Conditional Normalizing Flow Nets

• Conditional normalizing flow nets are invertible generative models that map data to latent space via context-dependent bijections, enabling exact conditional likelihood evaluation.
• They employ diverse architectures—autoregressive, coupling, spline-based, and continuous flows—to effectively model complex conditional distributions in applications like detector simulation and probabilistic forecasting.
• These models support efficient maximum likelihood training, fast conditional sampling, and rigorous diagnostic checks with theoretical guarantees on tractable Jacobians and symmetry properties.

A conditional normalizing flow net is a trainable invertible generative model for conditional density estimation. It represents a target conditional law such as $p(x\mid c)$ by applying a context-dependent bijection between observations and a simple base distribution, typically Gaussian, so that likelihood evaluation, sampling, and diagnostics remain explicit through change of variables. Across the literature, the term encompasses both discrete compositions of invertible blocks—such as Masked Autoregressive Flows, affine coupling layers, and spline transforms—and continuous-time flows defined by ODEs or continuity equations. In all cases, the defining features are conditioning on auxiliary variables, tractable Jacobian terms, and exact or explicitly computable conditional likelihoods (Xu et al., 2023, Wen et al., 2022, Nguyen et al., 2019, Geshkovski et al., 9 Feb 2026).

1. Conditional density modeling and change of variables

In its standard form, a conditional normalizing flow specifies an invertible map $f(\cdot;c)$ from data space to latent space, with latent variable $z=f(x;c)$ drawn from a simple base distribution. The conditional density then follows from the change-of-variables identity

$$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$

This formulation appears across conditional flows for detector simulation, probabilistic forecasting, decorrelation, density reweighting, and conditional transport (Xu et al., 2023, Wen et al., 2022, Klein et al., 2022, Algren et al., 2023).

The same principle is also used when the conditioning enters through the base density rather than the invertible map itself. In AP-CDE, for example, the flow transform $T_\theta(y)$ is independent of $x$, while the latent law is decomposed as $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$, so that

$$p(y\mid x)=p_Z(T_\theta(y)\mid x)\left|\det J_{T_\theta}(y)\right|.$$

This produces a conditional model with a supervised latent subspace $z_P$ and an unsupervised residual component $z_N$ (Zeng et al., 6 Jul 2025). InfoCNF similarly conditions a continuous flow through a partitioned latent prior, using a supervised code $f(\cdot;c)$0 and an unconditional code $f(\cdot;c)$1 rather than conditioning the ODE dynamics directly (Nguyen et al., 2019).

A continuous-time variant replaces the finite composition of bijections by an ODE

$f(\cdot;c)$2

with density evolution

$f(\cdot;c)$3

This yields continuous normalizing flows and conditional continuous normalizing flows with exact likelihoods up to numerical ODE integration (Nguyen et al., 2019). A more constructive perspective formulates conditional transport by a continuity equation

$f(\cdot;c)$4

and studies how piecewise-constant neural velocity fields can approximate both a diffeomorphism $f(\cdot;c)$5 and its pushforward $f(\cdot;c)$6 (Geshkovski et al., 9 Feb 2026).

2. Architectural patterns and conditioning mechanisms

The dominant architectural distinction is between autoregressive, coupling-based, spline-based, and continuous-flow constructions. In detector-response modeling for $f(\cdot;c)$7, the core architecture is a chain of Masked Autoregressive Flows with interleaved permutations and a final $f(\cdot;c)$8 bijection, trained to model a six-dimensional detector-response vector conditioned on truth-level kinematics and pile-up (Xu et al., 2023). In one MAF block,

$f(\cdot;c)$9

with $z=f(x;c)$0 and $z=f(x;c)$1 produced by masked autoregressive networks. Conditioning is injected by concatenating $z=f(x;c)$2 to the network inputs at each block, and permutations mitigate sensitivity to any single variable ordering (Xu et al., 2023).

