low-dimensional latent task vector into the parameters of a classical ISP pipeline (Kim et al., 2022). The pipeline contains five ordered stages—global scaling or digital gain, white balance, color correction, gamma correction, and tone mapping—and uses only 19 parameters in total. The test-time control variable is

$\theta_x\in\mathbb{R}^{10\times 3}$2

with the paper choosing $\theta_x\in\mathbb{R}^{10\times 3}$3. The image formation rule is summarized as

$\theta_x\in\mathbb{R}^{10\times 3}$4

at test time, while training replaces $\theta_x\in\mathbb{R}^{10\times 3}$5 with parameters inferred from paired low-quality and high-quality images.

The two papers share several properties. Both predict compact parameter sets rather than dense output images; both support arbitrary output resolution because prediction occurs on a low-resolution proxy while the final transform is applied at full resolution; and both retain strong structural priors that reduce hallucination and spatial artifacts. Their limits are equally clear. Because the transformation is global, neither framework can natively express sky-only edits, local dodge-and-burn, or object-selective enhancement. CRISP makes this explicit by treating controllability as repeated navigation in a small latent space, even allowing a greedy search over $\theta_x\in\mathbb{R}^{10\times 3}$6 that can surpass 3DLUT PSNR in five steps under one evaluation setup.

4. Content-aware personalization and local enhancement fields

Content-aware personalization enters the literature most directly in masked style modeling, proposed by Kosugi and Yamasaki (Kosugi et al., 2023). Instead of averaging all preference examples into one user vector, the method represents each example as a content embedding and a style embedding, with style defined as a residual: $\theta_x\in\mathbb{R}^{10\times 3}$7 A Transformer then predicts the missing style for a query image by masking the query-style slot and conditioning on the user’s other content-style pairs: $\theta_x\in\mathbb{R}^{10\times 3}$8 The final enhancer is a PieNet-style modified U-Net conditioned on this predicted style. The method is trained in two stages and is paired with a Flickr-based training scheme comprising 100 images from each of 1,000 users, converted into pseudo input-retouched pairs using a degrading model. Reported results show best performance on both FiveK and PPR10K, and the user study records 79.5\% / 20.5\% preference for the proposed method over PieNet.

A complementary route to content-aware customization is local parameter fields rather than gallery-conditioned style prediction. “Crowd-Powered Photo Enhancement Featuring an Active Learning Based Local Filter” models enhancement as a dense per-pixel parameter map

$\theta_x\in\mathbb{R}^{10\times 3}$9

but predicts it from only a few key pixels selected by EMOC active learning (Kosugi et al., 2023). With Gaussian Process Regression, the full field can be written as

$I_{en} = L_{en} \cdot R_{de}.$0

or, equivalently, as a weighted combination of key-pixel parameter vectors. Pixel features are formed from spatial coordinates $I_{en} = L_{en} \cdot R_{de}.$1 and an illumination value $I_{en} = L_{en} \cdot R_{de}.$2 computed by LIME, and the reported best operating point is $I_{en} = L_{en} \cdot R_{de}.$3 key pixels.

This local-filter literature addresses a different misconception: content awareness need not require explicit semantic segmentation. In the reported system, workers optimize only brightness, saturation, and contrast at selected key pixels through Sequential Line Search, using a single slider at a time, while the regression model transfers those edits to the rest of the image. The approach is not real-time—enhancement of 5 target images finishes in about five hours under the crowd workflow—but it shows that sparse, interpretable local parameterization can outperform both global filters and several existing enhancement baselines in user preference.

5. Objective-driven enhancement, priors, and domain alignment

Some CustomEnhancer systems are customized not to a user style but to an external objective. “Classification Driven Dynamic Image Enhancement” integrates enhancement into an end-to-end recognition stack, with EnhanceNet generating image-specific luminance filters and ClassNet consuming the enhanced image (Sharma et al., 2017). The filter parameters are dynamic: $I_{en} = L_{en} \cdot R_{de}.$4 and training couples reconstruction to classification: $I_{en} = L_{en} \cdot R_{de}.$5 The multi-branch Dyn-CNN variant uses five enhancement methods—WLS, BF, GF, HistEq, and Imsharp—plus weighted fusion. The guiding claim is explicit: enhancement should be optimized for recognition, not necessarily human perception.

