Dual-Reference Normalization

• Dual-reference normalization is a method that uses two complementary references to preserve distinct statistical information and overcome limitations of single-reference approaches.
• It is applied in areas such as cross-modality medical image segmentation, holographic imaging, and cross-modal retrieval, enhancing model stability and style-awareness.
• Empirical results show improved performance metrics, including higher Dice scores and lower reconstruction errors, compared to traditional normalization techniques.

Searching arXiv for recent and foundational papers related to “dual-reference normalization” across the contexts represented in the provided material. In current literature, dual-reference normalization is used for a family of methods that replace a single normalization, conditioning, or recovery reference with two complementary references. The most explicit formulation appears in generalizable cross-modality medical image segmentation, where a model trained on one labeled source modality stores separate normalization statistics for source-similar and source-dissimilar style groups and then routes each target image to the closer branch at test time (Zhou et al., 2021). Closely related constructions appear in holographic coherent diffraction imaging, where block and pinhole references are combined in one structured inversion (Barmherzig et al., 2019), in cross-modal retrieval via Dual Bank Sinkhorn Normalization (Pan et al., 4 Aug 2025), and in several broader frameworks that normalize with respect to two regimes, two branches, or two complementary reference signals rather than one averaged distribution (Lan et al., 18 Jun 2026, Benatti et al., 2024, Bansal et al., 5 Feb 2026). This suggests a unifying interpretation: dual-reference normalization is not one universal operator, but a recurrent strategy for preserving complementary reference information that would otherwise be collapsed.

1. Cross-domain rationale

A recurring motivation for dual-reference designs is the failure of a single reference distribution under heterogeneity, multimodality, or large shift. In cross-modality medical imaging, standard normalization layers, especially BatchNorm, estimate statistics that reflect the training distribution; when the target modality has a very different appearance distribution, particularly a gray-value or style shift, those source statistics can be a poor match and can cause serious performance degradation. In holographic coherent diffraction imaging, the classical block and pinhole references have different noise behavior: the block reference tends to work better for low-frequency-dominant specimens, while the pinhole reference tends to work better for flat-spectrum data. In cross-modal retrieval, balancing target probabilities alone is described as inadequate because the query distribution may also be biased; hence balancing both sides becomes necessary. In proportional feature spaces, normalization must be aligned with whether the comparison is translation invariant or scale invariant. In ReLU-dual cut generation for stochastic mixed-integer programs, multiple optimal dual solutions can yield weak cuts, so normalization is used to select stronger representatives in an extended space (Zhou et al., 2021, Barmherzig et al., 2019, Pan et al., 4 Aug 2025, Benatti et al., 2024, Bansal et al., 5 Feb 2026).

Across these settings, the two references are not interchangeable duplicates. One typically captures a regime that is more local, direct, source-near, or low-frequency, while the other captures a regime that is more global, source-far, cumulative, or distribution-corrective. A plausible implication is that dual-reference normalization is most useful when the two references have complementary failure modes and when the downstream procedure can preserve that complementarity instead of averaging it away.

2. Dual-reference normalization in cross-modality medical image segmentation

The paper "Generalizable Cross-modality Medical Image Segmentation via Style Augmentation and Dual Normalization" formulates dual-reference normalization as a response to a hard domain-generalization problem: a model is trained only on one labeled source modality or domain and must be applied directly to a different unseen target modality without access to target data during training and without re-training afterward (Zhou et al., 2021). The method begins from the observation that many medical imaging modalities are gray-scale and that modality differences often manifest as intensity-distribution differences rather than semantic changes. To simulate appearance shifts from a single source domain, the paper applies monotonic nonlinear intensity remapping with cubic Bézier curves,

$$B(t) = \sum\limits_{i=0}^{n}\binom{n}{i}P_i(1-t)^{n-i}t^i,\quad n=3,\quad t\in[0,1],$$

with domain and range normalized to $[-1,1]$. The transformations are applied only to foreground regions.

The augmented images are partitioned into two style groups. The source-similar domain $\mathcal{D}^{ss}$ consists of images produced by increasing transformations, which remain closer to the source style. The source-dissimilar domain $\mathcal{D}^{sd}$ consists of images produced by decreasing transformations, which induce a larger gray-distribution gap from the source. These two groups are the dual references of the method. The segmentation backbone is a U-Net, but all BatchNorm layers are replaced by dual-normalization layers with a shared backbone and independent batch normalization layers for the two branches. The backbone parameters are shared, while the normalization statistics and affine parameters are independent.

