Photon-native Quantum Generative Models

• Photon-native quantum generative models are architectures that leverage native photonic hardware—Fock-state inputs, linear interferometers, and photon-number measurements—to generate both quantum states and classical data.
• They employ diverse methods including photonic QGANs, variational linear-optical models, and quantum-latent approaches, demonstrating robustness against noise and hardware imperfections.
• Efficient training and deployment are achieved via adversarial minimax games, MMD discrepancy minimization, and boson-sampling protocols, advancing integrated photonics research.

Searching arXiv for papers on photon-native quantum generative models and photonic QGANs. Photon-native quantum generative models are generative-learning architectures whose latent states, parametrized transformations, and sampling procedures are implemented directly in photonic hardware rather than by emulating qubit circuits. In the literature represented by photonic quantum generative adversarial networks, variational Born-style models, and quantum-latent deep generators, the common substrate is linear optics: Fock-state inputs or on-chip entangled-photon sources, meshes of Mach–Zehnder interferometers and phase shifters, and photon-number or click-pattern measurements. These models target both quantum-state generation and classical-data generation, and they are formulated to operate under near-term photonic constraints such as shot noise, phase noise, detector limitations, and fabrication defects (Wang et al., 2023, Sedrakyan et al., 2024, Gottlieb et al., 9 Mar 2026, Bacarreza et al., 27 Aug 2025).

1. Definition and scope

Photon-native quantum generative models exploit the natural building blocks of photonic hardware—Fock-state inputs, linear interferometers, and photon-number measurements—to learn probability distributions over fixed-photon-number bit-strings or to generate quantum states directly (Gottlieb et al., 9 Mar 2026). In this family, a model is not merely “quantum-inspired”; the photonic circuit itself constitutes the generative mechanism, and deployment corresponds to preparing photons, applying a programmable interferometer, and measuring output occupations or other photonic observables.

Several distinct but related formulations have been demonstrated. One line studies Quantum Generative Adversarial Networks in a programmable silicon quantum-photonic chip, where a generator and discriminator are both realized as photonic circuits and trained in a minimax game to reproduce a target two-ququart quantum state (Wang et al., 2023). A second line uses a linear-optical generator with Fock-space encoding and a classical discriminator to generate classical images on a single-photon quantum processor (Sedrakyan et al., 2024). A third line formulates photon-native quantum generative models as variational linear-optical circuits trained classically with a maximum mean discrepancy objective and deployed as boson-sampling devices (Gottlieb et al., 9 Mar 2026). A fourth line uses boson-sampling output distributions as quantum latent priors inside classical GANs, diffusion-style models, and flow-matching systems (Bacarreza et al., 27 Aug 2025).

The term therefore encompasses at least three operational regimes. In a quantum-data regime, the model reproduces a target density operator or pure state (Wang et al., 2023, Ma et al., 2024). In a classical-data regime, the photonic device defines a distribution over Fock outcomes that is mapped into pixel intensities or other classical observables (Sedrakyan et al., 2024). In a hybrid latent regime, the photonic processor supplies a discrete latent distribution $p_Q(z)$, and a classical downstream generator maps those samples into data space (Bacarreza et al., 27 Aug 2025). This suggests that “photon-native” refers less to a single architecture than to a design principle: photonic state preparation, optical interference, and photodetection are used in their native form as the generative substrate.

2. Photonic substrates and circuit constructions

The experimental platforms reported in this area are uniformly based on programmable linear optics, but they differ in encoding and physical realization. In the silicon-photonic QGAN for quantum-state learning, the chip generates on-chip signal–idler photon pairs and post-selects the two-photon maximally entangled state

$$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$

after which signal and idler propagate through two identical universal linear-optical circuits composed of Mach–Zehnder interferometers and thermo-optic phase shifters (Wang et al., 2023). The generator applies two independent single-ququart unitaries in the rectangular arrangement of Clements et al., while the discriminator uses a triangular MZI network implementing arbitrary single-ququart projective measurements (Wang et al., 2023).

For classical-data generation, the photonic QGAN based on Fock-space encoding works in an $m$-mode, $n$-photon Fock space with fixed input

$$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$

and output probabilities

$$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$

Its generator is an $m$-mode linear interferometer built from alternating variational and encoding layers, with classical noise re-uploaded through phase shifters (Sedrakyan et al., 2024). The circuit is mapped to a Clements rectangular mesh of MZIs on the Ascella processor, while the discriminator remains a classical fully connected network (Sedrakyan et al., 2024).

