Quantum Wasserstein GANs: Methods & Applications

• Quantum Wasserstein GANs are generative frameworks that replace conventional GAN discrepancies with Wasserstein-type objectives tailored for quantum-state and hybrid image generation.
• They utilize quantum semimetrics and Earth Mover’s distances with operator constraints to deliver smoother critic signals and enhance trainability under complex quantum settings.
• Empirical studies demonstrate improved fidelity in quantum-state learning and enhanced diversity in image generation while highlighting open challenges in critic design and latent modeling.

Searching arXiv for recent and foundational papers on Quantum Wasserstein GANs. Quantum Wasserstein GANs are generative adversarial frameworks that replace conventional GAN discrepancies with Wasserstein-type objectives in quantum settings. The term encompasses at least two distinct but related lines of work. One line concerns learning quantum data, where both target and generated objects are quantum states and the adversarial game is defined through a quantum Wasserstein distance or semimetric (Chakrabarti et al., 2019, Kiani et al., 2021, Jurasz et al., 2023). The other concerns classical image generation with a quantum generator, where a parameterized quantum circuit produces classical images and a classical critic is trained with a Wasserstein-1 objective, typically with gradient penalty (Thomas et al., 2024, Jäger et al., 27 Feb 2026). Across these settings, the common rationale is that Wasserstein formulations provide a smoother critic signal and improved trainability relative to overlap-based, Jensen–Shannon-based, or otherwise poorly conditioned objectives, while the differences lie in what is being generated, how the critic is constrained, and what approximation to “quantum Wasserstein” is actually used.

1. Historical emergence and problem setting

The earliest explicit formulation of a quantum Wasserstein GAN was introduced for learning quantum data in “Quantum Wasserstein Generative Adversarial Networks” (Chakrabarti et al., 2019). In that setting, the target is a density operator $Q$, the generator produces a density operator $P$, and the adversarial objective is built from a quantum Wasserstein semimetric defined by a semidefinite program over quantum couplings. The generator is an ensemble $\{(p_1,U_1),\dots,(p_r,U_r)\}$ acting on a fixed initial state, so that

$P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$

The critic is not a binary classifier but a pair of Hermitian observables $\phi\in H(X)$ and $\psi\in H(Y)$, measured on generated and target states respectively (Chakrabarti et al., 2019).

A second and now influential formulation appeared in “Learning quantum data with the quantum Earth Mover's distance” (Kiani et al., 2021). That work shifted from the semimetric of (Chakrabarti et al., 2019) to a quantum Earth Mover’s distance $D_{EM}$, described as a quantum analog of Wasserstein-1. Its central claim is that the loss geometry becomes better aligned with local quantum circuits, thereby reducing optimization pathologies such as poor local minima and exponentially decaying gradients. The resulting qWGAN is again a model for quantum-state learning, but now the critic is a sparse $k$-local Pauli Hamiltonian constrained by a tractable surrogate to a quantum Lipschitz condition (Kiani et al., 2021).

Subsequent work generalized this adversarial machinery to state preparation at unseen points of a phase diagram (Jurasz et al., 2023). There, the qWGAN is no longer used only to reproduce training states; instead, a classical model first predicts a vector of expectation values at an unseen parameter value, and the qWGAN then prepares a quantum state whose observables match that predicted vector. This reframes qWGANs as inverse-design tools from observable constraints to states (Jurasz et al., 2023).

In parallel, a separate line adapted Wasserstein GAN training to classical image generation with quantum generators. “VAE-QWGAN: Addressing Mode Collapse in Quantum GANs via Autoencoding Priors” defines a hybrid quantum-classical Wasserstein GAN in which the generator is a parameterized quantum circuit, the critic is classical, and the adversarial loss is the classical Wasserstein-1 objective with gradient penalty (Thomas et al., 2024). “Scaling Quantum Machine Learning without Tricks: High-Resolution and Diverse Image Generation” extends this image-generation line to full-resolution $32\times 32$ grayscale and color datasets using a single end-to-end quantum generator and a classical convolutional WGAN-GP critic (Jäger et al., 27 Feb 2026).