Spline flows replace affine transforms by monotone rational-quadratic splines. In probabilistic wind-power forecasting, the conditional flow uses autoregressive spline transforms with MADE-style conditioning and a conditional Gaussian base whose mean and covariance depend on contextual predictors (Wen et al., 2022). In the reweighting setting, rational-quadratic spline autoregressive flows are used as conditional density estimators over source variables $z=f(x;c)$3 given covariates $z=f(x;c)$4, thereby avoiding explicit density-ratio estimation (Algren et al., 2023). In decorrelation problems, rational-quadratic spline flows are applied to a one-dimensional discriminant conditioned on protected attributes, with the requirement that $z=f(x;c)$5 remain invertible and monotone for each $z=f(x;c)$6 (Klein et al., 2022).

Coupling-based conditional flows remain important when faster sampling is preferred. The generic affine conditional coupling layer keeps one part of the variable unchanged and updates the other part through scale and shift networks that depend on both the frozen coordinates and the conditioning signal, producing triangular Jacobians with additive log-determinants (Baranchuk et al., 2021). In surrogate radiative-transfer modeling, the flow is described as RealNVP-based, but the reported equations specialize to an element-wise affine conditional transformation

$z=f(x;c)$7

with a diagonal Jacobian and conditioner outputs given by a fully connected network (Seyedheydari et al., 27 Aug 2025). In cKRnet for domain-decomposed uncertainty quantification, the conditional affine coupling is modified to preserve triangular structure and stability through bounded scaling,

$z=f(x;c)$8

with $z=f(x;c)$9 (Li et al., 2024).

Continuous conditional flows parameterize the velocity field rather than a finite stack of bijections. InfoCNF uses a multiscale FFJORD architecture with a partitioned latent prior and adaptive solver tolerances learned by gating networks (Nguyen et al., 2019). Semi-equivariant conditional normalizing flows for 3D graphs define the vector field through EGNN-based mappings conditioned on a receptor graph, and then prove that semi-equivariance suffices for conditional invariance under joint rigid motions (Rozenberg et al., 2023).

A distinct constructive line develops conditional flows as finite compositions of incompressible and compressible transformations realized by neural ODEs. The construction in (Geshkovski et al., 9 Feb 2026) decomposes a target map into measure-preserving pieces $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$0 and a one-coordinate compressible part $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$1, then realizes the former by divergence-free shear flows and the latter exactly by ReLU slope-change stages. This places conditional normalizing flow nets in direct contact with Knöthe–Rosenblatt transport, cube permutations, and continuity-equation realizations (Geshkovski et al., 9 Feb 2026).

3. Likelihood training, inference, and diagnostics

A defining property of conditional normalizing flow nets is exact conditional likelihood. Training is generally maximum likelihood, or equivalently negative conditional log-likelihood minimization, without variational lower bounds. For data $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$2,

$$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$3

and the same template is used in detector simulation, probabilistic forecasting, decorrelation, reweighting, and radiative surrogates (Xu et al., 2023, Wen et al., 2022, Klein et al., 2022, Algren et al., 2023, Seyedheydari et al., 27 Aug 2025).

Sampling is the inverse operation. One draws $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$4 and computes $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$5. In autoregressive flows this inversion is sequential, while in coupling flows it is typically parallel within each block. For the detector-response model, conditional sampling proceeds autoregressively through the inverse MAF chain after drawing $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$6 (Xu et al., 2023). For radiative transfer, posterior samples are drawn by applying the learned diagonal affine inverse to Gaussian latent samples, then aggregating posterior means, standard deviations, and $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$7 intervals over an ensemble of cross-validation models (Seyedheydari et al., 27 Aug 2025).

Likelihood tractability also supports diagnostics and model selection. In the LHC detector application, early stopping is selected using minimal mean Wasserstein distance across six response dimensions on a validation set, even though the training objective itself is exact log-likelihood (Xu et al., 2023). In decorrelation, one trains the conditional flow on background-only samples so that the transformed score matches a base law independent of the protected attribute by construction; no explicit mutual-information or MMD penalty is required in the main method (Klein et al., 2022). In inverse problems, conditional flows can be pretrained on low-fidelity pairs and then reused as priors or initializations for more expensive high-fidelity variational objectives, again exploiting explicit conditional densities (Siahkoohi et al., 2021).

