Hybrid Quantum/Classical GANs

Updated 2 September 2025

Hybrid quantum/classical GANs are architectures that combine quantum parameterized circuits for generation with classical discriminators to achieve resource-efficient generative modeling.
They leverage design patterns such as patch architecture, latent space fusion, and quantum-correlated latent priors to enable scalable training and mitigate issues like barren plateaus.
Empirical results show these hybrid models generate high-quality outputs with fewer parameters and faster convergence compared to traditional GANs.

A hybrid quantum/classical GAN approach refers to a generative adversarial network (GAN) architecture in which quantum and classical computational elements are jointly employed—typically a quantum generator instantiated by a parameterized quantum circuit (PQC) and a classical (or sometimes quantum) discriminator, often with additional pre- or post-processing neural networks. Recent research demonstrates diverse design choices, ranging from PQC-based sample generation to quantum-correlated prior sampling and latent space fusion, to address the unique challenges and leverage the strengths of both quantum and classical paradigms for generative modeling tasks.

1. Hybrid GAN Architectures

Hybrid GANs integrate distinct quantum and classical components, with the most prevalent form featuring a quantum generator and a classical discriminator. The generator is typically built via a parameterized quantum circuit, where layers of single- and two-qubit gates (e.g., $R_x$ , $R_y$ , $R_z$ , controlled-phase, or CNOT gates) act on an initial product state $|0\rangle^{\otimes N}$ to generate a quantum state $|\psi(\vec \theta)\rangle$ . The output is produced either by measuring in the computational basis (for discrete samples such as bitstrings) or by extracting expectation values of observables (for continuous-valued features) (Situ et al., 2018, Romero et al., 2019, Tsang et al., 2022).

A classical discriminator, usually a feed-forward neural network or convolutional network, evaluates the authenticity of samples—either real or generated—and is trained by standard adversarial losses (binary cross-entropy or Wasserstein loss). Notably, in some variants, hybridization is extended to include quantum circuits in both the generator and discriminator roles (Boyle et al., 2023, Al-Othni et al., 13 Jul 2025), or by fusing classical neural modules (autoencoders, variational decoders, etc.) directly into the architecture (Chang et al., 4 Jun 2024, Vieloszynski et al., 22 Sep 2024, Thomas et al., 16 Sep 2024).

Key design patterns include:

Patch architecture: Divide large images into patches, each synthesized by a small PQC, improving scalability and mitigating barren plateaus (Tsang et al., 2022, Yang et al., 21 Nov 2024).
Latent space fusion: Employ classical autoencoders to compress data into a manageable latent space, on which quantum modules operate, allowing the synthesis of high-resolution data within current qubit limitations (Chang et al., 4 Jun 2024, Vieloszynski et al., 22 Sep 2024, Jiao et al., 26 Jun 2025).
Hybrid latent priors: Replace standard i.i.d. classical priors with quantum-correlated noise sampled from quantum circuits, often hybridized via convex mixing with the classical prior (Jin et al., 2 Jul 2025, Goh, 10 Aug 2025).

2. Quantum Generator and Encoding Methods

Quantum generators employ parameterized quantum circuits designed for NISQ compatibility, usually with shallow depth and native gate sets. Standard elements include:

Single-qubit rotation gates: $R_y(\theta)$ , $R_x(\theta)$ , $R_z(\theta)$ , with parameters modulated by classical noise or latent vectors.
Entangling gates: Controlled-phase (CP), CNOT, or controlled-Z gates, introducing learnable entanglement.
Encoding strategies: Angle encoding (direct mapping of classical latent values to rotation angles), amplitude encoding (embedding normalized vectors as amplitudes), or data re-uploading wherein latent noise is re-injected at each layer to maximize expressibility (Romero et al., 2019, Chang et al., 4 Jun 2024, Tsang et al., 2022).

Samples are produced by measuring the PQC. For discrete outputs, measurement in the computational basis directly yields samples. For continuous data, expectation values over a set of observables (e.g., Pauli $X$ , $Z$ operators) are computed and usually post-processed by shallow classical neural layers to increase nonlinearity (Romero et al., 2019, Shu et al., 2 Feb 2024, Tsang et al., 2022).

