Can real quantum hardware noise improve generative AI diffusion models?

Two protocols, one shared forward process

The study asks whether the noise-injection stage of a diffusion model can be replaced or supplemented by quantum dynamics, and whether real hardware noise can drive image generation. Diffusion models normally corrupt training samples through a classical Markov chain until they become pure noise, then train a neural network to reverse that corruption. Here the forward chain is the part being changed; the backward denoising network stays classical in both protocols.

The first protocol uses quantum stochastic walks, a formalism that interpolates between a purely quantum walk and a classical random walk through a single parameter, omega. Setting omega to 0 gives coherent quantum dynamics, omega equal to 1 gives classical dynamics, and intermediate values mix the two. The second protocol implements a quantum walk on a cycle graph in a quantum circuit and deliberately lets the device's intrinsic noise, increased by delay operations, act during the forward process.

This split matters for reading the results. The hybrid experiment is a controlled comparison of dynamics within a simulation-like setting, while the hardware experiment is a proof of concept that the pipeline runs on a real noisy processor. They answer different questions and should not be treated as one benchmark.

A mixed quantum-classical forward chain lowered FID

In the first protocol, the authors trained unconditional diffusion models on MNIST digit images and compared forward chains with different values of omega. The headline result is that a hybrid setting, omega equal to 0.3, generated image sets with lower Fréchet Inception Distance than a fully classical forward chain. FID compares the distribution of generated images with that of real images, so a lower value means the generated set is statistically closer to the real set.

The paper also reports that the hybrid model was statistically more robust across repetitions than the classical one. The comparison is between sets of generated images, not between individual samples, and the metric is computed on the whole generated dataset against the training data. The authors illustrate the difference with nine samples per setting and with pixel-value distributions, but the quantitative claim rests on FID over repeated runs.

The authors' explanation is mechanistic rather than proven: a balance between quantum coherence and stochastic noise is said to allow more efficient information transport across the graph, which yields better diffusion over the graph and a more complete exploration of the state space. They note that the eight-node cycle graph used here has diameter 4, above the threshold of 3 at which prior work reports efficient transport around omega near 0.1, and treat the observed optimum at 0.3 as consistent with that framework. This is an interpretation of the result, not a separately measured causal chain.

The graph turns pixel values into walker positions

To understand why a walk on a graph appears in an image model, it helps to see what the graph represents. Each pixel is encoded as the position of a walker on a cycle graph with eight nodes, so the eight possible positions correspond to eight gray levels. The forward process moves walkers around this graph, and the resulting distribution over positions defines the categorical distribution from which the next pixel value is drawn.

This encoding is what makes the quantum walk relevant to image generation at all. The walk is not operating on the image directly; it is operating on the discrete color values that each pixel can take. The cycle graph's rotation-invariant property lets the authors run each walker as if it started from the same initial state and then remap the outcome to the specific pixel value with a shift operation. That is the trick that lets the forward chain be run once for all walkers rather than separately per pixel.

The graph also sets a clear boundary. With eight nodes, the model covers eight gray levels, and the paper's MNIST images are normalized to that range. The number of qubits is tied to the number of gray levels, not to image dimensions, which is the basis for the authors' scalability claim.

Four qubits generated full-size MNIST images

The hardware protocol used only four qubits to generate 28 by 28 grayscale MNIST images: three qubits for the walker's position on the eight-node graph and one for the coin degree of freedom. The forward process was implemented in Qiskit, first simulated on a fake version of the processor with 100,000 shots, then run on the real IBM device with 10,000 shots. The shot count was reduced for the real device because of the computational resources available.

Noise was injected deliberately through delay operations whose length follows a schedule analogous to the cosine noise schedule used in classical diffusion models. The authors chose a coefficient so that the forward process converges to a uniform distribution within 20 time steps. The backward process was a classical multilayer perceptron, the same style of network used in the hybrid experiment.

The model was trained on 6,903 full-size MNIST images of the digit 0 as a proof of concept. The authors emphasize that the protocol suits the processor's connectivity, which has a maximum degree of 3, and that several walkers can run in parallel on different parts of the same processor. These are feasibility arguments, not evidence of image quality.

Real hardware noise trailed the hybrid simulation

The most decision-relevant comparison in the paper is between the hybrid quantum stochastic walk model and the real-hardware quantum walk model. The hybrid setting reached an FID of 114, while the IBM-hardware run reached 169, both computed between the original dataset and an equally sized generated dataset. The authors describe the hardware result as generating samples with higher FID and offer a possible explanation for the gap, but they do not establish a single cause.

This comparison is easy to misread. A lower FID is better, so 114 is the stronger result and 169 is the weaker one. The hardware run is not being presented as an improvement over the hybrid model; it is being presented as evidence that a real noisy processor can be used at all, with noise treated as part of the generative mechanism rather than an error to be corrected.

The paper also reports results on other digits and on the full MNIST set in its appendix. On the full set, a hybrid setting with omega equal to 0.4 produced better FID than classical dynamics, and the same general trend appeared on EMNIST letters, though the hybrid advantage was slightly less prominent there. Those runs used 100 training epochs instead of the 10,000 used in the digit-0 comparison, which the authors note results in higher FID values overall.

Noise as a resource, not yet a scaling claim

The paper's framing is that quantum noise need not be mitigated or corrected and can instead be exploited as a useful resource. That framing is supported by the fact that the hardware pipeline produced images and by the authors' argument that the forward protocol fits current noisy intermediate-scale quantum devices. It is not supported as a claim that real hardware noise outperforms classical or hybrid alternatives, because the reported hardware FID is higher than the hybrid FID.

Several boundaries follow directly from the design. The experiments use MNIST digits, with the hardware proof of concept limited to the digit 0 and 6,903 training images. The images are grayscale with eight levels, set by the eight-node graph. The backward network is classical in every configuration, so the results concern the forward corruption process only. The authors also note that they have evidence of the importance of noise characterization from comparing real and simulated processors, and they list error correction or mitigation as a future direction for better controlling noise levels.

The authors position the work as a starting point for large-scale quantum generative AI algorithms rather than a demonstration of one. Their own future directions include integrating processor topology more deeply with the circuit-level walk to widen the range of pixel values, and eventually realizing the quantum walk physically so the model could apply to quantum data with a quantum network in the reverse process.

The hybrid setting is the result to watch

If the question is whether real quantum hardware noise can improve generative diffusion models, the paper's answer is conditional. Real noise was used to generate images on a four-qubit processor, which supports the resource framing, but the real-hardware run scored FID 169 against 114 for the hybrid quantum stochastic walk model. The improvement claim over classical dynamics belongs to the hybrid setting at omega equal to 0.3, not to the hardware run.

The practical priority the paper establishes is therefore the intermediate mixing regime, not the fully quantum or fully classical extremes. The authors' transport argument explains why an intermediate omega could help, and the appendix trend at omega equal to 0.4 on the full MNIST set points in the same direction. Readers evaluating follow-up work should check which forward dynamics produced a given FID, on which dataset and digit subset, and whether the backward network was classical, because those conditions separate the paper's stronger hybrid result from its feasibility demonstration on real hardware.

Sources used in this answer