This split is essential. In the quantum-data line, “Quantum Wasserstein GAN” refers to adversarial learning directly on density operators (Chakrabarti et al., 2019, Kiani et al., 2021, Jurasz et al., 2023). In the image-generation line, it denotes a hybrid QGAN whose generator is quantum but whose loss is a classical Wasserstein objective on decoded images (Thomas et al., 2024, Jäger et al., 27 Feb 2026).

2. Quantum Wasserstein objectives and critic constraints

The 2019 formulation defines a quantum Wasserstein semimetric by analogy with classical optimal transport over couplings, but with the cost operator

$C := \frac12(I-\mathrm{SWAP}),$

so that

$P$0

subject to

$P$1

The choice of $P$2 via the complement of the symmetric projector ensures $P$3, and the paper proves nonnegativity, symmetry, and identity of indiscernibles, while also stating that the quantity is a semimetric because the triangle inequality fails numerically (Chakrabarti et al., 2019). Its dual form is

$P$4

subject to

$P$5

This dual constraint plays the role that a $P$6-Lipschitz condition plays in classical WGANs, although it is expressed as an operator inequality rather than as a gradient norm bound (Chakrabarti et al., 2019).

The 2021 qWGAN work instead adopts the quantum Earth Mover’s distance of De Palma, Marvian, and Lloyd, with dual form

$P$7

where $P$8 is a quantum Lipschitz constant that measures sensitivity to single-qubit changes (Kiani et al., 2021). This formulation is explicitly designed to reward local progress under shallow or moderately deep local circuits. For computational-basis-diagonal states it recovers the classical Earth Mover’s distance, and for basis states it reduces to Hamming distance: $P$9 The paper further emphasizes properties such as super-additivity and the bound

$\{(p_1,U_1),\dots,(p_r,U_r)\}$0

thereby tying the objective to trace-distance control while preserving locality sensitivity (Kiani et al., 2021).

Because exact optimization over all admissible observables is infeasible, the practical qWGAN restricts the critic to $\{(p_1,U_1),\dots,(p_r,U_r)\}$1-local Pauli operators and replaces the exact Lipschitz constant by a computable surrogate

$\{(p_1,U_1),\dots,(p_r,U_r)\}$2

yielding an approximate distance

$\{(p_1,U_1),\dots,(p_r,U_r)\}$3

The resulting critic step becomes a linear program with one constraint per qubit (Kiani et al., 2021).

The 2023 unseen-state work adopts the same De Palma-type Wasserstein-1 framework and similarly constrains the critic by linear inequalities on Pauli-string coefficients rather than by a gradient penalty (Jurasz et al., 2023). This is a notable point of divergence from the image-generation literature. In the quantum-state line, the critic constraint is implemented through observable-class restrictions and linear-program relaxations (Kiani et al., 2021, Jurasz et al., 2023); in the image-generation line, the critic is classical and the Lipschitz condition is enforced by WGAN-GP (Thomas et al., 2024, Jäger et al., 27 Feb 2026).

3. Architectures in quantum-state qWGANs

In the quantum-data setting, the generator is generally a variational quantum model that outputs pure or mixed states. In (Chakrabarti et al., 2019), the generator is a mixture of parameterized unitaries applied to a fixed initial state, with trainable probabilities $\{(p_1,U_1),\dots,(p_r,U_r)\}$4 and Pauli-rotation angles. The critic is parameterized either as linear combinations of Pauli strings or as parameterized measurement circuits, though the implementation emphasis is on Pauli expansions because they make expectation estimation and regularizer approximation experimentally manageable (Chakrabarti et al., 2019).