The MMD-trained photonic quantum generative model adopts a closely related linear-optical architecture but emphasizes scalable train-on-classical, deploy-on-quantum operation. The input is an $n$-photon Fock state $\lvert s\rangle\in\Phi_{m,n}$, the trainable interferometer is a universal $m\times m$ unitary $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$0, and the output distribution is

$$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$1

where $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$2 is the second-quantized extension of $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$3 (Gottlieb et al., 9 Mar 2026). Mesh choices include rectangular, triangular, 3-MZI, butterfly, and a Haar-compatible QR parametrization (Gottlieb et al., 9 Mar 2026).

A different photonic realization appears in quantum-latent generative models. There, a Parametric Down-Conversion source produces heralded identical photons, and the PT-2 device uses two sequential optical delay lines coupled to a single spatial channel by an electro-optic modulator. By feeding a train of single photons into these loops, one obtains an effective interferometer on up to $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$4 time-bin modes, with arbitrary $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$5 in a time-bin representation (Bacarreza et al., 27 Aug 2025). In the boson-sampling latent protocol, $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$6 single photons are injected into the first $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$7 channels of an $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$8-mode interferometer and the output photon-count vector $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$9 is measured (Bacarreza et al., 27 Aug 2025).

The most explicitly qubit-oriented photonic implementation is the two-qubit silicon quantum photonic chip with arbitrary controlled-unitary operations and arbitrary two-qubit pure-state generation (Ma et al., 2024). Its source layer uses spontaneous four-wave mixing in silicon spiral waveguides, amplitude-tunable entangled-state preparation via asymmetric Mach–Zehnder interferometers, and path-encoded controlled-unitary operations followed by tomography (Ma et al., 2024). Although this work is qubit-based, it remains photon-native in the sense that all state preparation, unitary evolution, and readout are implemented directly in integrated photonics.

3. Generative objectives and learning paradigms

The principal learning paradigms in photon-native generative modeling are adversarial training, discrepancy minimization, and quantum-latent pushforward generation.

In photonic QGANs for quantum data, the generator and discriminator “arm wrestle” over a measurement difference

$m$0

where $m$1 is the true data state, and overall learning is formulated as

$m$2

Alternating gradient updates maximize $m$3 with respect to the discriminator and minimize it with respect to the generator, with gradients evaluated on-chip via the parameter-shift rule (Wang et al., 2023). Fidelity to the target is quantified by the Uhlmann fidelity

$m$4

For pure $m$5, this reduces to $m$6 (Wang et al., 2023).

In photonic QGANs for classical data, the hybrid objective is the standard GAN loss

$m$7

The discriminator is updated by classical SGD or Adam, whereas the generator is updated by Simultaneous Perturbation Stochastic Approximation rather than exact parameter-shift rules. The SPSA estimate uses only two evaluations per update, independent of parameter dimension (Sedrakyan et al., 2024). No additional quantum divergence or regularizer is introduced beyond the standard adversarial loss (Sedrakyan et al., 2024).

The MMD-based framework replaces the adversarial game with a two-sample discrepancy objective. Maximum Mean Discrepancy is defined as

$m$8

with Gaussian kernel

$m$9

In the no-collision regime, the kernel can be rewritten as a product of linear-optical observables, which permits classical estimation via Gurvits’ permanent algorithm (Gottlieb et al., 9 Mar 2026). One gradient-descent iteration samples data, samples MMD observable weights, uses sign-vectors to compute a Glynn estimator for permanent-related quantities, constructs an unbiased MMD estimate, differentiates it by autodifferentiation or parameter-shift, and updates $n$0 (Gottlieb et al., 9 Mar 2026).

The quantum-latent formulation shifts the generative burden from the entire model to the prior distribution. The boson-sampling latent distribution over occupation vectors $n$1 with $n$2 is

$n$3

This discrete distribution is then supplied to a classical generator $n$4 inside WGAN-GP, DDGAN, or flow-matching pipelines (Bacarreza et al., 27 Aug 2025). The theoretical analysis introduces sampling classes $n$5 and $n$6, proves that if $n$7 and $n$8 is invertible, classically implementable, and both $n$9 and $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$0 are Lipschitz, then the pushforward $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$1, and derives a corollary about data distributions that cannot be approximated by any $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$2 with $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$3 (Bacarreza et al., 27 Aug 2025).

