Hybrid Quantum-Classical GAN Architectures

Updated 6 July 2026

Hybrid quantum-classical GANs are generative models that integrate parameterized quantum circuits for compact latent feature extraction with classical networks for high-dimensional data reconstruction.
They employ design patterns such as patchwise generation, latent prior sampling, and reversible architectures to overcome NISQ limitations and improve optimization stability.
Empirical results indicate reduced parameter counts and faster convergence compared to classical GANs, though challenges in scalability and training stability remain.

Searching arXiv for the cited works to ground the article. Hybrid quantum-classical GAN architectures are generative adversarial networks in which at least one major adversarial component is implemented with parameterized quantum circuits while the remaining computation stays classical. Across the literature, the most common pattern is a quantum generator coupled to a classical discriminator, although variants also place quantum modules in discriminators, latent priors, or both adversarial players (Shu et al., 2024). These architectures are motivated by two recurring constraints and opportunities: first, current noisy intermediate-scale quantum hardware favors shallow, low-qubit circuits rather than end-to-end quantum models; second, classical neural networks remain better suited for high-dimensional decoding, feature extraction, and optimization. The resulting systems typically use quantum circuits to generate low-dimensional latent features, probability vectors, or patchwise outputs, then rely on classical networks for image, tabular, or signal-space reconstruction and discrimination (Tsang et al., 2022). Recent work spans handwritten-image synthesis, latent-space image generation, unsupervised image-to-image translation, hyperspectral restoration, network-flow synthesis, and tabular data synthesis, with architectures differing primarily in where the quantum component is inserted and how quantum outputs are interfaced with classical layers (Vieloszynski et al., 2024).

1. Architectural taxonomy

Hybrid quantum-classical GANs are most naturally classified by which adversarial components are quantum. A standard reference taxonomy distinguishes fully quantum GANs, quantum-generator/classical-discriminator models, classical-generator/quantum-discriminator models, and broader “quantum-enhanced classical GANs” where only an auxiliary subroutine or latent sampler is quantum (Islam et al., 22 Jun 2025). In practice, the dominant design is the quantum-generator/classical-discriminator configuration, because it keeps the discriminator in a regime where backpropagation, large input dimensionality, and strong optimization heuristics are already well established (Shu et al., 2024).

A first major subtype replaces the generator with a variational quantum circuit followed by a small classical head. In the QC-GAN for MNIST, the generator consists of a quantum variational circuit together with a one-layer neural network, while the discriminator is a traditional neural network with two fully connected hidden layers of 64 and 32 units (Shu et al., 2024). A closely related pattern appears in QTabGAN, where a quantum variational circuit produces a $2^n$ -dimensional probability vector and a classical mapper converts it into a tabular row, again paired with a classical discriminator (Kumari et al., 13 Feb 2026). In both cases, the quantum part provides a compact latent transformation, and the classical part performs dimensional expansion into the target data space.

A second subtype uses the quantum circuit only as a latent prior. In HQCGAN, the generator and discriminator are classical neural networks, and the only quantum component is a fixed $n$ -qubit latent sampler with $n \in \{3,5,7\}$ , measured in the computational basis and linearly rescaled from bitstrings to $[-1,1]^n$ (Goh, 10 Aug 2025). A more elaborate instance replaces the standard Gaussian prior with a quantum-correlated Gaussian copula constructed from measurements of a 16-qubit entangling circuit on IBM hardware, while leaving WGAN, SNGAN, and BigGAN architectures unchanged (Jin et al., 2 Jul 2025). This suggests a distinct architectural regime in which quantum resources are used to shape the prior rather than to implement the generator map itself.