In (Kiani et al., 2021), the generator again has a mixed-state form,

$\{(p_1,U_1),\dots,(p_r,U_r)\}$5

but the emphasis shifts to the critic architecture. The critic Hamiltonian is parameterized as a Pauli expansion,

$\{(p_1,U_1),\dots,(p_r,U_r)\}$6

then truncated to $\{(p_1,U_1),\dots,(p_r,U_r)\}$7-local terms to maintain polynomial scaling in $\{(p_1,U_1),\dots,(p_r,U_r)\}$8 for fixed $\{(p_1,U_1),\dots,(p_r,U_r)\}$9 (Kiani et al., 2021). A key practical consequence is critic sparsity: the optimal linear-program solution has at most $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$0 active coefficients, so the critic signal sent back to the generator is a sparse Hamiltonian (Kiani et al., 2021).

The unseen-state extension (Jurasz et al., 2023) preserves this generator–critic structure but inserts a classical stage that maps either a control parameter $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$1 or a latent code to a target vector of observable expectations. In the labeled setting, one interpolates each expectation function $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$2; in the unlabeled setting, one trains a classical WGAN-GP on expectation vectors. The qWGAN then prepares a state whose measured Pauli observables match the supplied target vector (Jurasz et al., 2023). This makes the qWGAN functionally analogous to a quantum decoder constrained by observable statistics.

A concise comparison of the principal quantum-state formulations is useful.

Paper	Data type	Critic mechanism
(Chakrabarti et al., 2019)	Quantum states	Dual SDP with Hermitian observables and entropic regularization
(Kiani et al., 2021)	Quantum states	Sparse $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$3-local Pauli Hamiltonian with LP-constrained quantum Lipschitz surrogate
(Jurasz et al., 2023)	Quantum states at unseen points	Same qWGAN machinery driven by predicted observable expectations

These formulations are not interchangeable. The 2019 semimetric uses a SWAP-based cost and entropic regularization (Chakrabarti et al., 2019), whereas the 2021 and 2023 papers rely on a De Palma-style Wasserstein-1 geometry with locality-sensitive observables (Kiani et al., 2021, Jurasz et al., 2023).

4. Hybrid QWGANs for classical image generation

For classical images, the term “Quantum Wasserstein GAN” usually denotes a hybrid quantum-classical Wasserstein GAN in which the generator is quantum and the critic is classical. In the baseline setting described in (Thomas et al., 2024), the generator is a parameterized quantum circuit $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$4, the critic $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$5 is classical, and the objective is

$P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$6

with the critic assumed to be 1-Lipschitz. The paper explicitly states that it does not use weight clipping. Instead, it follows WGAN-GP: $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$7 This line is motivated by the claim that Wasserstein training supplies a smoother and more informative signal when model and data distributions have little overlap, and that this is especially valuable for capacity-constrained NISQ-compatible generators (Thomas et al., 2024).

The patch-based image generator in (Thomas et al., 2024) is inherited from prior patch-QWGAN work. The full generator is

$P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$8

a concatenation of $P = \sum_{i=1}^r p_i\, U_i \rho_0 U_i^\dagger.$9 quantum sub-generators. Each sub-generator is an $\phi\in H(X)$0-qubit circuit with angle encoding

$\phi\in H(X)$1

followed by repeated variational layers

$\phi\in H(X)$2

where $\phi\in H(X)$3 is a CNOT entangling layer and $\phi\in H(X)$4 is a U3 rotation (Thomas et al., 2024). Output patches are obtained by a nonlinear projective measurement on ancillas, partial trace to a mixed state on data qubits, computational-basis measurement, and normalization to pixel values (Thomas et al., 2024).

