Regional Feature Harmonization

• Regional feature harmonization is the process of aligning features within defined regions—such as image foregrounds or anatomical ROIs—to reduce inconsistencies while preserving key attributes.
• It involves partitioning a domain, estimating region-specific statistics (e.g., means, variances), and applying targeted transformations that adjust only the mismatched region.
• Applications span composite image editing, medical imaging, and hydrometeorological studies, with performance measured by metrics like PSNR, MSE, and ROC-AUC.

Regional feature harmonization denotes a family of methods that align features within explicitly defined regions—spatial masks, semantic segments, anatomical ROIs, subgroup partitions, or feature-space subdomains—so that regional incompatibilities are reduced while task-relevant structure is preserved. In image harmonization, the canonical formulation is a composite image $I$ with foreground mask $M_f$ and background mask $M_b=1-M_f$, where the objective is to transform foreground features so that they become compatible with the background or with real harmonious images while leaving the background unchanged (Niu et al., 2023). The same organizing idea appears in medical imaging, where harmonization removes site or acquisition effects from region-derived features while preserving biology (Yang et al., 22 Jul 2025), in regional explanation frameworks that partition feature space to minimize disagreement between attribution methods (Herbinger et al., 30 Jan 2026), and in hydrometeorological regionalization, where standardized time-series features make catchments comparable across climates (Papacharalampous et al., 2022).

1. Core concept and formal scope

Regional feature harmonization is defined by three recurring operations. First, a domain is partitioned into regions. In composite-image harmonization these are usually foreground and background masks, submasks obtained from RGB clustering, or semantic regions inferred by a segmentation prior (Niu et al., 2023). In interactive portrait harmonization, the relevant region is a user-selected reference region $R$ in the background, encoded by a guide mask $M_R$ (Valanarasu et al., 2022). In medical imaging, regions may be cortical ROIs, tract-wise DTI measures, regional fMRI features, myocardial masks, or benign-versus-malignant subgroups (Yang et al., 22 Jul 2025). In explanation theory, the regions are subsets $\Omega_k\subseteq\mathcal{X}$ of feature space itself (Herbinger et al., 30 Jan 2026).

Second, region-specific statistics, embeddings, or relations are estimated. A representative image-harmonization formalization decomposes encoder features $F$ into foreground and background components $F^f$ and $F^b$ using downsampled masks, then expresses mismatch through first- and second-order statistics such as $\mu_{f,l}$, $M_f$0, $M_f$1, and $M_f$2 at layer $M_f$3 (Niu et al., 2023). Other works construct region descriptors by content features, patch statistics $M_f$4, semantic embeddings, masked pooled channel descriptors, or learned style vectors (Zhu et al., 2022). In feature-level harmonization for MRI or radiomics, region-specific variation is modeled as additive and multiplicative batch effects, often through ComBat-style location-scale corrections (Yang et al., 22 Jul 2025).

Third, a region-targeted transformation is applied while preserving a complement. In image harmonization this often means modulating or replacing only foreground features and leaving background features unchanged (Chen et al., 2023). In MRI harmonization, anatomy-fixed modulation or segmentation-conditioned renormalization seeks to change scanner- or site-specific appearance while preserving morphology (Ren et al., 2021). In GRANITE, regionalization is not a pixel operation but a partition of feature space that minimizes a formal disagreement risk,

$M_f$5

thereby harmonizing different explanation methods within each region (Herbinger et al., 30 Jan 2026).

Domain	Region definition	Harmonization target
Image compositing	Foreground/background masks, semantic regions, submasks, reference region $M_f$6	Foreground appearance, color, luminance, texture (Niu et al., 2023)
Generative consistency/editing	Subject masks, facial semantic regions	Identity-consistent or harmonious regional styles (Gaur et al., 31 Jul 2025)
Medical imaging/radiomics	Anatomical ROIs, subgroup partitions, myocardium masks	Site, scanner, and acquisition effects (Yang et al., 22 Jul 2025)
Explainability/regionalization	Feature-space regions $M_f$7, catchments	Explanation agreement or cross-site comparability (Herbinger et al., 30 Jan 2026)

2. Regional feature harmonization in image compositing

The most direct use of the term appears in image harmonization, where global foreground adjustment is treated as insufficient whenever the foreground contains multiple appearance patterns or the background is spatially heterogeneous. A key early formulation matches each foreground location to content-related background regions. “Image Harmonization by Matching Regional References” decomposes harmonization into a deep, low-resolution Locations-to-Location Translation (LTL) and a high-resolution Patches-to-Location Translation (PTL). LTL computes $M_f$8 and fuses the result with foreground tokens, whereas PTL matches foreground content tokens $M_f$9 to background patch content tokens $M_b=1-M_f$0, then transfers regional appearance statistics by $M_b=1-M_f$1. The same system uses residual reconstruction, $M_b=1-M_f$2, to preserve high-frequency detail, and reports $M_b=1-M_f$3 PSNR, $M_b=1-M_f$4 MSE, and $M_b=1-M_f$5 fMSE on iHarmony4, with a wider variant reaching $M_b=1-M_f$6 PSNR and $M_b=1-M_f$7 MSE (Zhu et al., 2022).