For incomplete-data settings, the conditional law induced by a trained normalizing flow can itself become the target of MCMC. Projected Latent MCMC samples exactly from $$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$8 by operating in latent space and using a Metropolis–Hastings correction whose target is built from the flow density and an auxiliary factor on observed coordinates (Cannella et al., 2020). This enables Monte Carlo EM for normalizing flows under missing data (Cannella et al., 2020).

Continuous conditional flows introduce solver-specific computational issues. In CNFs, the cost is dominated by the number of function evaluations required by the ODE solver. InfoCNF therefore augments conditional modeling with per-layer gating networks that learn solver tolerances, trading off speed, likelihood, and predictive performance (Nguyen et al., 2019). This suggests that, in continuous formulations, “Conditional Normalizing Flow Net” denotes not only a conditional density model but also a coupled dynamical system and numerical integration pipeline.

4. Application domains and representative instantiations

Conditional normalizing flow nets have been deployed across a wide range of scientific and engineering tasks, but the underlying objective remains conditional density estimation with invertibility and tractable Jacobians.

In high-energy physics detector modeling, the method is used as a surrogate for parts of the Monte Carlo chain. A six-dimensional detector-response vector for two photons,

$$p(x\mid c)=p_Z\big(f(x;c)\big)\left|\det J_f(x;c)\right|.$$9

is conditioned on truth-level kinematics and pile-up,

$T_\theta(y)$0

with the goal of reproducing asymmetric detector responses and inter-photon correlations in $T_\theta(y)$1 events at $T_\theta(y)$2 TeV (Xu et al., 2023). In the reported study, seven million events are generated with MadGraph@NLO v2.3.7 plus Pythia 8.235 and CTEQ6L1 PDFs, and the learned conditional flow reproduces baseline Gaussian smearing, engineered correlations at $T_\theta(y)$3 and $T_\theta(y)$4, and asymmetric low-energy tails (Xu et al., 2023). The largest discrepancy in learned resolutions across scans is reported as under $T_\theta(y)$5, and the mean and standard deviation of $T_\theta(y)$6 in $T_\theta(y)$7 GeV agree within statistical uncertainty (Xu et al., 2023).

In analysis-side HEP workflows, conditional flows are used for decorrelation and for replacing reweighting. For decorrelation, a conditional flow transforms a discriminant $T_\theta(y)$8 into $T_\theta(y)$9 so that the background distribution becomes independent of mass or other protected attributes while preserving ranking at fixed $x$0 (Klein et al., 2022). For distribution correction, a conditional flow learns $x$1 from source samples and then generates corrected samples by drawing $x$2 from a target marginal and sampling conditionally, avoiding binning and explicit density-ratio estimation (Algren et al., 2023). In the toy studies reported there, statistical precision is up to three times greater than with reweighting at identical sample sizes, and in a $x$3 application the per-bin uncertainties are $x$4–$x$5 smaller than with binned reweighting (Algren et al., 2023).

In probabilistic forecasting, spline-based conditional flows learn continuous conditional densities and thereby avoid quantile crossing. In wind-power forecasting, the model is described as distribution-free and produces continuous densities, coherent quantiles, and multivariate scenarios, with the proposed method attaining the best or tied-best CRPS on several GEFCom 2014 farms and competitive multivariate scores in temporal and spatial scenario generation (Wen et al., 2022). This suggests a broader interpretation of “net” as an end-to-end probabilistic forecaster rather than merely a generator.

In inverse problems and uncertainty quantification, conditional flows play two distinct roles. One is amortized posterior approximation, where the output distribution over latent or physical parameters depends on observations; the multi-fidelity scheme in (Siahkoohi et al., 2021) uses a block-triangular conditional flow pretrained on low-fidelity data and then fine-tuned for a high-fidelity posterior. The second is conditional density estimation inside larger Monte Carlo pipelines: cKRnet in CKR-DDUQ estimates $x$6 so that importance weights become

$x$7

which removes a joint-density bottleneck in domain-decomposed PDE uncertainty propagation (Li et al., 2024).