A third subtype inserts quantum circuits into both adversarial players. HyperKING uses hybrid generators and hybrid discriminators for hyperspectral image restoration, with convolutional read-in and read-out shells around quantum cores (Lin et al., 16 Apr 2025). A transfer-learning study on CIFAR-10 compares four variants—fully classical, quantum discriminator only, quantum generator only, and fully hybrid—and reports the strongest FID and KID for the fully hybrid setting (Al-Othni et al., 13 Jul 2025). By contrast, a conceptual Hamiltonian-based framework models both generator and discriminator as quantum unitaries controlled by classical optimization, but remains at the level of formalism rather than an implementation-ready gate architecture (Nokhwal et al., 2023).

A fourth subtype is specialized to reversibility or latent-space compression. LatentQGAN couples a classical convolutional autoencoder with a bank of quantum sub-generators and a classical discriminator, so that the quantum system models a low-dimensional latent representation rather than raw $28\times 28$ pixels (Vieloszynski et al., 2024). iHQGAN uses two mutually approximately reversible quantum generators with shared parameters, classical assisted reconstruction networks, and classical critics to address unsupervised image-to-image translation (Yang et al., 2024). VAE-QWGAN fuses a classical VAE encoder with a quantum Wasserstein generator and classical critic, using an autoencoding prior to mitigate mode collapse (Thomas et al., 2024).

2. Quantum generator designs and encoding schemes

Most concrete hybrid QGAN generators use shallow variational circuits with angle encoding, local rotations, and sparse entangling layers. In QC-GAN, the latent noise $z=(z_1,\dots,z_N)$ is encoded by $R_y(z_i)$ on each qubit, with $N=5$ qubits and $M=4$ repeated variational blocks in the MNIST experiments (Shu et al., 2024). The variational blocks apply parameterized $R_x$ and $n$ 0 rotations plus entangling gates, giving a full unitary $n$ 1, after which measurement yields a small real feature vector that is sent to a dense layer (Shu et al., 2024). The notable architectural choice is that the quantum circuit is not asked to emit the entire image directly; instead it provides a compact, expressive intermediate representation.

A closely related but patch-based design appears in PQWGAN. There, the generator is fully quantum and consists of $n$ 2 sub-generators, each operating on $n$ 3 qubits with an encoding layer of $n$ 4 gates and $n$ 5 repeated layers of arbitrary single-qubit rotations $n$ 6 plus a nearest-neighbor CNOT chain (Tsang et al., 2022). The ancilla qubits are post-selected in the $n$ 7 state, yielding a nonlinear map from latent vector to a mixed state on the data qubits; the resulting computational-basis probabilities are normalized and truncated into image patches (Tsang et al., 2022). This patch decomposition is explicitly motivated by NISQ constraints and is used to synthesize full $n$ 8 images without dimensionality reduction (Tsang et al., 2022).

LatentQGAN modifies this pattern by shifting quantum generation to latent space. Each of its $n$ 9 sub-generators runs on $n \in \{3,5,7\}$ 0 qubits, with $n \in \{3,5,7\}$ 1 generator qubits and $n \in \{3,5,7\}$ 2 ancilla, depth $n \in \{3,5,7\}$ 3, angle embedding via $n \in \{3,5,7\}$ 4, and a one-dimensional chain of CZ entanglers (Vieloszynski et al., 2024). Partial measurement on the ancilla subsystem and renormalization produce an 8-dimensional probability vector per sub-generator, and concatenating these vectors yields a $n \in \{3,5,7\}$ 5 latent matrix that matches the row-normalized autoencoder representation (Vieloszynski et al., 2024). The architectural implication is that the quantum circuit’s natural output—a normalized probability vector—can be made structurally compatible with a classical latent manifold if the encoder is designed accordingly.