Gradients and Optimization

Gradients with respect to PQC parameters are computed via the parameter-shift rule: $\frac{\partial P_{\theta}(x)}{\partial \theta} = \frac{1}{2}\left[P_{\theta+\frac{\pi}{2}}(x) - P_{\theta-\frac{\pi}{2}}(x)\right]$ which allows exact gradient estimation with two shifted-circuit evaluations (Situ et al., 2018, Romero et al., 2019). Integration with classical post-processing (e.g., neural networks) is achieved via hybrid automatic differentiation or by approximating gradients (finite differences) when full backpropagation is not supported.

3. Integration Strategies and Training Schemes

Hybrid loss calculation applies adversarial objectives, typically: $\mathcal{L}(D, G) = \mathbb{E}_{x\sim P_{\mathrm{data}}}[\log D(x)] + \mathbb{E}_{z\sim P(z)}[\log (1 - D(G(z)))]$ or its Wasserstein or WGAN-GP analogues, sometimes generalized to quantum trace expressions if both networks are quantum (Nokhwal et al., 2023, Jurasz et al., 2023).

Practical training involves:

Alternating parameter updates: Classical optimizer (e.g., Adam, SGD) is used for both classical and quantum parameters, with hybrid automatic differentiation frameworks supporting the chain of derivatives across quantum and classical layers (Romero et al., 2019, Tsang et al., 2022).
Multi-circuit batch/patch processing: Batch learning using quantum superposition (amplitude encoding multiple data in a single quantum state) or patch-based parallelization, which accelerates training and gradient computation (Huang et al., 2020, Tsang et al., 2022, Yang et al., 21 Nov 2024).
Transfer learning: Injection of pretrained classical feature extractors (e.g., ResNet-18), especially effective in discriminators, improves convergence and stability (Al-Othni et al., 13 Jul 2025).

4. Performance Metrics and Empirical Results

Hybrid quantum/classical GANs are evaluated on standard metrics:

Frechet Inception Distance (FID): $FID=x_g=\|\mu_x-\mu_g\|_2^2 + Tr(\sigma_x + \sigma_g - 2(\sigma_x\sigma_g)^{1/2})$
Kernel Inception Distance (KID), Inception Score (IS), Jensen-Shannon Divergence (JSD), Wasserstein Distance, Number of Distinct Bins (NDB), Structural Similarity (SSIM), and Peak Signal-to-Noise Ratio (PSNR)
KL/JS Divergences: used for distributional similarity between generated and real sample distributions (Boyle et al., 2023).

Empirical findings include:

Quantum generators can generate images of comparable or improved quality over classical GANs with dramatically fewer parameters (e.g., $10^3$ versus $10^6$ ) (Tsang et al., 2022, Shu et al., 2 Feb 2024, Jiao et al., 26 Jun 2025).
Training efficiency (convergence iterations, parameter count, and time-to-convergence) is often superior in hybrid models (Shu et al., 2 Feb 2024, Tsang et al., 2022).
The use of quantum-correlated latent priors derived from hardware circuits (rather than classical i.i.d noise) consistently improves FID, achieving up to 17% improvement in the case of BigGAN on CIFAR-10 (Jin et al., 2 Jul 2025).
Modular, patch-based, and invertible/reversible circuit architectures enable scalable and stable training, reduce the risk of mode collapse and barren plateaus, and provide parameter efficiency (Yang et al., 21 Nov 2024, Ma et al., 9 Apr 2024).

5. Applications and Domain-Specific Adaptations

Hybrid quantum/classical GANs are being advanced for:

Synthetic data generation for classification, especially with scarce or imbalanced data (skin disease images, MNIST, Fashion-MNIST) (Jiao et al., 26 Jun 2025, He et al., 30 May 2025, Huang et al., 2020, Tsang et al., 2022).
Data augmentation for quantum and classical neural classifiers, dynamically generating samples most beneficial for downstream task performance (He et al., 30 May 2025).
Latent representation modeling (e.g., image-to-image translation, anomaly detection, style transfer, risk management) by fusing with autoencoder representations (Chang et al., 4 Jun 2024, Vieloszynski et al., 22 Sep 2024, Yang et al., 21 Nov 2024).
Unsupervised image-to-image translation using invertible quantum generators with unidirectional cycle consistency aided by classical neural networks (Yang et al., 21 Nov 2024).
Learning quantum or hybrid quantum-classical generative models for high-dimensional classical and quantum data (state preparation, phase diagrams, quantum simulation) (Jurasz et al., 2023, Ma et al., 9 Apr 2024).