In scientific surrogate modeling beyond HEP, a conditional flow can replace expensive Monte Carlo simulation by returning full posterior predictive distributions. The radiative-transfer surrogate in (Seyedheydari et al., 27 Aug 2025) conditions on wavelength-dependent optical coefficients and particle-size information to predict reflectance, absorbance, and transmittance spectra, using Adam with learning rate $x$8, $x$9 epochs, five-fold cross-validation, and $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$0 posterior samples per fold at inference (Seyedheydari et al., 27 Aug 2025). The paper explicitly notes that energy conservation and non-negativity are not enforced, so predicted intervals can extend below zero (Seyedheydari et al., 27 Aug 2025).

Additional applications further broaden the concept. Conditional flows have been used for BSM parameter scans through differentiable conditional likelihoods (Saito et al., 2024), proposal construction in differentiable particle filters (Chen et al., 2021), individualized survival curves with CNF-based time-to-event densities (Ausset et al., 2021), lattice field-theory sampling through locality-constrained autoregressive conditional flows (R., 2023), receptor-aware ligand generation with semi-equivariant CNFs on 3D graphs (Rozenberg et al., 2023), and functional Bayesian inversion in PDEs through CNF-iVI (Zhao et al., 2024).

5. Theoretical guarantees and structured properties

The literature contains several distinct kinds of guarantees for conditional normalizing flow nets: exact likelihood identities, structural invariances, constructive approximation results, and convergence or consistency results inside larger algorithms.

Triangular and autoregressive structure gives immediate Jacobian tractability. For MAF, the Jacobian is triangular and the log-determinant is a sum of one-dimensional scale terms (Xu et al., 2023). For Knothe–Rosenblatt-type maps, the Jacobian determinant is the product of diagonal terms,

$p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$1

which is why triangular maps are central to constructive conditional transport and conditional sampling (Geshkovski et al., 9 Feb 2026). This is also the basis for cKRnet, whose conditional affine couplings remain triangular and bounded away from singularity through the $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$2 scaling (Li et al., 2024).

Constructive approximation theorems show that conditional sampling can be built from explicit neural flows. In (Geshkovski et al., 9 Feb 2026), a conditional transport map is approximated by a continuity-equation flow whose velocity is a width-1 ReLU perceptron with piecewise-constant controls. One theorem gives simultaneous approximation of a $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$3 diffeomorphism $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$4 and its pushforward $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$5 in $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$6 and total variation, albeit with a worst-case switch complexity scaling like $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$7 (Geshkovski et al., 9 Feb 2026). A second theorem yields a Maurey-in-time construction under Sobolev regularity with at most

$p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$8

switches, providing a dimension-independent $p_Z(z\mid x)=p(z_N)\,p(z_P\mid x)$9 rate for both map and pushforward approximation (Geshkovski et al., 9 Feb 2026). This places conditional normalizing flow nets within rigorous approximation theory rather than purely empirical architecture design.

Symmetry guarantees appear prominently in graph and physics applications. The semi-equivariant CNF for 3D graphs proves that if the complex-to-complement mapping is rotation semi-equivariant and permutation semi-equivariant, then the induced conditional density satisfies

$$p(y\mid x)=p_Z(T_\theta(y)\mid x)\left|\det J_{T_\theta}(y)\right|.$$0

This is achieved by coupling EGNN-based equivariant dynamics with an explicit affine transformation that removes translation dependence (Rozenberg et al., 2023). In lattice field theory, locality-constrained autoregressive conditional flows exploit the fact that a slice’s conditional law depends only on a low-dimensional dependency surface, not the full lattice, and the resulting sampler is corrected by independent Metropolis–Hastings to ensure asymptotic exactness (R., 2023).

Algorithmic consistency theorems arise when conditional flows are embedded in larger estimators. CKR-DDUQ proves that, under domain-decomposition convergence, support coverage, i.i.d. target samples, and a universal approximation assumption for cKRnet, the probability estimates obtained with cKRnet-based importance weights converge to the true probabilities (Li et al., 2024). PL-MCMC proves that a Metropolis–Hastings chain in latent space converges to the exact conditional distributions associated with a trained normalizing flow (Cannella et al., 2020). CNF-iVI establishes measure-theoretic conditions under which infinite-dimensional flow transforms remain equivalent to the Gaussian prior and proves discretization invariance for several functional flow families (Zhao et al., 2024).