Other generator designs emphasize alternative resource trade-offs. The adversarial-network-flow QC-GAN uses $n \in \{3,5,7\}$ 6 qubits, three Successive Unitary Data Injection blocks, a ring topology of CNOT gates, and Pauli- $n \in \{3,5,7\}$ 7 expectation values as a 4-dimensional quantum feature vector, followed by a small classical post-processing MLP that outputs the selected flow features synack, ct_state_ttl, sbytes, and smean (Paudel et al., 7 May 2026). QTabGAN similarly measures all qubits repeatedly to obtain a full $n \in \{3,5,7\}$ 8-dimensional empirical probability vector, which is then passed to a classical mapper (Kumari et al., 13 Feb 2026). By contrast, HQCGAN dispenses with trainable quantum parameters entirely and uses a fixed $n \in \{3,5,7\}$ 9 circuit with realistic noise as a discrete latent source (Goh, 10 Aug 2025).

The earliest quantum-generator/classical-discriminator pattern in the supplied literature is the discrete-distribution QGAN, where the generator is a parameterized quantum circuit built from one-qubit rotations and controlled-phase gates, and measurement directly yields discrete classical samples $[-1,1]^n$ 0 with probability $[-1,1]^n$ 1 (Situ et al., 2018). That architecture was introduced for discrete distributions specifically because the generator’s underlying Born probabilities remain differentiable with respect to gate parameters even though the observed outputs are discrete (Situ et al., 2018). This suggests a persistent theme in hybrid QGAN design: measurement-induced discreteness does not preclude gradient-based training if the quantum layer is treated as a differentiable probabilistic generator.

3. Classical interfaces, decoders, and discriminators

The defining practical problem of hybrid GAN design is the interface between low-dimensional quantum outputs and high-dimensional classical data. The most direct solution is a classical decoder or mapper. In QC-GAN, five measured quantum features are mapped by a single fully connected layer $[-1,1]^n$ 2 and bias $[-1,1]^n$ 3 to a flattened $[-1,1]^n$ 4 image (Shu et al., 2024). The paper reports 60 parameters in the QuantumLayer and 25,872 in the linear layer, a parameter budget far smaller than the classical fully connected and DCGAN baselines used in the same study (Shu et al., 2024). The quantum layer thus acts as a compressed front end, while classical decoding bears the dimensional burden.

LatentQGAN adopts a more structured classical interface via an autoencoder. A convolutional encoder compresses each MNIST image into a $[-1,1]^n$ 5 latent matrix, then a row-wise normalization enforces compatibility with the quantum sub-generators’ probability outputs (Vieloszynski et al., 2024). The decoder, trained with mean squared reconstruction loss, expands the 40-dimensional latent vector back to image space through dense and transpose-convolution layers (Vieloszynski et al., 2024). Because the discriminator operates on the latent code or decoded image through a classical FCNN with input 40, hidden layers 64 and 16, and one sigmoid output, the adversarial task is transferred from pixel space to a structured latent space that is feasible for four-qubit circuits (Vieloszynski et al., 2024).

QTabGAN pushes the same interface principle into tabular data. After preprocessing continuous attributes through clipping, z-score normalization, and rescaling to $[-1,1]^n$ 6, and categorical attributes through one-hot encoding, the generator’s classical mapper consumes $[-1,1]^n$ 7 concatenated with a one-hot label vector $[-1,1]^n$ 8, and emits $[-1,1]^n$ 9 through a ReLU MLP (Kumari et al., 13 Feb 2026). This explicitly separates latent expressivity from heterogeneous output handling: the quantum circuit learns a compact probability representation, and the classical network learns how to render it as continuous and categorical fields (Kumari et al., 13 Feb 2026).

Discriminators remain predominantly classical even in strongly hybrid models. QC-GAN uses a 784-input network with hidden sizes 64 and 32 (Shu et al., 2024). LatentQGAN uses a 40-input FCNN with hidden sizes 64 and 16 (Vieloszynski et al., 2024). The adversarial-flow QC-GAN uses a classical MLP with hidden layers 128 and 64, batch normalization, Leaky ReLU, dropout 0.3, and spectral normalization under a WGAN-GP objective (Paudel et al., 7 May 2026). PQWGAN uses a fully connected critic 784 $28\times 28$ 0 512 $28\times 28$ 1 256 $28\times 28$ 2 1 with LeakyReLU activations and no activation on the final layer (Tsang et al., 2022). These choices reflect a common judgment: discrimination over large classical inputs is better delegated to conventional neural architectures.