The 2026 full-resolution work departs sharply from patching. It proposes a single end-to-end quantum generator aligned to FRQI and MCRQI encodings, with a classical CNN WGAN-GP critic (Jäger et al., 27 Feb 2026). For grayscale FRQI, a $\phi\in H(X)$5 image uses $\phi\in H(X)$6 address qubits plus one color qubit: $\phi\in H(X)$7 with

$\phi\in H(X)$8

Each generator layer contains three stages: noise upload by parameterized $\phi\in H(X)$9 rotations, entanglement among address qubits via structured two-qubit blocks, and controlled $\psi\in H(Y)$0 rotations on the color qubit (Jäger et al., 27 Feb 2026). Address qubits are ordered in Morton (Z) order, and the entangling pattern alternates between nearest-neighbor and next-nearest-neighbor couplings. The paper attributes much of its performance to this task-specific inductive bias rather than to Wasserstein training alone (Jäger et al., 27 Feb 2026).

The critic in (Jäger et al., 27 Feb 2026) is classical and fully convolutional: 3 convolutional layers, kernel size $\psi\in H(Y)$1, stride 2, leaky ReLU activations, and a final linear scalar output, explicitly following the WGAN-GP style of Gulrajani et al. (Jäger et al., 27 Feb 2026). This again underscores that the “quantum” component lies in the generator, not in the critic.

5. Mode collapse, latent geometry, and learned priors

A major theme in the image-generation literature is that Wasserstein training alone does not eliminate mode collapse. “VAE-QWGAN” makes this point explicit by arguing that poor sample diversity in prior patch-QWGANs is driven not only by the adversarial objective but also by the use of uninformed prior distributions such as $\psi\in H(Y)$2 or $\psi\in H(Y)$3 (Thomas et al., 2024).

Its proposed remedy is a hybrid model combining a classical VAE encoder $\psi\in H(Y)$4, a shared quantum decoder/generator $\psi\in H(Y)$5, and a classical critic $\psi\in H(Y)$6 (Thomas et al., 2024). The encoder defines

$\psi\in H(Y)$7

and the VAE decoder is collapsed into the same quantum generator used by the QWGAN: $\psi\in H(Y)$8 The combined objective is

$\psi\in H(Y)$9

with the critical modification that during training the adversarial term samples latent vectors from the encoder posterior $D_{EM}$0, not from a free prior (Thomas et al., 2024).

The paper describes this as supplying two complementary signals to the shared quantum generator: a reconstruction signal from the VAE and a distribution-matching signal from the Wasserstein critic. In practice, the encoder is updated only with the VAE loss, while the shared quantum decoder/generator is updated using a weighted reconstruction term minus the QGAN-GP objective (Thomas et al., 2024). At inference time, because there is no input image $D_{EM}$1, a Gaussian Mixture Model is fit to training latent vectors, and generation proceeds via

$D_{EM}$2

Experimentally, the model uses 50 mixture components (Thomas et al., 2024).

A related but architecturally distinct concern appears in (Jäger et al., 27 Feb 2026), where diversity is improved through a learnable multimodal Gaussian mixture injected layerwise into the FRQI-aligned generator. A shared $D_{EM}$3 is transformed by learned means and variances $D_{EM}$4 for a sampled mode $D_{EM}$5, giving

$D_{EM}$6

Flattened across layers, this yields

$D_{EM}$7

The paper argues that unimodal noise leads to class morphing and mode blending, while tuned multimodal noise provides a better quality–diversity tradeoff (Jäger et al., 27 Feb 2026).

These two papers address the same pathology from different angles. (Thomas et al., 2024) treats latent mismatch as a representational problem and imports structure through a classical VAE and a post hoc GMM prior. (Jäger et al., 27 Feb 2026) treats it as a generative-noise design problem and uses a learnable multimodal latent mechanism embedded directly into the quantum circuit. This suggests that, within hybrid QWGANs for images, latent-space design is as consequential as critic design.

6. Empirical behavior and benchmark regimes

The empirical record of Quantum Wasserstein GANs is heterogeneous because the task domains differ substantially.