A recurring misconception is that all “conditional normalizing flow nets” are architecturally similar. The literature instead supports a broader taxonomy: some condition the bijection directly (Xu et al., 2023, Wen et al., 2022), some condition only the latent prior (Zeng et al., 6 Jul 2025, Nguyen et al., 2019), some operate in continuous time (Nguyen et al., 2019), and some are constructed via transport equations or infinite-dimensional quasi-invariant transforms (Geshkovski et al., 9 Feb 2026, Zhao et al., 2024). The shared object is the conditional density model with explicit invertibility, not a single canonical network blueprint.

6. Limitations, trade-offs, and future directions

The main limitations reported across the literature concern rare tails, extrapolation, dimensional scaling, physical constraints, and computational cost. In detector-response modeling, extreme tails and systematic shifts absent from the training targets remain challenging, and robustness under domain shift such as altered pile-up regimes is explicitly identified as an open issue (Xu et al., 2023). In the radiative surrogate, the lack of built-in constraints means the model can place probability mass outside $$p(y\mid x)=p_Z(T_\theta(y)\mid x)\left|\det J_{T_\theta}(y)\right|.$$1 and violate $$p(y\mid x)=p_Z(T_\theta(y)\mid x)\left|\det J_{T_\theta}(y)\right|.$$2; the authors identify logistic, logit, softmax, or physics-informed layers as natural remedies (Seyedheydari et al., 27 Aug 2025). In distribution correction, out-of-support target covariates can still cause extrapolation problems even though the flow avoids weight explosion (Algren et al., 2023).

Sampling speed and tractability remain an architectural trade-off. MAF is preferred in density-estimation-centric settings because log-likelihood evaluation is fast and exact, but conditional sampling is slower due to the autoregressive inverse (Xu et al., 2023). IAF inverts this trade-off, while RealNVP-style couplings are faster to sample but may require deeper stacks or richer spline parameterizations for comparable expressivity (Xu et al., 2023). Continuous flows introduce solver overhead and sensitivity to numerical tolerances; InfoCNF explicitly treats tolerances as learnable control variables because the number of function evaluations dominates compute (Nguyen et al., 2019).

Several papers emphasize that conditional-flow performance depends strongly on how conditioning is represented. AP-CDE argues that conditioning via a structured latent base can yield better interpretability than directly modulating every layer with the external variable (Zeng et al., 6 Jul 2025). CNF-iVI shows that, in infinite-dimensional inverse problems, admissibility constraints force the conditioning mechanism to be lightweight and basis-based rather than arbitrarily expressive (Zhao et al., 2024). Semi-equivariant graph flows compress the receptor into invariant signatures rather than operating on the full joint receptor–ligand graph, because the receptor can be one to two orders of magnitude larger than the ligand (Rozenberg et al., 2023).

Future work is correspondingly diverse. Detector-response modeling points toward higher-dimensional observables, variable multiplicities, spline flows, transformer-based conditioners, and direct integration of selection efficiencies and uncertainty propagation into analysis pipelines (Xu et al., 2023). Constructive theory suggests exploiting triangularity, multiscale grids, and Maurey-type approximations to reduce switch complexity (Geshkovski et al., 9 Feb 2026). Domain applications repeatedly call for comparisons with GANs, VAEs, and diffusion models to clarify trade-offs among stability, sampling speed, and exact-likelihood advantages (Xu et al., 2023, Seyedheydari et al., 27 Aug 2025).

Taken together, the literature presents the conditional normalizing flow net as a general paradigm rather than a single architecture: a conditioned invertible model with explicit density evaluation, adaptable to discrete flows, ODE-based flows, constructive transport schemes, and infinite-dimensional settings. Its central appeal is unchanged across domains—conditional sampling, exact or tractable likelihoods, and structured inductive biases such as triangularity, equivariance, or measure preservation—but the technical realization depends strongly on the geometry and constraints of the target problem (Xu et al., 2023, Geshkovski et al., 9 Feb 2026, Rozenberg et al., 2023, Zhao et al., 2024).