There are, however, notable discriminator-side quantum insertions. The transfer-learning architecture on CIFAR-10 keeps a modified ImageNet-pretrained ResNet-18 backbone, projects the 512-dimensional feature vector to five dimensions, sends it through a VQC with RY data encoding, a CNOT chain, and RX–RY–RZ per qubit, and then maps the resulting expectation values to a single logit (Al-Othni et al., 13 Jul 2025). HyperKING similarly embeds a highly entangled four-qubit classifier inside a classical downsampling and sigmoid head, using amplitude encoding of 16-dimensional feature vectors, CRX entanglement modules, and Pauli- $28\times 28$ 3 measurement (Lin et al., 16 Apr 2025). These examples indicate that a quantum discriminator is most practical when classical preprocessing first compresses the input.

4. Training objectives and hybrid differentiation

The adversarial objectives in hybrid QGANs are, in most cases, inherited directly from classical GANs. QC-GAN uses the usual minimax formulation and writes the generator loss as $28\times 28$ 4, while the discriminator loss is presented in a form that is conceptually the standard real-vs-fake objective despite a noted typographical issue in the second term (Shu et al., 2024). The essential point is that the loss functional remains classical; only the parametrization of $28\times 28$ 5 changes.

Wasserstein variants are especially common because quantum generator gradients are expensive and training stability is at a premium. PQWGAN uses WGAN-GP with critic loss

$28\times 28$ 6

generator loss $28\times 28$ 7, $28\times 28$ 8, and five critic steps per generator step (Tsang et al., 2022). VAE-QWGAN extends the same WGAN-GP principle to a hybrid objective in which the quantum generator is simultaneously the VAE decoder and the GAN generator, and the critic is classical (Thomas et al., 2024). The adversarial-network-flow QC-GAN also uses WGAN-GP, spectral normalization, and $28\times 28$ 9 discriminator updates per generator update (Paudel et al., 7 May 2026). These repeated uses of WGAN-GP suggest that smooth critic geometry and gradient-penalty regularization are especially valuable when generator gradients are mediated by quantum measurements.

The standard differentiation mechanism is the parameter-shift rule. QC-GAN states

$z=(z_1,\dots,z_N)$ 0

for observables depending on a gate $z=(z_1,\dots,z_N)$ 1, and explicitly decomposes $z=(z_1,\dots,z_N)$ 2 into discriminator backpropagation, dense-layer derivatives, and quantum-layer derivatives (Shu et al., 2024). LatentQGAN likewise uses parameter-shift for the generator PQCs after backpropagating through the classical discriminator (Vieloszynski et al., 2024). The discrete-distribution QGAN derives an analogous shift identity for the Born probabilities $z=(z_1,\dots,z_N)$ 3, giving a gradient estimator in terms of discriminator evaluations under shifted generator parameters (Situ et al., 2018). The continuous-distribution variational quantum generator formalism generalizes the same idea to expectation values and emphasizes automatic differentiation across the quantum-classical graph (Romero et al., 2019).

Some later works use alternative hybrid differentiation mechanisms. The network-flow QC-GAN employs PennyLane’s lightning.qubit adjoint differentiation for the clean model and default.mixed for the noisy model, rather than explicit parameter-shift, while preserving the same chain-rule structure from discriminator loss through classical head to quantum measurement outputs (Paudel et al., 7 May 2026). The transfer-learning study notes that the VQC is implemented as a differentiable layer and that gradients are obtained through standard automatic differentiation mechanisms compatible with variational quantum algorithms (Al-Othni et al., 13 Jul 2025). Across these designs, the constant architectural fact is that optimization remains classical even when forward evaluation is quantum.