4. Training protocols, initialization, and deployment

Training procedures depend strongly on whether the objective is adversarial or discrepancy-based, but several recurring themes appear: alternating optimization, shot-based estimation, and architecture-sensitive initialization.

In the silicon photonic QGAN for quantum states, initialization uses a randomly chosen maximally entangled two-ququart target state $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$4, identity initialization for the generator unitaries $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$5 and $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$6, and random initialization for discriminator projectors $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$7 and $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$8 (Wang et al., 2023). Training on the real photonic chip uses 150–160 rounds of alternating gradient steps (Wang et al., 2023). In the first five rounds, the measurement difference $$|\psi_{\rm in}\rangle = |n_1^{\rm in}, \dots, n_m^{\rm in}\rangle,$$9 widens on the discriminator’s turn and shrinks on the generator’s turn, giving a direct physical signature of the minimax dynamics (Wang et al., 2023).

In the photonic QGAN for images, one epoch consists of sampling a batch of latent noise and real data, performing one discriminator update, then seven SPSA generator steps, repeated for 1500 total epochs (Sedrakyan et al., 2024). Typical hyperparameters are batch size $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$0, discriminator learning rate $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$1, SPSA gain $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$2, and 10500 SPSA iterations overall (Sedrakyan et al., 2024). Convergence is identified when $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$3 and $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$4 stabilize and oscillate around $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$5, while generated images acquire human-recognizable features (Sedrakyan et al., 2024).

The MMD-based photonic QGM introduces a train-on-classical, deploy-on-quantum workflow. Each permanent estimate via Gurvits costs $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$6 for error $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$7, and overall cost per batch scales as

$$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$8

With modest $$p(s)=\bigl|\langle s|\,U_g(\boldsymbol\theta)\,|\psi_{\rm in}\rangle\bigr|^2.$$9, $m$0, and automatic differentiation, training up to $m$1 parameters is efficient on a classical laptop (Gottlieb et al., 9 Mar 2026). Once $m$2 is trained, deployment is simply preparing $m$3, applying the photonic circuit with phases $m$4, and measuring in the Fock basis; this is precisely an $m$5-photon boson-sampling experiment (Gottlieb et al., 9 Mar 2026).

Initialization and ansatz selection are treated as substantive design variables in the MMD framework. Identity-near initialization reduces variance, warm-starts based on progressively smaller kernel bandwidths marginally improve final loss, and butterfly and Haar-compatible ansätze showed best expressivity and trainability (Gottlieb et al., 9 Mar 2026). Input-state choice—such as populating the first $m$6 modes or using random placement—also functions as a hyperparameter (Gottlieb et al., 9 Mar 2026).

In quantum-latent models, deployment is separable: one first samples $m$7 from a boson-sampling device or simulator and then feeds those latents into a classical generator (Bacarreza et al., 27 Aug 2025). The paper recommends moderate latent dimensions $m$8, approximately invertible and Lipschitz generators, and circuit randomization across seeds to avoid over-fitting to one unitary (Bacarreza et al., 27 Aug 2025). This suggests that, in the latent setting, photonic hardware can be treated as a reusable stochastic primitive rather than an end-to-end trainable model.

5. Empirical demonstrations and performance

The experimental record for photon-native quantum generative models spans quantum-state learning, classical distribution loading, image generation, and latent-space enhancement.

For quantum-data generation in a programmable silicon photonic chip, the first photonic experimental QGAN demonstrated that after 150 rounds on a noisy chip with $m$9 and total two-photon count $n$0, the measurement difference converged to $n$1 and the fidelity climbed from $n$2 to $n$3 (Wang et al., 2023). In a defective-chip experiment with 16 out of 30 generator phase shifters disabled, the model converged after $n$4 rounds with final fidelity $n$5 and $n$6 (Wang et al., 2023). The same study reports from numerical sweeps that for phase noise $n$7, the average end-of-training fidelity exceeds $n$8, while at $n$9 it falls below $\lvert s\rangle\in\Phi_{m,n}$0; likewise, even with up to $\lvert s\rangle\in\Phi_{m,n}$1 of the generator’s phase shifters broken, the expected fidelity remains $\lvert s\rangle\in\Phi_{m,n}$2 (Wang et al., 2023).

