Conditional Invertible Neural Networks

• Conditional Invertible Neural Networks (cINNs) are conditional normalizing flows that create bijective mappings between data and latent space, enabling exact likelihood evaluation and efficient inference.
• They leverage conditional affine coupling layers with specialized subnetworks to incorporate auxiliary data while preserving invertibility, making them adaptable across applications.
• cINNs support multimodal conditional distributions and rapid sampling, proving ideal for uncertainty quantification in inverse problems and generative tasks in various domains.

Searching arXiv for relevant cINN papers to ground the article. Conditional invertible neural networks (cINNs) are conditional normalizing flows that realize an invertible, continuously differentiable mapping between target variables and a tractable latent space under explicit conditioning on observations or auxiliary variables. In the standard formulation, a cINN learns a family of bijections $f(\cdot;c)$ such that $z=f(x;c)$ and $x=f^{-1}(z;c)$, with $z$ typically constrained to follow a standard Gaussian. This permits exact conditional likelihood evaluation through the change-of-variables formula, efficient posterior sampling by inversion, and direct representation of multimodal conditional distributions in inverse problems, generative modeling, and conditional synthesis (Ardizzone et al., 2019).

1. Definition and mathematical formulation

A cINN models a conditional density $p(x\mid c)$ by combining an invertible transformation with a simple latent prior. In the notation used across multiple works, the defining relations are

$z = f(x; c), \qquad x = f^{-1}(z; c),$

with $p_Z(z)=\mathcal N(0,I)$ in the common case. The conditional density then follows from change of variables: $$p_X(x \mid c) = p_Z(f(x;c)) \cdot \left|\det \frac{\partial f(x;c)}{\partial x}\right|.$$ Equivalent forms appear throughout the literature for targets $y$, parameters $\theta$, or latent video codes $z=f(x;c)$0, but the underlying construction is the same: the condition affects the transformation while invertibility is preserved with respect to the variable of interest (Phipps et al., 2023).

With a Gaussian base density, maximum-likelihood training reduces to minimizing a negative log-likelihood consisting of a quadratic latent term and a Jacobian log-determinant term. Representative expressions include

$z=f(x;c)$1

for conditional forecasting, and

$z=f(x;c)$2

for conditional inverse design or image translation (Phipps et al., 2023). Several papers emphasize that this exact likelihood is a defining difference relative to adversarial or variational alternatives. In some applications, the cINN is trained as a direct posterior model $z=f(x;c)$3; in others, it models a conditional latent distribution that is subsequently decoded by another network, as in stochastic image-to-video synthesis (Dorkenwald et al., 2021).

Because the mapping is bijective for fixed condition, sampling and inference are symmetric. One may encode an observed target into latent space for analysis, or sample $z=f(x;c)$4 and invert to obtain conditional samples. This supports posterior estimation in scientific inverse problems, diverse conditional synthesis in vision, and probabilistic forecasting without modifying the underlying deterministic forecaster (Haldemann et al., 2022).

2. Architectural principles

Most cINNs in the cited literature are built from conditional affine coupling layers in the RealNVP or Glow family. A standard conditional affine coupling layer splits an input $z=f(x;c)$5 and transforms only one subset: $z=f(x;c)$6 with inverse

$z=f(x;c)$7

Because the Jacobian is triangular, the log-determinant is simply the sum of the scale outputs, making exact density evaluation computationally tractable (Ardizzone et al., 2021).

Conditioning enters through the non-invertible subnetworks that produce the scale and shift parameters. This is a crucial design principle: the conditioning variable can be injected into $z=f(x;c)$8 and $z=f(x;c)$9 without breaking invertibility because these subnetworks are evaluated only in the forward direction. The condition may be concatenated directly, processed by a dedicated feature extractor, or provided at multiple scales. Image-based cINNs use feed-forward conditioning networks such as ResNet-18, VGG-like trunks, or task-specific CNNs to produce feature pyramids that modulate coupling layers at each resolution (Ardizzone et al., 2019).

Architectural variants reflect domain requirements. Image models combine conditional coupling blocks with Haar wavelet downsampling, random or learned channel permutations, split-merge multiscale routing, and convolutional coupling subnetworks (Ardizzone et al., 2021). Scientific inverse models for low-dimensional parameters typically use fully connected subnetworks and random orthogonal permutations in FrEIA-based implementations (Haldemann et al., 2022). Spectral and time-series applications often use dense conditioners because both inputs and outputs are vector-valued (Phipps et al., 2023). In video synthesis, the flow operates on a compact latent $x=f^{-1}(z;c)$0 rather than directly on pixels, with 20 invertible blocks and conditioning features extracted from the initial frame by a pretrained image autoencoder (Dorkenwald et al., 2021).

Several stabilization devices recur across papers. These include actnorm, fixed or invertible permutations, clamped or $x=f^{-1}(z;c)$1-parameterized scale outputs, small-noise dequantization for images, and multi-scale factorizations that reduce compute and memory. This suggests that practical cINN design is less about a single canonical architecture than about preserving exact invertibility while adapting the conditioning pathway and coupling subnetworks to the structure of the conditioned variable.