In compressed-image enhancement, the dominant concern becomes enhancement bias toward the compression domain. The proposed remedy is a conditional discriminator

$I_{en} = L_{en} \cdot R_{de}.$6

that receives the compressed image as a condition, together with a domain-divergence regularizer

$I_{en} = L_{en} \cdot R_{de}.$7

which enforces that enhanced images should be at least as far from the compressed image as the raw image is in VGG-feature space (Xing et al., 2024). This design is explicitly training-time only and is reported to incur no inference overhead. On BPG at QP 37, adding the method to Real-ESRGAN changes FID / PSNR from $I_{en} = L_{en} \cdot R_{de}.$8 to $I_{en} = L_{en} \cdot R_{de}.$9.

A third strand seeks interpretability through customized learnable priors inside a Retinex-style unfolding model. CUE models the image as

$$L = L_{his} + w_{1} L_{gram} + w_{2} L_{chr} + w_{3} L_{spa} + w_{4} L_{per},$$0

and introduces two MAE-based priors: an illumination prior trained toward bilateral-filtered illumination maps, inserted into the illumination proximal operator, and a noise prior trained toward HOG gradient representations, used as a regularization term

$$L = L_{his} + w_{1} L_{gram} + w_{2} L_{chr} + w_{3} L_{spa} + w_{4} L_{per},$$1

(Zheng et al., 2023). The model reports best performance at $$L = L_{his} + w_{1} L_{gram} + w_{2} L_{chr} + w_{3} L_{spa} + w_{4} L_{per},$$2 unfolding stages and has 0.25M parameters. This prior-based formulation is more transparent than unconstrained end-to-end mapping, but it still does not become fully analytic; the reflectance and noise proximals remain small CNNs.

6. Diffusion-era CustomEnhancer

A later diffusion-based framework explicitly names itself CustomEnhancer and redefines enhancement as a plug-in for personalized photo generation rather than classical retouching (Ren et al., 25 Sep 2025). The target systems are training-based personalization models such as PhotoMaker and InstantID, which the paper characterizes as suffering from degraded scenes, insufficient control, and suboptimal perceptual identity. The solution is a zero-shot enhancement pipeline that combines scene-rich representations from pretrained SDXL, perceptual identity cues from face swapping, and latent fusion inside the personalized model.

The pipeline is staged. SDXL first generates an identity-agnostic scene image under the user prompt. A face-swapping model then injects the reference identity into that image. The resulting image is inverted into the personalized model using ResInversion, which replaces Null-text Inversion’s optimization loop with residual-noise rectification: $$L = L_{his} + w_{1} L_{gram} + w_{2} L_{chr} + w_{3} L_{spa} + w_{4} L_{per},$$3 Generation then proceeds through triple-flow fused PerGeneration or BiMD, which fuses a pivot DDIM inversion flow, a backward reconstruction flow, and a forward customized generation flow. The paper’s “training-free control” means that existing SDXL controls such as OpenPose and Canny are reused without controller retraining for each personalized model.

This is the most explicit use of the word “CustomEnhancer” in the literature block, but the meaning is specialized. It is not a generic photo retoucher; it is an augmentation framework for personalized diffusion models. Reported gains on CelebA-HQ include, for the PhotoMaker plug-in, Scene Div. 60.0 vs 52.2, Face Sim. 67.3 vs 55.7, and FID 416.3 vs 473.9. The runtime advantage of ResInversion is also substantial: PhotoMaker + NTI: 7283 s versus PhotoMaker + ResInversion: 102 s. At the same time, “training-free” does not imply architectural simplicity. The framework still requires auxiliary SDXL generation, face swapping, inversion, and latent fusion.

Taken together, these works indicate that CustomEnhancer is best understood as a family of enhancement systems organized around explicit conditioning. In one branch, customization is attribute-level and interpretable; in another, it is a compact parameterization over global transforms or ISP operators; in another, it is set-conditioned inference over user galleries or sparse local control points; and in another, it is a plug-in objective or prior that changes what “enhancement” is meant to optimize. The most stable technical lesson is that customization becomes practically useful when the conditioning signal—reference image, task vector, preference set, downstream objective, or latent control flow—is made structurally legible rather than hidden inside a single undifferentiated mapping.