Training uses only a segmentation objective, namely a sum of soft Dice losses over the two groups: $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$ Running statistics are updated separately for each branch with exponential moving averages, so each branch accumulates its own layerwise mean and variance statistics. The paper describes this as a lightweight ensemble: the network weights are mostly shared, but the branches encode distinct source-style distributions.

At test time, the model computes instance statistics at each BN layer for the target sample, forming a target style embedding

$$e_t=[e_t^1,\dots,e_t^L]=[(\mu_t^1,{\sigma_t^1}^2),\dots,(\mu_t^L,{\sigma_t^L}^2)],$$

and compares it with each branch embedding

$$e_d=[e_d^1,\dots,e_d^L]=[(\mu_d^1,{\sigma_d^1}^2),\dots,(\mu_d^L,{\sigma_d^L}^2)].$$

The branch is selected by nearest distance across BN layers,

$c=\mathop{\arg\min}\limits_d \text{Dist}(e_t,e_d).$

The method is therefore explicitly style-aware at inference: it does not use one frozen normalization state and does not average multiple branches.

The reported experiments cover BraTS, Cross-Modality Cardiac, and Abdominal Multi-Organ. On BraTS, with T2 as source, the method reaches an average Dice of 54.44% and average HD of 19.05 mm; with T1CE as source, it reaches an average Dice of 57.98% and HD of 18.88 mm. On the cardiac dataset, it achieves 51.70% Dice for MRI$\rightarrow$CT and 32.19% Dice for CT$\rightarrow$MRI. On the abdominal multi-organ dataset, the reported CT$[-1,1]$0MRI result is 70.22% Dice and 2.94 mm HD (Zhou et al., 2021).

The ablations are central to the definition of the method. Training only on the source-similar branch or only on the source-dissimilar branch does not generalize uniformly across target modalities; style-based selection is better than either branch alone and better than naïve ensembling; and the number of Bézier control-point pairs does not dramatically change performance within a reasonable range. The paper explicitly states that the novelty is not merely “using multiple BN layers,” but using two intentionally constructed style references and choosing between them by comparing style embeddings at inference (Zhou et al., 2021).

3. Dual references as stabilization in holographic coherent diffraction imaging

The paper "Dual-Reference Design for Holographic Coherent Diffraction Imaging" introduces a dual-reference design consisting of a block reference and a pinhole reference placed adjacent to the imaging specimen (Barmherzig et al., 2019). The composite object is

$[-1,1]$1

where $[-1,1]$2 is the unknown specimen, $[-1,1]$3 is the all-ones $[-1,1]$4 block reference, and $[-1,1]$5 is a single nonzero pixel at one corner. The geometry places $[-1,1]$6 in the upper-left quadrant, $[-1,1]$7 in the upper-right, $[-1,1]$8 in the lower-left, and $[-1,1]$9 in the lower-right of a $\mathcal{D}^{ss}$0 array.

The recovery algorithm uses oversampled Fourier magnitude data, forms an autocorrelation by inverse DFT, extracts two cross-correlation blocks, and solves a stacked overdetermined linear system. The two references induce different linear maps: $\mathcal{D}^{ss}$1 The block channel is cumulative or integration-like, while the pinhole channel is identity-like. Their stacked least-squares solution yields a linear map from the data to the estimate, so the phase retrieval problem is reduced to a single structured linear inversion after autocorrelation estimation.

The paper does not use the phrase “dual-reference normalization” as a formal preprocessing operator. Instead, the normalization or stabilization interpretation is implicit in the error analysis. The block and pinhole references produce different scalings and conditioning in the linearized recovery; the pinhole stabilizes local amplitudes, while the block provides strong low-frequency or global information. The key claim is that the dual-reference choice makes the weight matrix governing Poisson shot-noise propagation behave approximately like the pointwise minimum of the single-reference weights. Under the Poisson model,

$\mathcal{D}^{ss}$2

and the dual-reference design is argued to be uniformly superior to either single-reference scheme in expected reconstruction error (Barmherzig et al., 2019).

The simulations use 9 specimen images, each resized to $\mathcal{D}^{ss}$3, oversampled Fourier data of size $\mathcal{D}^{ss}$4, and Poisson noise with nominal photon count

$\mathcal{D}^{ss}$5

The reported relative squared error is

$\mathcal{D}^{ss}$6

For Mimivirus, the dual-reference method gives 1.51 versus 3.70 for block and 46.9 for pinhole; for Influenza, 4.64 versus 18.7 and 50.7; for mCherry proteins, 0.927 versus 1.84 and 139.5 (Barmherzig et al., 2019). In this setting, the dual-reference effect is a conditioning mechanism rather than a normalization layer.