3. Training objectives and inference modes

The canonical training criterion is conditional maximum likelihood. In practice, cINNs are optimized by minimizing the empirical negative log-likelihood over paired samples $x=f^{-1}(z;c)$2, or over $x=f^{-1}(z;c)$3 in inverse problems. Since the latent prior is usually standard Gaussian, the loss has a simple form involving the latent norm and the Jacobian log-determinant. This yields a single-objective likelihood-based training procedure in image generation (Ardizzone et al., 2019), probabilistic forecasting (Phipps et al., 2023), domain transfer (Dreher et al., 2023), and thin-film inverse design (Luce et al., 2022).

Some applications augment the pure flow objective. In exoplanet formation inference, the total loss combines the exact conditional flow likelihood with an $x=f^{-1}(z;c)$4 reconstruction penalty,

$x=f^{-1}(z;c)$5

which was reported to stabilize training on relatively small datasets (Burn et al., 5 Dec 2025). In unsupervised sim-to-real transfer, maximum-likelihood terms are combined with LSGAN discriminator losses, while cycle consistency is obtained exactly from invertibility rather than by an auxiliary penalty (Dreher et al., 2023). In stochastic image-to-video synthesis, the cINN is trained on a latent code produced by a conditional VAE-GAN, so the flow likelihood is only one stage of a larger two-stage model (Dorkenwald et al., 2021).

Inference proceeds in several distinct modes. In conditional generation, one samples $x=f^{-1}(z;c)$6 and computes $x=f^{-1}(z;c)$7. In amortized Bayesian inversion, repeated latent draws produce approximate posterior samples $x=f^{-1}(z;c)$8, from which marginal estimates, credible intervals, or MAP approximations can be derived (Candebat et al., 2024). In analysis-oriented applications, the forward map is used to extract latent codes from observed targets; these may then be edited, transferred, or perturbed before inversion, as in motion transfer for video or style transfer in image colorization (Ardizzone et al., 2021). In probabilistic forecasting, a deterministic point forecast $x=f^{-1}(z;c)$9 is first mapped into latent space, perturbed locally by Gaussian noise, and then mapped back to obtain a predictive distribution centered around the deterministic model’s output (Phipps et al., 2023).

A plausible implication is that cINNs are most distinctive when both directions of the mapping are operationally useful: forward evaluation supplies exact likelihoods or latent analyses, while inversion supplies samples, reconstructions, or posterior draws.

4. Major application domains

cINNs have been applied across a wide range of conditional synthesis and inverse-problem settings. In natural image generation, early work demonstrated class-conditional MNIST generation and diverse ImageNet colorization with a maximum-likelihood-trained cINN that combines an invertible backbone with a non-invertible conditioning network (Ardizzone et al., 2019). A later formulation extended the same design to natural image-to-image translation, including day-to-night translation and colorization, with eight conditional coupling blocks and five Haar wavelet downsamplings for day-to-night translation, and a deeper multiscale model for colorization (Ardizzone et al., 2021).

In video generation, cINNs have been used to impose a one-to-one mapping between a residual dynamics code and a compact video latent given the first frame. The method in “Stochastic Image-to-Video Synthesis using cINNs” uses a conditional VAE for video latent autoencoding and a conditional flow on the resulting $z$0-dimensional latent, enabling stochastic synthesis, controllable synthesis via an auxiliary control $z$1, and motion transfer by inversion of the learned residual bridge (Dorkenwald et al., 2021).

In probabilistic forecasting, cINNs have been used as post hoc uncertainty models that convert arbitrary deterministic forecasts into predictive distributions. The conditioning vector includes calendar features and optionally exogenous inputs, while the deterministic forecast itself is mapped into latent space and locally perturbed. Reported results show systematic CRPS improvements over Gaussian residual baselines and competitive performance against DeepAR, QRNN, and NNQF across Electricity, Bike, OPSD, and Price datasets (Phipps et al., 2023).

Scientific inverse problems form another major application class. In exoplanet characterization, cINNs approximate the posterior over interior structure parameters conditioned on planetary mass, radius, and refractory ratios, reproducing Metropolis–Hastings posteriors for K2-111 b while yielding much faster inference once trained (Haldemann et al., 2022). In exoplanet formation inference, cINNs map observed planet properties to posterior samples over disk and formation parameters using a synthetic training population from a global planet formation model (Burn et al., 5 Dec 2025). In stellar spectroscopy, cINNs have been used to infer full posterior distributions of atmospheric parameters and abundances from GIRAFFE or 4MOST-like spectra, with reported accuracies of $z$2 K in $z$3, $z$4 dex in $z$5, and $z$6 dex in $z$7 for high-quality Gaia-ESO spectra in one study, and survey-scale throughput for 4MOST synthetic NLTE spectra in another (Candebat et al., 2024).