Subsequent models made the regional prior more explicit. HDNet introduces hierarchical adaptation through a Local Dynamic (LD) module, which matches foreground local representations to the $M_b=1-M_f$8-nearest neighbor background regions using cosine similarity, and a Mask-aware Global Dynamic (MGD) module, which applies distinct convolutions to foreground and background in the decoder (Chen et al., 2022). FRIH performs a global coarse harmonization and then adaptively clusters the foreground into $M_b=1-M_f$9 RGB-homogeneous submasks using CFSFDP, with $R$0, refining each sub-region through a lightweight cascaded module and fusion head. On iHarmony4, FRIH reports PSNR $R$1 dB, MSE $R$2, and a total model size of $R$3 M parameters (Peng et al., 2022). A complementary line introduces explicit regional discrimination: region-wise contrastive learning selects positive samples from the ground-truth background region and negative samples from the current harmonized foreground region, while external background style fusion injects masked background statistics into decoder features; its appendix reports the best patch-sampling result at $R$4, with MSE $R$5, PSNR $R$6, and SSIM $R$7 (Liang et al., 2022).

More recent formulations emphasize global context, semantics, and intermediate supervision. “Deep Image Harmonization with Globally Guided Feature Transformation and Relation Distillation” formulates the task as transforming foreground encoder features so that their statistics are compatible with background features, uses a bottleneck global vector $R$8, and predicts channel-wise scales for modulated convolution through GIFT. The regional transformation is explicit: $R$9 with the background branch left unchanged (Niu et al., 2023). “Segment Anything Model Meets Image Harmonization” replaces hard region loops with Semantic-guided Region-aware Instance Normalization (SRIN), in which SAM-derived semantic maps guide cross-attention and generate spatially varying $M_R$0 for foreground-only modulation at the bottleneck. On iHarmony4, SRIN reports average MSE $M_R$1 and PSNR $M_R$2 dB, and on HAdobe5k at $M_R$3 reports PSNR $M_R$4 dB, MSE $M_R$5, and fMSE $M_R$6 (Chen et al., 2023). These results make explicit that regional harmonization can be realized through correspondence, clustering, contrastive discrimination, global modulation, or semantic normalization, provided the transformation is restricted to the target region.

3. Interactive, generative, and style-based extensions

Regional feature harmonization also underlies interactive and generative systems in which the guiding region is not simply “the background.” Interactive Portrait Harmonization introduces a user-selected reference region $M_R$7 and a guide mask $M_R$8, encoded with partial convolutions, to produce a style code $M_R$9 that conditions the harmonizer both by AdaIN layers and by concatenation with the masked foreground input (Valanarasu et al., 2022). The losses are also regionalized: luminance matching aligns highlight, mid-tone, and shadow statistics,

$\Omega_k\subseteq\mathcal{X}$0

and style consistency plus two triplet losses align the harmonized foreground style code with the selected guide region and the ground truth. On PortraitTest, the portrait-specialized IPH++ reports PSNR $\Omega_k\subseteq\mathcal{X}$1, SSIM $\Omega_k\subseteq\mathcal{X}$2, and MSE $\Omega_k\subseteq\mathcal{X}$3 (Valanarasu et al., 2022).

In diffusion-based and training-free settings, the same principle appears as region-restricted feature sharing. StorySync builds subject masks from cross-attention maps, allows self-attention queries in one image to attend to keys and values from other images only within subject regions, and then applies Regional Feature Harmonization (RFH) by matching each subject-region descriptor $\Omega_k\subseteq\mathcal{X}$4 to its best correspondence in another image. The update is residual and masked,

$\Omega_k\subseteq\mathcal{X}$5

with confidence gating to avoid over-smoothing (Gaur et al., 31 Jul 2025). On an SDXL backbone, StorySync reports CLIP-I $\Omega_k\subseteq\mathcal{X}$6 and DreamSim $\Omega_k\subseteq\mathcal{X}$7, and an ablation without RFH drops CLIP-I to $\Omega_k\subseteq\mathcal{X}$8 and raises DreamSim to $\Omega_k\subseteq\mathcal{X}$9 (Gaur et al., 31 Jul 2025).