6. Technical and Resource Considerations

Significant operational, resource, and algorithmic constraints inform hybrid model design:

Qubit count and patching: Hybrid methods often patch images into segments or operate in compressed latent spaces to minimize per-circuit qubit requirements, enabling the training of full images or higher-dimensional data on NISQ devices (Tsang et al., 2022, Chang et al., 4 Jun 2024).
Parameter count reduction: Hybrid quantum/classical generators often achieve orders-of-magnitude fewer parameters than classical GANs with only modest reduction (or sometimes improvement) in output quality (Tsang et al., 2022, Shu et al., 2 Feb 2024, Jiao et al., 26 Jun 2025).
Mitigation of NISQ constraints: Modular circuit depth, trainable encoding blocks, and careful circuit parameter initialization (e.g., small-angle initialization) mitigate the impact of noise, barren plateaus, and convergence challenges (Boyle et al., 2023, Chang et al., 4 Jun 2024, Tsang et al., 2022).
Integration with standard machine learning frameworks and HPC/quantum hardware co-design: Training leverages existing classical automatic differentiation (AD) stacks, mini-batch SGD, Qiskit/PyTorch/MindSpore Quantum, and cloud-based quantum processors (Boyle et al., 2023, Shu et al., 2 Feb 2024, Tsang et al., 2022, Jiao et al., 26 Jun 2025).

7. Current Challenges and Future Research Trajectories

Open challenges recognized in current literature include:

Expressivity and mode collapse: Effective latent space design and regularization (e.g., VAE-QWGAN with a Gaussian mixture prior fit to the training latent vectors) are employed to address missing mode coverage and enhance sample diversity (Thomas et al., 16 Sep 2024).
Sample diversity and nonlinearity: Hybrids that insert classical neural nonlinearities either before or after the quantum generator, or that repeatedly “re-upload” the noise across PQC layers, increase expressive capacity to mitigate issues associated with the linearity of quantum evolution (Romero et al., 2019, Ma et al., 9 Apr 2024, Chang et al., 4 Jun 2024).
Integration complexity: Ensuring seamless gradient propagation between classical and quantum modules, managing varying resource constraints, and jointly optimizing hybrid loss landscapes across disparate computational substrates (Romero et al., 2019, Tsang et al., 2022, Shu et al., 2 Feb 2024).
Scaling to larger and more realistic datasets: Research is extending patch models, latent compression, and distributed implementations for tasks such as skin disease color image generation, earth observation imagery, and high-dimensional tabular data (Jiao et al., 26 Jun 2025, Chang et al., 4 Jun 2024, Vieloszynski et al., 22 Sep 2024).
Noise resilience and quantum hardware deployment: Quantitative experiments now demonstrate robustness of hybrid models under real device noise, including IBM 127-qubit Eagle (Jiao et al., 26 Jun 2025, Vieloszynski et al., 22 Sep 2024).

Future research is focused on improving expressibility, harnessing quantum-correlated priors for inductive bias (Jin et al., 2 Jul 2025, Goh, 10 Aug 2025), developing fully invertible flows with quantum blocks (Zhang et al., 22 May 2024), optimizing for patch shapes and circuit features (Tsang et al., 2022), and integrating advanced quantum autoencoding and cycle-consistency paradigms for domain translation and outlier detection (Yang et al., 21 Nov 2024, Chang et al., 4 Jun 2024, Thomas et al., 16 Sep 2024).

Hybrid quantum/classical GANs, by leveraging parameterized quantum circuits within established adversarial learning frameworks and interfacing with classical models for discrimination, autoencoding, and nonlinear transformation, have demonstrated the unique potential to efficiently synthesize both discrete and continuous data at reduced resource cost and improved generative diversity. They circumvent obstacles such as vanishing gradient in discrete output domains (Situ et al., 2018), facilitate scalable training under NISQ constraints (Tsang et al., 2022, Huang et al., 2020), and provide a blueprint for increasingly complex generative modeling tasks in the quantum machine learning landscape.