For quantum-state learning, the 2019 qWGAN reports smooth fidelity increase for pure states with 1, 2, 4, and 8 qubits, mixed-state learning up to 3 qubits, and resilience to simulated additive Gaussian sampling noise with standard deviations $D_{EM}$8 (Chakrabarti et al., 2019). It also demonstrates a circuit-compression application via the Choi–Jamiołkowski isomorphism: a qWGAN learns a 52-gate circuit that approximates a 3-qubit Hamiltonian-simulation circuit otherwise requiring about $D_{EM}$9 Pauli-rotation gates under the cited Suzuki formula, achieving average output fidelity $k$0 but with worse worst-case error than the product-formula reference (Chakrabarti et al., 2019).

The 2021 qWGAN gives two especially salient trainability results. First, in a GHZ-learning toy model, the EM distance decreases under local progress even when fidelity can remain blind, and the paper proves monotone improvement bounds for intermediate GHZ-construction states (Kiani et al., 2021). Second, in a teacher–student experiment based on a mixing circuit, it reports that overlap-based gradients decay exponentially with qubit number even for shallow constant-depth circuits, whereas qWGAN gradients remain approximately constant with system size in the studied regime (Kiani et al., 2021). The work also notes that deeper target circuits become difficult when low-order Pauli critics are insufficient, making critic locality a central bottleneck (Kiani et al., 2021).

The unseen-state paper reports that, for a topological phase-transition circuit, qWGAN-generated states at unseen values of $k$1 achieve high fidelity and low Wasserstein distance relative to true target states, and reproduce nonlocal string order parameters

$k$2

thereby tracking the phase transition at $k$3 (Jurasz et al., 2023). In the unlabeled setting, it reports that the Wasserstein distance between target expectation vectors and those realized by the generated quantum states drops quickly below $k$4, though it typically plateaus above zero because the outer classical WGAN-GP does not perfectly model the distribution of expectation vectors (Jurasz et al., 2023).

For classical image generation, (Thomas et al., 2024) evaluates on MNIST and Fashion-MNIST using $k$5 grayscale images, but only two classes per dataset: digits 0 and 1 for MNIST, and T-Shirt and Trouser for Fashion-MNIST. It trains on randomly selected 2400 training samples from two classes and reports lower Wasserstein distance during training than baseline PQWGAN with Gaussian or uniform priors. For final generation with GMM inference, it reports the following JSD and NDB results over 5 runs (Thomas et al., 2024):

Dataset / metric	PQWGAN priors	VAE-QWGAN + GMM
MNIST JSD	$k$6 Gaussian; $k$7 Uniform	$k$8
Fashion-MNIST JSD	$k$9 Gaussian; $32\times 32$0 Uniform	$32\times 32$1
MNIST NDB	$32\times 32$2 Gaussian; $32\times 32$3 Uniform	$32\times 32$4
Fashion-MNIST NDB	$32\times 32$5 Gaussian; $32\times 32$6 Uniform	$32\times 32$7

The paper states that lower JSD and NDB indicate better diversity and less collapse, and qualitatively reports that Gaussian-prior PQWGAN produces noisy artifact-ridden samples, uniform-prior PQWGAN remains problematic, and VAE-QWGAN yields clearer reconstructions and better class separation (Thomas et al., 2024).

The 2026 full-resolution study broadens the image benchmark regime substantially. It trains on full 10-class MNIST and Fashion-MNIST, resized to $32\times 32$8, and on color SVHN. Its main grayscale models use 64 layers, 40 modes, and about 50,000 generator updates, achieving reported FIDs of 118 on MNIST and 91 on Fashion-MNIST in the showcased large models, while elsewhere also stating 109 and 70 under different settings or checkpoints (Jäger et al., 27 Feb 2026). For SVHN on the subset where the central digit is 0, it reports a 32-layer QGAN with 3 modes, nearly 100,000 iterations, and FID 84 (Jäger et al., 27 Feb 2026). Against the cited patch-QGAN baseline of Tsang et al., it reports FID 152 versus 207 on an MNIST 3-class subset, and FID 60 versus 179 on a Fashion-MNIST 2-class subset (Jäger et al., 27 Feb 2026).