The two-qubit silicon-photonic GAN with maximum expressibility reports high-fidelity state learning in a different architecture. For a pure target $\lvert s\rangle\in\Phi_{m,n}$3, the final fidelity is $\lvert s\rangle\in\Phi_{m,n}$4, and for a mixed target $\lvert s\rangle\in\Phi_{m,n}$5, the final fidelity is $\lvert s\rangle\in\Phi_{m,n}$6 (Ma et al., 2024). In hybrid quantum-classical GAN mode, the same work loads classical normal, log-normal, and bimodal distributions with final average $\lvert s\rangle\in\Phi_{m,n}$7 over five trials, and reports compressed MNIST digit generation with critic loss tending to zero and KLD tending to zero in approximately 200 epochs (Ma et al., 2024).

For classical image generation on a single-photon processor, the photonic QGAN based on Fock-space encoding uses Structural Similarity Index as a quality and diversity measure. On $\lvert s\rangle\in\Phi_{m,n}$8 real versus generated images, the average similarity score is approximately $\lvert s\rangle\in\Phi_{m,n}$9, and the diversity score $m\times m$0 is approximately $m\times m$1 (Sedrakyan et al., 2024). Compared to a small classical GAN with approximately 332 parameters, the photonic QGAN with approximately 208 parameters reaches comparable similarity but slightly higher diversity (Sedrakyan et al., 2024). Human-recognizable contours for digits emerge by iteration 1500 in ideal simulation and around 1000 in experiment (Sedrakyan et al., 2024).

In the MMD-trained photonic QGM, convergence is reported to be smooth, with larger estimator batches reducing noise (Gottlieb et al., 9 Mar 2026). On a boson-sampling dataset with $m\times m$2, the QCBM outperforms a classical RBM and a uniform fixed-weight baseline, closely matching “test-to-test” MMD (Gottlieb et al., 9 Mar 2026). On real-world datasets—sushi preferences, movie preferences, and a bioinformatics CMap task—the QCBM matches or marginally underperforms the RBM but remains competitive (Gottlieb et al., 9 Mar 2026). A 2000-step training run on $m\times m$3 takes approximately 2 hours on a laptop, corresponding to 3–4 seconds per step (Gottlieb et al., 9 Mar 2026).

The quantum-latent program reports gains from boson-sampling priors in downstream deep generators. On a synthetic quantum dataset generated from 8 photons in a 16-mode Haar-random interferometer, the boson sampler achieves the best score on the paper’s discrete-structure metric, with $m\times m$4, while classical latents perform worse (Bacarreza et al., 27 Aug 2025). On the QM9 molecular dataset, using FCD as the main metric, the boson sampler at $m\times m$5 achieves $m\times m$6, compared with $m\times m$7 for distinguishable-photon sampling, $m\times m$8 for Gaussian latents, and $m\times m$9 for Bernoulli latents; similar ranking is reported at dimensions 32 and 48, and PT-2 hardware matches simulation (Bacarreza et al., 27 Aug 2025).

6. Noise, defects, scalability, and open problems

A central theme in photon-native generative modeling is robustness under realistic photonic imperfections. The photonic QGAN for quantum data models phase-shifter noise by replacing each ideal phase $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$00 with $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$01, models detection shot noise by replacing ideal coincidences $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$02 with Poisson-distributed $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$03, and models defects as stuck thermo-optic phases fixed at unknown values (Wang et al., 2023). The reported robustness to phase noise up to $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$04 and to severe damage affecting up to half of the generator’s phase shifters is therefore not an abstract noise tolerance statement but one tied to explicit device-level perturbations (Wang et al., 2023).

For deployable photonic QGMs, photon loss and circuit depth become major constraints. Loss reduces photon-number conservation and biases output distributions, while deep interferometers accumulate phase errors (Gottlieb et al., 9 Mar 2026). The work notes that loss mitigation via postselection or classical postprocessing and zero-noise extrapolation can help but increase sampling cost, and that alternative fault-tolerant meshes such as fractal or recurrent MZIs improve resilience (Gottlieb et al., 9 Mar 2026). It also identifies the no-collision regime as the tractable classical-simulation setting and states that handling higher occupations requires new classical algorithms (Gottlieb et al., 9 Mar 2026).

