"Quantum" has started appearing next to "generative AI" in vendor decks, and marketers are beginning to ask whether the images behind their next campaign will be produced by a quantum computer. It is a fair question, and it has a precise answer rather than a speculative one.
Quantum image generation is not vapour. It exists, it runs on real hardware, and it produces real pictures. The useful thing to know is what those pictures look like, because that single fact settles the planning question for any brand.
What a quantum GAN actually is
A normal generative adversarial network has two halves. A generator invents images, a discriminator judges whether each one is real or fake, and the two train against each other until the generator wins often enough to be useful.
A quantum GAN, or QGAN, replaces the generator with a parameterised quantum circuit. The discriminator stays a classical neural network. That detail matters more than it first appears: current QGANs are hybrid systems, not quantum end to end, and the judging half is ordinary machine learning.
There is a second architectural compromise. Quantum processors available today are small, far too small to hold a whole image, so the image is cut into patches. Each patch gets its own sub-generator circuit, latent noise is fed into each one, the output distribution of each circuit is mapped onto pixels, and the patches are stitched back into a single picture.
What has actually been built
The clearest working example comes from the photonic quantum company Quandela, whose QGAN algorithm generates artificial images from existing image datasets, implemented on their photonic processor and trained end to end. The architecture is exactly the patch-based design above: noise into variational photonic circuits, sub-generator outputs mapped to pixels, recombination into a finished image, classical neural network as discriminator. They have published the implementation openly.
The stated commercial purpose is synthetic training data. Generating artificial images for many scenarios reduces the cost and time of acquiring specialised image datasets and increases diversity in the data used to train other AI models, which matters in sectors like automotive and aerospace where real edge-case footage is expensive or dangerous to collect.
That is a genuine and sensible use case. It is also a completely different job from producing a campaign image, and the difference is where brands get confused.
The output, stated plainly
Published quantum image generation works at benchmark scale. As described by Quandela and ZKSF.org, whose practical overview of quantum machine learning on simulators covers why this training is so expensive, the standard datasets are the ones the whole field uses:
| What is generated | Size | Values per image |
|---|---|---|
| MNIST handwritten digit, the usual QGAN benchmark | 28 x 28 greyscale | 784 |
| Fashion-MNIST clothing item | 28 x 28 greyscale | 784 |
| CIFAR-10 object | 32 x 32 colour | 3,072 |
| One Instagram post | 1080 x 1080 colour | 3,499,200 |
| One 4K portrait | 3840 x 2160 colour | 24,883,200 |
A single social post carries roughly 4,400 times more values than one MNIST digit. A 4K portrait carries about 31,700 times more. That is the honest scale of the distance, and it is not a distance that closes with one hardware generation.
Resolution is not even the hard part
If the only obstacle were image size, the field would simply wait for bigger processors. Two deeper problems say otherwise, and neither is discussed in marketing material.
They learn the average, not the distribution. Published work finds that QGANs struggle to generalise across datasets, converging on merely the average representation of the training data. For a brand, that is fatal in a specific way: an AI influencer campaign needs variety, distinctiveness and a consistent character, and a model that collapses toward the mean of its training set produces the blandest possible output. Generic is the one thing campaign creative cannot be.
There is an analytical ceiling on the quantum half. When the generator's output is a pure quantum state, the achievable discriminator quality is bounded by the fidelity between that state and the target data distribution. This is a mathematical limit on the architecture rather than an engineering shortfall, which means better hardware does not remove it.
And training is brutally expensive. Quantum machine learning needs thousands of circuit evaluations per training run, and variational circuits of this kind run into barren plateaus, regions where the gradient effectively vanishes and the optimiser stops learning. Anyone who has trained a classical model knows what a dead gradient costs. Here it arrives by construction as circuits get wider.
Why the vendor pages do not carry these numbers
In fairness to the vendors, a use case page is a use case page. It exists to describe a capability to a commercial audience, not to publish benchmarks, and the underlying research is usually open for anyone who wants the detail.
The practical consequence still lands on the buyer. If you read a page describing artificial image generation on a quantum processor, nothing on it tells you the output is a 28 by 28 digit. You have to go to the research to find that out, and most people planning a marketing budget reasonably do not.
The rule that protects you is simple: ask what it produced, at what size, measured how. Any claim about quantum capability that cannot answer those three questions is a roadmap, not a result.
What a brand should do about this right now
Nothing, and that is a genuinely useful answer rather than a dismissive one.
- Do not budget for quantum creative. There is no product to buy. Classical diffusion models already produce campaign grade imagery, and the gap between them and quantum generation is measured in four orders of magnitude.
- Do not let it delay a decision. "We are waiting to see what quantum does" is not a content strategy. The timeline is research, not procurement.
- Do treat it as a real field. This is not a scam. Photonic processors really are generating images, and the synthetic-training-data use case is sound. It simply is not a creative production tool.
- Do watch two specific things. Not qubit counts, which say little. Watch whether anyone demonstrates generalisation beyond the training-set average, and whether anyone generates above benchmark resolution. Those are the two walls, and until one falls nothing else matters.
You can check this for yourself, free
The pleasant surprise is that verifying any of this costs nothing to begin. Quantum circuits can be built and run in a browser with nothing to install, so the claims in this article are checkable rather than something you have to take on trust. There is a free online quantum circuit simulator that runs in your browser, which is enough to see how few values a small circuit actually emits and why patching an image across many of them is necessary.
Beyond the free tier, simulation runs at a fraction of a cent per job, and time on real quantum processors is rented by the shot, in the region of $0.30 per task plus fractions of a cent per shot. An afternoon of genuine curiosity costs less than a coffee, which is a better basis for a strategy view than any vendor deck.
The short version
Quantum computers generate images today. They generate handwritten digits at 28 by 28 pixels, in greyscale, using a classical neural network to do the judging and splitting the picture into patches because the processors cannot hold a whole one.
The distance to a photorealistic AI influencer is about four orders of magnitude in raw values, and the obstacles that matter most are not about size at all. They are generalisation, an analytical fidelity ceiling, and gradients that vanish as the circuits grow.
It is a fascinating research field and it is worth watching. It is not a production tool, it will not be one soon, and any campaign planned around it today would be planned around a 28 by 28 greyscale digit.