A plausible implication is that “Quantum Wasserstein GANs” should not be discussed as a single benchmark family. Their empirical performance depends strongly on whether the task is quantum-state learning, image generation from decoded quantum states, or hybrid state preparation from observable targets.

7. Limitations, misconceptions, and research directions

A common misconception is that all Quantum Wasserstein GANs instantiate a single canonical quantum Wasserstein distance. The literature does not support that view. The 2019 qWGAN is built on a quantum Wasserstein semimetric defined through the SWAP operator and entropic regularization (Chakrabarti et al., 2019). The 2021 and 2023 qWGANs use a quantum Earth Mover’s / Wasserstein-1 distance based on neighboring states and a quantum Lipschitz constant (Kiani et al., 2021, Jurasz et al., 2023). The image-generation papers called QWGANs do not define a Wasserstein distance on quantum states at all; they use a classical Wasserstein-1 objective with gradient penalty applied to classical decoded images, while keeping only the generator quantum (Thomas et al., 2024, Jäger et al., 27 Feb 2026).

A second misconception is that Wasserstein training by itself resolves trainability and diversity issues. The evidence is more qualified. In quantum-state learning, the 2021 work attributes improved trainability to the locality-aware geometry of the quantum EM distance, but also identifies the critic approximation $32\times 32$9 as the main limitation, especially when higher-order correlations matter (Kiani et al., 2021). In hybrid image generation, (Thomas et al., 2024) argues that mode collapse persists under Wasserstein training if the latent prior is uninformed, while (Jäger et al., 27 Feb 2026) argues that task-specific inductive bias and multimodal latent design are decisive for scaling to full-resolution images.

A third misconception is that current QWGAN results demonstrate hardware-ready quantum advantage. None of the cited papers make that claim. The 2019 work is motivated by near-term implementation but reports only classical simulations and explicitly notes sampling overhead and approximation complexity for the regularizer (Chakrabarti et al., 2019). The 2026 image-generation work studies finite measurement shot noise and reports encouraging results under a 2048-shot example, but all main training remains in state-vector simulation; the circuits for best-performing models are very deep, with compiled depth estimate 9379 for $C := \frac12(I-\mathrm{SWAP}),$0 grayscale, 64-layer models (Jäger et al., 27 Feb 2026). The 2024 VAE-QWGAN experiments are also conducted in the infinite-shot limit (Thomas et al., 2024).

Several open directions recur across the literature. One is critic design: better relaxations of quantum Wasserstein dual problems, improved operator selection, and richer observable classes remain open in the quantum-state line (Kiani et al., 2021, Jurasz et al., 2023). Another is ansatz design and encoding methods, especially for scalability and generalization to more complex datasets, explicitly identified by (Thomas et al., 2024). A third is measurement efficiency and readout, highlighted in the full-resolution image work, which suggests compressed sensing, Fourier-space measurements via QFT on address qubits, and shadow tomography adapted to structured image states as future directions (Jäger et al., 27 Feb 2026).

Taken together, the literature indicates that Quantum Wasserstein GANs are best understood not as a single model class but as a family of adversarial quantum generative methods unified by Wasserstein-inspired loss geometry. In quantum-state learning, their distinguishing feature is a locality-aware discrepancy between density operators and a critic formulated through observables (Chakrabarti et al., 2019, Kiani et al., 2021, Jurasz et al., 2023). In hybrid image generation, their distinguishing feature is the use of Wasserstein critic training to stabilize optimization of quantum generators acting on classical data domains (Thomas et al., 2024, Jäger et al., 27 Feb 2026). The strongest general conclusion supported by the current record is not merely that Wasserstein objectives are beneficial, but that their value depends on how they are combined with critic constraints, latent modeling, and architecture-specific inductive biases.