The image-generation QGAN faces a related but distinct bottleneck: photon loss and sampling overheads constrain the attainable spatial dimension, even though patch-based and distribution-based mappings let moderate $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$05 values generate $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$06 images (Sedrakyan et al., 2024). The use of threshold detectors instead of ideal photon-number-resolving detectors further necessitates postselection and binning rules (Sedrakyan et al., 2024).

The quantum-latent literature frames scalability partly as a hardware advantage in sampling throughput. Classical boson-sampling simulation reaches approximately $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$07 only at significant cost, whereas a photonic device can stream more than $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$08 samples in minutes (Bacarreza et al., 27 Aug 2025). At the same time, the paper notes that classical approximations can match low-order statistics but incur quadratic or higher overheads and cannot efficiently produce i.i.d. samples at scale (Bacarreza et al., 27 Aug 2025). A plausible implication is that photon-native latent samplers may become attractive before fully trainable end-to-end photonic generators do, because the hardware burden is concentrated in repeated sampling rather than iterative differentiation.

Several future directions recur across the literature. As photonic integration scales up, cross-talk and thermal management become more complex, requiring improved phase-shifter designs and active calibration (Wang et al., 2023). Richer ansätze with more photons, qudits, or entangled inputs will increase parameter-space dimensionality and may require natural gradients or adaptive learning rates (Wang et al., 2023). Proposed extensions include exact photonic parameter-shift rules, analytic gradient formulas, superposition and entangled photon inputs, programmable photonic processors with error-mitigation protocols, conditional QGANs, fully quantum discriminators, and multi-photon or continuous-variable encodings (Gottlieb et al., 9 Mar 2026, Sedrakyan et al., 2024, Bacarreza et al., 27 Aug 2025). These are not yet standardized methods; they remain active development directions.

7. Conceptual significance and relation to neighboring paradigms

Photon-native quantum generative models occupy a distinct position relative to qubit-based QGANs, classical deep generators, and generic “quantum-inspired” probabilistic models. Their defining claim is not merely that a quantum circuit can replace a classical layer, but that photonic primitives naturally implement distributions and transformations relevant to generative learning. In the MMD-based formulation, deployment is literally boson sampling (Gottlieb et al., 9 Mar 2026). In the latent-prior formulation, the quantum advantage argument is tied to pushforwards of distributions in the class $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$09 through approximately invertible Lipschitz generators (Bacarreza et al., 27 Aug 2025). In photonic adversarial learning, near-term integrated hardware performs the optimization loop itself, including parameter-shift-based gradient evaluation or SPSA-based shot-efficient updates (Wang et al., 2023, Sedrakyan et al., 2024).

A common misconception is that all photon-native generative models are adversarial. The literature shows otherwise: adversarial minimax training is one family, but MMD minimization and quantum-latent priors constitute separate paradigms with different computational trade-offs (Gottlieb et al., 9 Mar 2026, Bacarreza et al., 27 Aug 2025). Another misconception is that photon-native models are restricted to quantum outputs. In fact, the reported applications include learning two-ququart target states, loading classical one-dimensional distributions, generating compressed MNIST images, modeling preference datasets, bioinformatics signatures, and improving molecular graph generation on QM9 (Wang et al., 2023, Ma et al., 2024, Sedrakyan et al., 2024, Gottlieb et al., 9 Mar 2026, Bacarreza et al., 27 Aug 2025).

Taken together, the published results indicate a field that has progressed from proof-of-principle photonic QGANs toward a broader taxonomy of photon-native generative methods. The experimentally demonstrated ability to attain fidelities above $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$10 for quantum-state generation under realistic defects (Wang et al., 2023), the classical-laptop trainability of MMD-based linear-optical generators up to $$|\psi_0\rangle = \tfrac14 \sum_{k=1}^4 |k\rangle_s \otimes |k\rangle_i,$$11 parameters (Gottlieb et al., 9 Mar 2026), and the empirical improvement of boson-sampling latents over Gaussian and Bernoulli baselines on QM9 (Bacarreza et al., 27 Aug 2025) show that photon-native generative modeling is no longer a single algorithmic proposal. It is a developing research area organized around a common physical principle: photonic interference and measurement are treated not as implementation details, but as the native source of expressive generative structure.