Further applications include photoacoustic imaging uncertainty quantification (Nölke et al., 2020), radio map reconstruction from time-ordered data (Zhang et al., 2023), sound-field reconstruction in reverberant environments (Karakonstantis et al., 2024), detector unfolding in high-energy physics (Backes et al., 2022), domain transfer in spectral imaging (Dreher et al., 2023), photonic and thin-film inverse design (Frising et al., 2022, Luce et al., 2022), and cosmic-ray source-property inference from propagated observables (Bister et al., 2021). This breadth reflects a common pattern: the observation-to-parameter map is non-unique, but paired simulations or training data are available.

5. Advantages, expressive capabilities, and recurrent claims

Across the literature, three properties recur. First, cINNs provide exact conditional likelihoods. This follows from the change-of-variables formula and tractable Jacobians of coupling layers, and is explicitly contrasted with GAN-based methods and with approximate variational objectives in multiple works (Ardizzone et al., 2021).

Second, cINNs support direct representation of multimodal conditional distributions. This is especially salient in inverse problems where many distinct parameter settings produce nearly identical observations. In nanophotonic inverse design, the posterior over slit–grating geometries is at least bimodal because mirror-reflected asymmetric devices generate identical spectra; the cINN resolves the corresponding modes in the conditional distribution $z$8 and samples from both (Frising et al., 2022). In photoacoustic imaging, multi-modal posteriors arise when different tissue configurations yield similar multispectral signatures, and the cINN exposes this ambiguity rather than collapsing it into a point estimate (Nölke et al., 2020). In video, stochastic futures are represented through a residual latent that remains in one-to-one correspondence with generated motion under a fixed initial frame (Dorkenwald et al., 2021).

Third, cINNs are repeatedly described as efficient at inference time once training is complete. This is central in exoplanet characterization, where the database is computed once and the trained cINN then amortizes posterior inference across many planets (Haldemann et al., 2022); in cosmic-ray inference, where a trained cINN evaluates posteriors in seconds while MCMC requires multiple chains and hours per scenario (Bister et al., 2021); and in sound-field reconstruction, where amortized posterior sampling or MAP estimation becomes far faster than hierarchical Bayesian inference based on MCMC (Karakonstantis et al., 2024).

Several papers also stress exact cycle consistency or bidirectionality. In unsupervised domain transfer, invertibility makes $z$9 exact, obviating an explicit cycle-consistency loss (Dreher et al., 2023). In image translation and colorization, the same bidirectionality enables latent interpolation, style transfer, and latent editing (Ardizzone et al., 2019). This suggests that cINNs are particularly valuable when uncertainty quantification, sample diversity, and reverse analysis are all required simultaneously.

6. Limitations, failure modes, and relation to adjacent methods

Despite their breadth, cINNs inherit several limitations noted repeatedly in the cited papers. A central issue is dependence on simulation quality or training coverage. In detector unfolding, a cINN trained once inherits the Monte Carlo prior and can be biased when simulation mismodels the true distribution, motivating the iterative cINN variant with classifier-based reweighting (Backes et al., 2022). In exoplanet formation inference, single-planet training covers only a small range of observable space and yields overconfident extrapolation in unsampled regions (Burn et al., 5 Dec 2025). In stellar spectroscopy, out-of-distribution regimes such as very metal-poor stars or cool giants show degraded accuracy, albeit with inflated uncertainties (Candebat et al., 2024).

Another limitation is that invertibility does not remove the need for adequate support coverage. In thin-film inverse design, single-shot cINN proposals are often only approximate and require local optimization for high-precision designs; the paper explicitly advocates a hybrid pipeline in which cINN samples seed Nelder–Mead or OpenFilters refinement (Luce et al., 2022). In photonic inverse design, flow continuity can create low-density bridges between distinct modes because a connected Gaussian base is continuously deformed into a connected manifold (Frising et al., 2022). In photoacoustic and sound-field reconstruction, uncertainty estimates remain sensitive to mismatches between the training simulator and real experimental noise (Nölke et al., 2020, Karakonstantis et al., 2024).

The relation to adjacent methods is also clarified in the literature. Relative to cGANs, cINNs are repeatedly associated with exact likelihoods, stable non-adversarial training, and resistance to mode collapse (Ardizzone et al., 2019). Relative to cVAEs, cINNs avoid the need to approximate multimodal posteriors with simple latent families and retain exact invertibility, which can be decisive in highly multimodal inverse design (Frising et al., 2022). Relative to MCMC, cINNs trade an up-front training cost for rapid amortized inference; several papers report close agreement between cINN and MCMC posteriors while emphasizing that the cINN becomes advantageous when many targets share the same forward model (Haldemann et al., 2022). Relative to classical optimization or iterative inverse methods, cINNs provide a sampleable conditional posterior rather than a single solution.

A plausible implication is that cINNs are best understood not as universal replacements for Bayesian computation or conditional generation, but as a specific likelihood-based family that is especially effective when three conditions hold simultaneously: paired training data or simulators are available, posterior multimodality matters, and repeated inference justifies amortization.