Several works modify the training loop to manage quantum instability or limited latent structure. VAE-QWGAN trains the encoder only on the VAE objective, while the quantum generator is updated with a weighted combination of reconstruction and adversarial losses, using $z=(z_1,\dots,z_N)$ 4 to balance them (Thomas et al., 2024). iHQGAN alternates critic updates with generator updates and copies shared parameters between forward and inverse quantum generators so that $z=(z_1,\dots,z_N)$ 5 implements the adjoint ordering of $z=(z_1,\dots,z_N)$ 6 (Yang et al., 2024). HyperKING pretrains the generator, then alternates training $z=(z_1,\dots,z_N)$ 7 and $z=(z_1,\dots,z_N)$ 8 in 60-epoch phases, using RMSProp rather than Adam because of instability in adversarial training (Lin et al., 16 Apr 2025). These strategies indicate that quantum-classical GAN training often benefits from explicit asymmetry or staged optimization rather than purely simultaneous minimax updates.

5. Major design patterns: patches, latent spaces, reversibility, and priors

One of the most persistent design patterns is patchwise or modular generation. PQWGAN generates a $z=(z_1,\dots,z_N)$ 9 image by dividing it into $R_y(z_i)$ 0 patches, each produced by a separate quantum sub-generator, and shows that increasing the number of patches while keeping total parameter count roughly fixed improves trainability and image quality (Tsang et al., 2022). The same modular logic appears in iHQGAN, which pads MNIST images to $R_y(z_i)$ 1, splits them into $R_y(z_i)$ 2 patches of size $R_y(z_i)$ 3, amplitude-encodes each patch into a five-qubit sub-circuit, and recomposes the measured outputs into the translated image (Yang et al., 2024). Modular patching keeps qubit count low, but it also imposes a locality prior on image generation.

A second major pattern is latent-space generation through classical compression. LatentQGAN is the clearest example: a convolutional autoencoder learns a $R_y(z_i)$ 4 latent manifold, and the quantum generator only models that latent distribution (Vieloszynski et al., 2024). VAE-QWGAN adopts a related but probabilistic variant, using a classical VAE encoder $R_y(z_i)$ 5 to generate an informed latent distribution for the quantum Wasserstein generator and then fitting a Gaussian mixture model with $R_y(z_i)$ 6 components to the learned latents at inference time (Thomas et al., 2024). In both cases, classical representation learning is used to relocate the adversarial task into a dimension where current quantum circuits are viable.

A third pattern is reversible or approximately reversible architecture. iHQGAN explicitly argues that unsupervised image-to-image translation is approximately reversible and exploits the unitarity of quantum circuits by pairing two generators $R_y(z_i)$ 7 and $R_y(z_i)$ 8 with shared parameters, where each sub-circuit in $R_y(z_i)$ 9 is the adjoint of the corresponding sub-circuit in $N=5$ 0 (Yang et al., 2024). Content consistency is not enforced by a classical CycleGAN-style dual-generator cycle, but by assisted classical neural networks $N=5$ 1 and $N=5$ 2 that reconstruct the source images through one-sided cycles $N=5$ 3 and $N=5$ 4 (Yang et al., 2024). This suggests that quantum reversibility can replace parameter duplication, while classical networks compensate for the limited expressivity of shallow PQCs.

A fourth pattern is learned or structured priors. The earliest discrete QGAN already framed the quantum generator as a differentiable sampler over discrete distributions (Situ et al., 2018). HQCGAN uses a noisy fixed quantum circuit as the latent source, and the hardware-derived latent-prior approach on CIFAR-10 goes further by constructing a quantum-correlated Gaussian copula from a 16-qubit entangling circuit, optionally mixed with the classical Gaussian prior as $N=5$ 5 (Jin et al., 2 Jul 2025). In that latter work, the generator and discriminator remain completely classical, so the hybridization is entirely in the latent source (Jin et al., 2 Jul 2025). A plausible implication is that quantum structure can improve a GAN even when no trainable quantum layer appears inside the neural architecture proper.