4. Probability balancing and Dual Bank Sinkhorn Normalization

In cross-modal retrieval, the paper "Hubness Reduction with Dual Bank Sinkhorn Normalization for Cross-Modal Retrieval" interprets hubness reduction as a probability-balancing problem over the query–target similarity matrix

$\mathcal{D}^{ss}$7

A hubness problem arises when a small number of targets are retrieved by many queries. The paper reinterprets Inverted Softmax as target-probability balancing: it adjusts similarity scores with a target-specific compensation term and can be written as a column-normalized softmax. This balances target marginals only. The paper argues that this is insufficient because the query distribution remains unconstrained (Pan et al., 4 Aug 2025).

The proposed Sinkhorn Normalization (SN) enforces both row and column marginals through an entropy-regularized optimal transport formulation. With

$\mathcal{D}^{ss}$8

the objective constrains both $\mathcal{D}^{ss}$9 and $\mathcal{D}^{sd}$0, and the solution is computed by the Sinkhorn-Knopp algorithm on the Gibbs kernel

$\mathcal{D}^{sd}$1

The resulting transport plan can be written as

$\mathcal{D}^{sd}$2

so the normalization includes both a query-side and a target-side compensation term.

In query-agnostic retrieval, the test queries are unavailable at inference, so the normalization terms cannot be computed from the true test set. Existing methods use a query bank $\mathcal{D}^{sd}$3, but the paper argues that this can induce a substantial distributional gap between the query bank, the true test queries, and the target gallery. The proposed Dual Bank Sinkhorn Normalization (DBSN) adds a target bank $\mathcal{D}^{sd}$4 and estimates hubness over the extended target set $\mathcal{D}^{sd}$5. The paper states this formally with

$\mathcal{D}^{sd}$6

capturing the claim that a target bank narrows the query–target distributional gap (Pan et al., 4 Aug 2025).

The empirical evaluation covers image-text retrieval, video-text retrieval, and audio-text retrieval. On Flickr30k, zero-shot CLIP with SN reaches R@1 = 69.3 versus 66.8 for IS; on MS-COCO, zero-shot CLIP with SN reaches R@1 = 40.8 versus 38.6 for IS; on MSR-VTT, CLIP4Clip with SN reaches R@1 = 49.2 versus 46.4 for IS; on AudioCaps, SN reaches R@1 = 43.6 versus 41.5 for IS. In the query-agnostic dual-bank setting, on Flickr30k with training query bank and training target bank, DBSN obtains R@1 = 67.1, above SN (66.3) and IS (65.1) (Pan et al., 4 Aug 2025). Here the dual-reference idea is explicitly probabilistic: normalization balances both sides of retrieval rather than only one.

5. Dual-reference control in generative and neural normalization systems

The paper "FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining" studies style-content dual-reference generation, where a model receives a content reference image $\mathcal{D}^{sd}$7, a style reference image $\mathcal{D}^{sd}$8, and a text prompt $\mathcal{D}^{sd}$9, and must generate an image that preserves the structure and semantics of $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$0, adopts the visual style of $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$1, and follows the instruction in $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$2 (Lan et al., 18 Jun 2026). The paper identifies a central failure mode: semantic leakage from the style reference. Its response is a two-stage curriculum with stage-specific disentanglement mechanisms.

In Stage 1, style-reference generation is regularized by an attention-level enrichment constraint. For semantic group $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$3, the enrichment score is

$$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$4

and the paper regularizes the first transformer block with a two-sided hinge. It also adds entropy regularization over style-reference attention. In Stage 2, dual-reference generation uses frequency-aware RoPE modulation on the style-reference branch,

$$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$5

where $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$6 suppresses high-frequency local copying and $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$7 amplifies low-frequency global style guidance. The paper explicitly interprets this as balancing the two reference streams so that style is transferred without semantic contamination and content is preserved without style collapse.

