Researchers have demonstrated a way to generate images using quantum computing with a dramatic reduction in required resources. The work uses the inherent noise produced during quantum measurement, typically seen as a hindrance, to replace randomness deliberately added during the image generation process.

On an eight-dimensional Gaussian mixture, the new method achieved comparable image quality with just 303 measurement shots per sample, a reduction of approximately 5300x compared to a reference requiring 1.6 x $10^{6}$ shots. This approach, detailed in a recent paper, formulates diffusion processes on a hyperplane to allow for adaptive noise compensation during sampling.

Quantum Measurement Noise as Diffusion Sampler Resource

The work demonstrates a 5300 × 5300 × reduction in computational resources needed for image generation on an eight-dimensional Gaussian mixture, achieved by repurposing a traditionally unwanted element of quantum computing: measurement noise. This counterintuitive approach fundamentally shifts how measurement noise is viewed, from a liability to an asset in quantum neural networks. The team formulated the diffusion process on a bounded zero-sum hyperplane mapped to the probability simplex, a mathematical framework that allows them to connect the probabilities generated by the quantum circuit to the task of denoising images.

This formulation enables a “compensation noise” that dynamically adjusts to the level of determinism within the sampler itself, effectively redistributing the need for randomness. By accounting for the covariance of the measurement-induced perturbation in the reverse update step, the method reduces the amount of externally injected diffusion noise, creating a more efficient process.

The researchers explain that under-sampling is no longer simply an approximation of exact quantum readout, but a mechanism for strategically allocating randomness between quantum measurement and classical noise injection. The core contribution of this work, according to the paper, is a quantum computing mechanism for diffusion model applications that establishes the conditions under which measurement noise can be effectively “spent” to reduce the need for additional noise.

Conventional sampling methods rely on a large number of shots to minimize measurement noise before adding the full prescribed diffusion noise. This new sampler uses fewer shots and correspondingly reduces the added noise, acknowledging the fluctuations already present in the measurement. The process begins with random noise, with denoising serving as the prediction at each step and complete reversal of that process resulting in image generation.

Experiments using the MNIST dataset in a learned latent space further validate the proposed mechanism for generation, demonstrating its applicability beyond simple Gaussian mixtures. Table III in the paper details the per-sample resources for various quantum configurations, highlighting the substantial reduction in shots per sample achieved with the adaptive under-sampling method. 6 × 10^6. “We have shown that finite-shot quantum measurement noise can supply part of the randomness a stochastic diffusion sampler would otherwise inject,” the paper concludes, suggesting a pathway toward more efficient quantum image generation.

Multinomial Sampling Noise in Trainable Readouts

The formulation of diffusion processes on a bounded zero-sum hyperplane allows for a dynamic adjustment of noise compensation, matching the level of determinism within the quantum sampler. This approach directly addresses the inherent trade-off between computational cost and image quality in quantum-enhanced generative modeling, a challenge previously requiring extensive measurement repetitions to mitigate noise. By characterizing measurement error as multinomial sampling noise propagated through a trainable linear readout, the research establishes a framework for actively using, rather than suppressing, the randomness of quantum measurement.

The work moves beyond simply approximating exact quantum readout with under-sampling; it redefines the role of finite-shot noise as an integral component of the generative process. This alignment is achieved by explicitly propagating the covariance of the measurement-induced perturbation through the linear readout, enabling a precise “compensation noise” that balances the stochasticity of the reverse diffusion process.

The method’s effectiveness hinges on satisfying a key condition, that the covariance of the sampling noise, after being reshaped by the readout, does not exceed the covariance of the noise prescribed by the reverse step. The researchers demonstrate that a state-dependent shot budget is derived by solving the condition for matching noise covariance, meaning fewer measurements are needed as the sampler becomes more confident in its output.

The team’s method uses the noise prescribed by a stochastic sampler as a budget for finite-shot measurement error, a strategy that builds on recent work combining quantum circuits with denoising diffusion probabilistic models. The mean and covariance of the noise are calculated exactly, while normality is approximate, ensuring a robust and reliable framework for noise compensation.

Diffusion Process Formulation on a Zero-Sum Hyperplane

The researchers detail how this approach addresses the inherent tension between the need for repeated measurements to suppress quantum noise and the desire to minimize computational cost. The team’s method tackles the challenge of finite-shot sampling noise, traditionally viewed as a limitation in quantum computing, by repurposing it as a resource.

The method’s success depends on satisfying a positive-semidefinite condition, ensuring the total conditional noise covariance matches that of the ideal, noise-free reverse step, under the assumption of Gaussian readout error. The encoding used operates on the probability simplex, bridging the gap between practical data and the quantum circuit by placing the data on the bounded zero-sum hyperplane.

The team’s work extends beyond theoretical formulation, detailing a specific implementation using a denoiser circuit operating on three qubits. Latent diffusion experiments on MNIST further validate the proposed mechanism for generation in a learned latent space.

Shot-Efficient Diffusion with Adaptive Measurement Budgets

The work details a method for determining an adaptive shot budget during the image generation process, increasing as the sampler becomes more deterministic. This contrasts with conventional quantum diffusion models that require a fixed, often high, number of shots to suppress measurement noise before injecting additional randomness for each step of the reverse diffusion process. Instead, the researchers demonstrate a way to use the inherent noise from quantum measurement as a resource, effectively reducing the amount of noise the sampler needs to add independently.

This allows the circuit to use the injected-noise covariance of the stochastic diffusion sampler as a budget for finite-shot readout noise, a key step toward reducing computational cost. Validation of this mechanism came through latent diffusion experiments performed on the MNIST dataset, confirming the proposed approach for generation within a learned latent space.

The team’s hybrid encoding, designed with specific constraints, further facilitates this process by connecting the circuit’s Born probabilities to the denoising task. The researchers emphasize that the shot count is not a hyperparameter to be tuned, but rather an output of the condition established by balancing measurement noise and injected diffusion noise.

This adaptive allocation of shots represents a shift in how quantum diffusion models are designed and implemented, potentially enabling more scalable quantum image generation. The method’s efficacy stems from turning a traditionally viewed liability, measurement noise, into an asset that actively contributes to the generative process, streamlining the overall computation.

Latent Diffusion Validation on MNIST Dataset

Latent diffusion experiments using a learned latent space confirmed the viability of this noise-replacement strategy with the MNIST dataset. The team achieved a minimum mean squared distance (MMD^2) of 0.0143 in image space, a metric of generation quality, using only 993 shots per sample. The approach successfully tied the quantum resource requirement to the latent dimension rather than the image resolution, a key step toward scalability.

The team’s work suggests that even deterministic diffusion sampling can produce useful samples, though it requires careful consideration of the covariance structure of the quantum measurement estimator. This adaptive approach represents a step toward more efficient quantum image generation, potentially unlocking broader applications for quantum neural networks.

👉 More information
🗞 Leveraging Quantum Measurement Randomness for Shot-Efficient Diffusion Models
✍️ Tianhao Guo, Dianxin Zheng, Yuhua Qian, Yaohui Zheng, Yuzhi Yang, Merouane Debbah and Chau Yuen
🧠 ArXiv: https://arxiv.org/abs/2610.41721

More like this