6. Empirical behavior, resource trade-offs, and comparative results

Across the supplied studies, hybrid architectures are repeatedly claimed to improve parameter efficiency relative to classical baselines. QC-GAN reports better FID than a classical GAN using much fewer training parameters and fewer iterations for convergence, with 60 quantum parameters and 25,872 linear-layer parameters in the generator, compared with approximately 52k parameters for GAN1, approximately 535k for GAN2, and over 3M for DCGAN (Shu et al., 2024). PQWGAN reports comparable image quality to classical WGAN-GP baselines while using roughly $N=5$ 6 fewer generator parameters, with typical quantum-generator configurations in the 5,000–7,500 parameter range versus approximately $N=5$ 7 parameters in the classical generator (Tsang et al., 2022). iHQGAN similarly reports 5,760 parameters for its shared quantum generator versus $N=5$ 8 parameters for the CycleGAN generator pair in Edge Detection and $N=5$ 9 in Font Style Transfer (Yang et al., 2024).

Image-quality comparisons are more mixed and architecture dependent. QC-GAN states that it reaches lower FID scores with fewer iterations than fully connected classical GANs and performs similarly to DCGAN despite being much shallower and smaller (Shu et al., 2024). LatentQGAN reports lower Fréchet Distance than QPatchGAN and MosaiQ in simulation, requires 676× fewer training iterations than MosaiQ to reach comparable FD, and on IBM ibm-quebec obtains FD values of 38, 43, and 34 for classes 0, 5, and 9, versus simulator values 33, 39, and 29 (Vieloszynski et al., 2024). The transfer-learning study reports FID/KID/IS values of 349.6/0.321/1.822 for the fully classical baseline, 328.4/0.251/2.351 for quantum discriminator only, 257.4/0.145/2.722 for quantum generator only, and 198.7/0.088/2.393 for the fully hybrid model (Al-Othni et al., 13 Jul 2025). These results support the specific claim that placing the VQC in both generator and discriminator can improve distributional metrics, although generator-only quantum insertion gave the highest Inception Score in that study (Al-Othni et al., 13 Jul 2025).

Mode collapse is explicitly addressed in VAE-QWGAN. On MNIST, the study reports JSD values of $M=4$ 0 for PQWGAN with Gaussian prior, $M=4$ 1 for PQWGAN with uniform prior, and $M=4$ 2 for VAE-QWGAN with a GMM prior; normalized NDB values are $M=4$ 3, $M=4$ 4, and $M=4$ 5, respectively (Thomas et al., 2024). On Fashion-MNIST, the corresponding JSD values are $M=4$ 6, $M=4$ 7, and $M=4$ 8, with NDB values $M=4$ 9, $R_x$ 0, and $R_x$ 1 (Thomas et al., 2024). These numbers support a narrow but important conclusion: learned latent priors can improve diversity and mode coverage in hybrid QGANs.

Hardware-noise effects are not uniformly detrimental. The network-flow QC-GAN compares a clean and noisy QC-GAN and reports that the noisy QC-GAN attains the best sample-wise MSE, 0.7033 versus 0.7271 for the clean QC-GAN and 0.7508 for the classical GAN, while remaining competitive in attack success against XGBoost (Paudel et al., 7 May 2026). The hardware-latent BigGAN work reports FID 4.06 for the classical baseline, 3.51 for the simulator quantum prior, 3.45 for IBM–Sherbrooke, 3.82 for IBM–Kingston, and 3.37 for dual-layout parallel hardware sampling, the last corresponding to a 17.20% reduction from baseline (Jin et al., 2 Jul 2025). These findings suggest that, in some architectures, device noise can broaden overly concentrated quantum sampling distributions and act as a beneficial inductive bias rather than a purely destructive perturbation.