Generative image harmonization has likewise been reframed as region-to-region injection. R2R combines Clear-VAE with an Adaptive Filter for detail preservation, a Harmony Controller equipped with Mask-aware Adaptive Channel Attention (MACA), and a latent diffusion model fine-tuned for foreground-focused denoising (Zhang et al., 13 Aug 2025). MACA explicitly separates foreground and background feature maps, computes masked channel descriptors, regresses per-channel scale and shift, and reinjects them only into the foreground: $F$0 On iHarmony4, this model reports overall PSNR $F$1, MSE $F$2, and fMSE $F$3 (Zhang et al., 13 Aug 2025). A related but broader facial synthesis formulation treats harmonization as cross-region style compatibility. “Towards Harmonized Regional Style Transfer and Manipulation for Facial Images” extracts per-region style embeddings, adapts them through a multi-region style attention module, and evaluates coherence with a “harmony score” that combines boundary consistency, color/illumination consistency, and texture compatibility (Wang et al., 2021). Across these works, regional harmonization functions as a constraint on controllability: it permits local transfer or subject consistency without collapsing the full image to a single global code.

4. Medical imaging, radiomics, and anatomy-aware harmonization

In MRI and radiomics, regional feature harmonization addresses site, scanner, protocol, or acquisition heterogeneity rather than foreground-background mismatch. A recent survey defines region-based MRI features as cortical thickness or ROI volumes from T1-weighted morphometry, tract- or ROI-wise DTI metrics such as FA and MD, and regional fMRI measures such as mean ROI connectivity and network graph features (Yang et al., 22 Jul 2025). Feature-level harmonization is commonly expressed by ComBat: $F$4 with empirical Bayes shrinkage for the site-specific location and scale effects $F$5 and $F$6. The survey distinguishes ComBat, ComBat-GAM, CovBat, LongComBat, linear mixed-effects models, and deep approaches such as cVAE, gcVAE, DeepComBat, and DeepResBat, and recommends covariate-aware ComBat as a strong baseline, with CovBat when cross-feature covariance matters for machine learning (Yang et al., 22 Jul 2025).

Several studies show that subgroup-aware regionalization is essential when acquisition effects interact with biology. In pulmonary-nodule radiomics, ComBat was evaluated under pooled harmonization, harmonization with subgroup covariates, and separate subgroup-specific harmonization for benign and malignant nodules. The mean proportion of acquisition-independent features was $F$7 for pooled harmonization, $F$8 with covariates, and $F$9 with separate harmonization; screening ROC-AUCs were $F^f$0, $F^f$1, and $F^f$2, respectively (Huchthausen et al., 2024). In echocardiography, self-supervised ConvNeXt-V2 kernels were repurposed as a fixed preprocessing module before myocardium-ROI radiomics extraction. The filtered image $F^f$3 is the average of $F^f$4 feature maps, and the method reports the lowest mean JSD, $F^f$5, for cross-manufacturer feature alignment, together with downstream HHD-vs-HCM AUC $F^f$6 (Lee et al., 2023).

At the image level, anatomy-conditioned modulation has become a direct analogue of regional feature harmonization. Segmentation-Renormalized Deep Feature Modulation conditions generator normalization layers on learned anatomical segmentation embeddings, replacing fixed affine parameters with FiLM-style scale and shift: $F^f$7 so that image translation is region-aware with respect to anatomy rather than purely global style (Ren et al., 2021). The method improves FID, KID, downstream Dice, and robustness to self-adversarial perturbation across T1w MRI, FLAIR MRI, and OCT. IHF-Harmony generalizes this anatomy-preserving perspective to multi-modality MRI through an invertible hierarchy flow that subtractively removes artefact-related features and an artefact-aware normalization that transfers target characteristics while keeping anatomy fixed: $F^f$8 Its full model reports RMSE $F^f$9, MS-SSIM $F^b$0, PSNR $F^b$1, and LPIPS $F^b$2 in the reported ablation table (Zhu et al., 25 Feb 2026). In this medical literature, “regional” usually means anatomically or biologically meaningful partitions, and harmonization is judged by whether non-biological variability is reduced without degrading anatomy or disease signal.

5. Regional harmonization beyond image appearance

The concept also appears outside classical image harmonization. GRANITE treats disagreement between feature-based explanation methods as a problem of regional feature-space incompatibility. Given two explanations that differ because of interaction handling or masking distributions, it partitions the feature space into disjoint regions $F^b$3 where those disagreement sources are minimized (Herbinger et al., 30 Jan 2026). The framework shows that, when masking is fixed, the difference between two regional explanations is a weighted sum of higher-order pure interactions (Theorem 1), and when interaction handling is fixed, conditional-versus-marginal disagreement is entirely induced by regional distributional differences (Theorem 3). On Bikesharing, remaining disagreement after depth-3 partitioning drops from $F^b$4 to $F^b$5 for ICE vs. PDP with GBT and from $F^b$6 to $F^b$7 for CFI vs. PFI with GBT (Herbinger et al., 30 Jan 2026). Here harmonization means agreement between explanation operators, not appearance transfer.