The data pipeline mines community LoRAs from Civitai, TensorArt, and Liblib, built on Illustrious, FLUX-dev, and Qwen-Image, and reports final dataset sizes of 273k triplets with FLUX, 33k with Qwen, and 172k with Illustrious. The benchmark includes 200 content references and 200 style references and evaluates style fidelity, content fidelity, instruction following, VLM-based verification, and aesthetics. On SRef, the model ranks first on VLM-Style = 7.142 and Ver-S = 0.482; on CRef+SRef, it again ranks first on VLM-Style = 5.467 and Ver-S = 0.409. Ablations report that enabling the attention enrichment loss changes the VLM leakage score from 2.674 to 0.522, and enabling RoPE modulation changes the leakage score from 1.047 to 0.453 (Lan et al., 18 Jun 2026).

A broader neural-normalization antecedent is "Mode Normalization" (Deecke et al., 2018). That paper does not impose exactly two references, but it generalizes BatchNorm by learning multiple mode-specific means and variances on the fly: $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$8 Its central claim is that BN is too restrictive on multi-modal or heterogeneous data because it uses one set of batch statistics for the whole mini-batch. Mode Normalization instead detects modes dynamically and jointly normalizes samples that share common features. This provides a useful technical backdrop for later dual-reference architectures: dual-reference normalization can be viewed as the $$\mathcal{L}_{seg} = \mathcal{L}_{Dice}(S_{\theta}(x_{ss}), y_{ss}) + \mathcal{L}_{Dice}(S_{\theta}(x_{sd}), y_{sd}).$$9 end of a broader family in which multiple references are preserved rather than collapsed (Deecke et al., 2018).

6. Broader extensions, scope, and common misconceptions

The phrase extends beyond neural-network normalization layers. In "Normalization in Proportional Feature Spaces," normalization is organized around a duality between uniform and proportional feature spaces. Uniform comparisons satisfy

$$e_t=[e_t^1,\dots,e_t^L]=[(\mu_t^1,{\sigma_t^1}^2),\dots,(\mu_t^L,{\sigma_t^L}^2)],$$0

while proportional comparisons satisfy

$$e_t=[e_t^1,\dots,e_t^L]=[(\mu_t^1,{\sigma_t^1}^2),\dots,(\mu_t^L,{\sigma_t^L}^2)],$$1

The paper introduces a duality transformation $$e_t=[e_t^1,\dots,e_t^L]=[(\mu_t^1,{\sigma_t^1}^2),\dots,(\mu_t^L,{\sigma_t^L}^2)],$$2, under which translation in the uniform domain becomes scaling in the proportional domain. It then argues that uniform features should be normalized by standardization, whereas proportional features should be normalized by scaling only, using Separated Proportional Normalization (SPN) or Joint Proportional Normalization (JPN) based on non-centralized dispersion. Here the two references are two invariance regimes—translation and scaling—rather than two learned branches (Benatti et al., 2024).

In "Database Normalization via Dual-LLM Self-Refinement," the phrase is used in a looser but structurally related sense. The framework Miffie uses a generation module and a verification module in an iterative loop for schema normalization up to 3NF. The implementation uses GPT-4 for generation and o1-mini for verification, with the maximum practical refinement limit set to 3 loops. The dual-model version reaches an elimination rate of 72%, compared with 45% for GPT-4 only and 57% for o1-mini only. This is not feature normalization in the classical sense; rather, the “dual-reference” effect comes from judging one output against a second model’s verification feedback (Jo et al., 25 Aug 2025).

In "Normalization of ReLU Dual for Cut Generation in Stochastic Mixed-Integer Programs," normalization is performed in an extended lifted space for the ReLU dual. The paper introduces a normalized dual with coefficients constrained by

$$e_t=[e_t^1,\dots,e_t^L]=[(\mu_t^1,{\sigma_t^1}^2),\dots,(\mu_t^L,{\sigma_t^L}^2)],$$3

and proves that the resulting cuts are tight and Pareto-optimal in the original state space under stated conditions. It also proves that normalization can recover any cut obtained by regularization, whereas the converse does not hold. Here the term refers neither to BN-style statistics nor to preprocessing; it denotes a geometric mechanism for selecting stronger dual representatives under degeneracy (Bansal et al., 5 Feb 2026).

A common misconception is therefore to treat dual-reference normalization as a single standardized technique. The literature does not support that interpretation. In some papers it is a literal normalization layer with two stored statistical references; in others it is a dual-reference design for inversion, a doubly stochastic probability-balancing method, a two-branch disentanglement mechanism, a feature-space invariance framework, or a normalized dual formulation. The stable common element is narrower: two complementary references are retained because a single reference is judged insufficient for robustness, conditioning, or correctness (Benatti et al., 2024, Jo et al., 25 Aug 2025, Bansal et al., 5 Feb 2026).