The most extensive tabular synthesis results are reported by QTabGAN. Averaged over five classification datasets, it reports an accuracy difference of 2.16% and F1-score difference of 0.048 between models trained on real versus synthetic data, compared with 5.23% and 0.090 for CTAB-GAN+, 8.9% and 0.107 for CTAB-GAN, and 23.2% and 0.216 for TabularQGAN (Kumari et al., 13 Feb 2026). For regression, it reports EVS and $R_x$ 2 differences of 0.02, compared with 0.03 and 0.04 for CTAB-GAN+ and 0.16 and 3.85 for TabularQGAN (Kumari et al., 13 Feb 2026). While these results are domain specific, they illustrate a broader hybrid design principle: keeping the quantum core fixed-width and feature-agnostic can scale more gracefully than allocating qubits per feature.

7. Limitations, controversies, and future directions

The most persistent limitation is scalability under NISQ constraints. Nearly all image-generation studies either use MNIST-scale tasks, low-dimensional latent spaces, or patch decomposition (Tsang et al., 2022). The transfer-learning CIFAR-10 study uses only five qubits in each VQC block and small batch sizes because of simulation cost (Al-Othni et al., 13 Jul 2025). HQCGAN restricts itself to binary MNIST and low-dimensional latent spaces imposed by current quantum hardware (Goh, 10 Aug 2025). This suggests that present-day hybrid QGANs remain bottlenecked not by the concept of hybridization but by the number of quantum degrees of freedom that can be simulated or executed robustly.

A second limitation concerns the lack of formal quantum advantage. QC-GAN explicitly notes that it provides heuristic and empirical evidence but no rigorous quantum advantage proof (Shu et al., 2024). LatentQGAN likewise states that the evidence is empirical and limited to better FD and visual quality under fixed parameter budgets (Vieloszynski et al., 2024). The Hamiltonian-based conceptual work is even more explicit that its claims of speedup and improved generative quality are projected rather than experimentally demonstrated (Nokhwal et al., 2023). The supplied literature therefore supports only careful empirical statements, not asymptotic superiority claims.

A third limitation is training instability and barren plateaus. QC-GAN notes strong oscillations in generator and discriminator losses at early training (Shu et al., 2024). HyperKING reports that quantum adversarial learning succeeds only about 70% of the time across seeds and hardware because of hyperparameter sensitivity, although that claim is specific to its own setting (Lin et al., 16 Apr 2025). PQWGAN interprets the failure of single-patch large-circuit configurations as evidence that global cost functions on larger circuits are more vulnerable to barren plateaus, and recommends many smaller patches instead (Tsang et al., 2022). These works do not invalidate hybrid QGANs, but they underscore that trainability is still an architectural criterion, not a solved optimization problem.

There is also a recurring controversy over what counts as a “quantum GAN.” Some works, such as HQCGAN and the hardware-latent prior framework, keep both generator and discriminator classical and place the quantum circuit only in the latent sampling process (Goh, 10 Aug 2025). Other works reserve the label for architectures in which at least one adversarial network itself contains trainable quantum layers (Shu et al., 2024). This suggests a useful distinction between “quantum-latent GANs” and “quantum-network GANs,” though that terminology is not standardized in the cited papers.

Future directions in the supplied literature are comparatively consistent. Several works propose scaling to more qubits and deeper circuits as hardware improves (Shu et al., 2024). Others emphasize more complex datasets, color images, and videos (Shu et al., 2024). The transfer-learning study explicitly identifies higher-resolution synthesis, alternative quantum circuit designs, and experimentation with emerging hardware as open paths (Al-Othni et al., 13 Jul 2025). The survey literature frames the broader program as the search for hybrid architectures in which small quantum modules act as expressive engines embedded inside large classical pipelines, while classical models continue to handle optimization, high-dimensional decoding, and robust discrimination (Islam et al., 22 Jun 2025). A plausible implication is that the near-term future of hybrid quantum-classical GANs lies less in replacing classical GANs wholesale than in identifying the most leverage-bearing bottlenecks—latent modeling, compact feature transforms, or structured priors—where a small quantum module can materially improve a classical adversarial system.