Hydrometeorological regionalization offers another variant. A large-scale streamflow study computes $F^b$8 standardized time-series features—autocorrelation, entropy, seasonality, trend, lumpiness, stability, nonlinearity, linearity, spikiness, curvature, and others—on precipitation, temperature, and streamflow from $F^b$9 catchments, after z-score scaling and STL decomposition with $\mu_{f,l}$0 seasons per year (Papacharalampous et al., 2022). These descriptors are then merged with topographic, land-cover, soil, and geologic attributes in random-forest regionalization models. Relative RMSE improvements reach up to about $\mu_{f,l}$1 for some streamflow features, and spectral entropy, seasonality strength, and several autocorrelation features are reported as more regionalizable than others (Papacharalampous et al., 2022). This suggests that “harmonization” can also mean constructing scale-free, uniformly computed regional descriptors that support transfer to ungauged locations.

A related precursor in computer vision is DRFI, which formulates saliency as regression over a $\mu_{f,l}$2-dimensional regional feature vector combining regional contrast, backgroundness, and regional properties across multiple segmentation levels (Jiang et al., 2014). Although its primary task is salient object detection rather than harmonization, its regional contrast and backgroundness constructions explicitly normalize heterogeneous cues through region comparisons and multi-level fusion. The reported AUC reaches $\mu_{f,l}$3 on MSRA-B (Jiang et al., 2014). A plausible interpretation is that later harmonization work inherits this broader regional-processing tradition: region-specific descriptors are constructed first, and only then integrated into a coherent global decision.

6. Recurring mechanisms, evaluation criteria, and limitations

Across domains, several mechanisms recur. One is region restriction: many methods leave a complement unchanged, such as the background in image harmonization (Niu et al., 2023), non-subject regions in StorySync (Gaur et al., 31 Jul 2025), or anatomy-preserving latent content in MRI harmonization (Zhu et al., 25 Feb 2026). Another is region-conditioned modulation, implemented through AdaIN, FiLM, instance normalization with learned $\mu_{f,l}$4, masked dynamic convolution, or channel-wise affine transforms derived from regional descriptors (Valanarasu et al., 2022). A third is soft rather than hard regional correspondence: cross-attention, semantic priors, K-nearest neighbor fusion, or content-affinity matrices are repeatedly used to avoid brittle one-to-one matching (Chen et al., 2023). This suggests that regional feature harmonization is usually most effective when regions are treated as structured, probabilistic supports rather than rigid partitions.

Evaluation is similarly domain-specific but structurally aligned. Natural-image harmonization primarily uses PSNR, MSE, fMSE, and SSIM (Peng et al., 2022). Story and subject-consistency models use CLIP-I, DreamSim, CLIP-T, LPIPS, and human preference scores (Gaur et al., 31 Jul 2025). Medical harmonization uses site separability, JSD, ROC-AUC, Dice, FID, KID, LPIPS, MS-SSIM, and downstream task performance (Lee et al., 2023). GRANITE uses regional disagreement risk and remaining disagreement percentages (Herbinger et al., 30 Jan 2026). The shared criterion is not a single universal metric but the reduction of region-linked inconsistency without erasing desired signal.

Limitations also recur. Many image methods are sensitive to inaccurate masks, poor guide-region choice, weak content correspondence, or semantic segmentation failures (Valanarasu et al., 2022). In radiomics and MRI, severe confounding between site and biology, pathology-dependent acquisition effects, and unseen sites remain difficult (Huchthausen et al., 2024). GRANITE notes that smooth interaction or dependency effects may require many regions, reducing interpretability (Herbinger et al., 30 Jan 2026). Hydrometeorological regionalization remains sensitive to nonstationarity and regional bias (Papacharalampous et al., 2022). A plausible implication is that regional feature harmonization works best when the chosen regions correspond to actual generative or causal heterogeneity; if the partition is misaligned with the source of variation, harmonization may under-correct, over-correct, or simply move inconsistency to another level.

Taken together, regional feature harmonization is less a single algorithm than a recurring design principle: define a meaningful region, estimate region-specific mismatch, transform the target region with context-aware supervision or modulation, and preserve the variables that should not change. The principle spans composite-image editing, diffusion-based subject consistency, anatomical image translation, radiomics, explanation alignment, and hydrological regionalization, but in every case its central aim remains the same: comparability without loss of structure.