This Person Does Not Exist

This Person Does Not Exist is best understood today as a historical web demonstration of synthetic face generation, not as a currently verified face-generation service. On September 20, 2026, its home page returned HTML containing a domain-for-sale offer, while the old /image path returned 404. The page still returned a changing background image, so those checks do not prove that every historical asset or URL has disappeared.

What the name refers to

The name refers to a website that became known in 2019 for showing a synthetic portrait when a visitor loaded or refreshed the page. A 2019 Design Museum exhibition record names Phil Wang as the creator. We did not locate a preserved first-party statement from Wang during this review, so his employer, exact role, and launch date remain secondary-source history rather than facts verified from the site's current operator.

The more important identity distinction is technical. This Person Does Not Exist was a website and demonstration, not a model family developed by NVIDIA. Contemporary accounts associated it with NVIDIA's original StyleGAN research, but the primary NVIDIA sources do not document the website's production stack. Philip Wang should not be described as an NVIDIA researcher. That distinction matters when a reader tries to verify the code, training data, licence, or current availability. Each belongs to a different entity.

The current domain cannot support the old product description. Its HTML on September 20, 2026 contained a domain-sale message and referenced random-person.jpeg as a background. The historical /image route returned an ordinary 404 page. Three requests for the background returned different byte hashes, which shows that the responses changed. It does not reveal whether the server generated an image on demand, selected one from a collection, used a particular model, or granted any reuse right.

For a reader arriving from search, the practical verdict is therefore narrow. The historical demonstration remains useful as a case study in how synthetic faces entered public awareness. It is not a suitable starting point for a new production workflow unless the current operator, generation flow, and terms can be identified again.

How the historical demonstration worked

NVIDIA's official StyleGAN repository documents a pretrained generator for Flickr-Faces-HQ, usually shortened to FFHQ, at 1024 by 1024 pixels. A generator maps a sampled latent vector into an image. StyleGAN's contribution was a redesigned generator that allowed learned image attributes and fine stochastic details to enter the synthesis process at different scales. The associated 2018 research paper discusses separation of high-level attributes such as pose and identity from stochastic variation such as freckles and hair.

That account explains the historical demonstration without making claims about its present server. A visitor did not upload a portrait, specify an age, or enter a prompt in the familiar interface. The original experience was closer to sampling: load the page, receive a face image, and reload for another sample. It offered immediacy, but very little control. A random face generator and an identity-preserving portrait editor solve different jobs, even if both display a face.

The output was an image, not a reusable model checkpoint or an editable identity. The public interaction did not document a seed, prompt, dataset row, or provenance record for each result. Without those details, an image cannot be reproduced reliably or audited back to a generation event. That weak provenance is acceptable for a technology demonstration, but it is a serious limitation for a design system, public campaign, research corpus, or regulated workflow.

The current site check was not a hands-on test of the historical generator. We did not observe a working control that promised a new generated face, inspect server code, or obtain an output under dated terms. Any present-tense statement about generation speed, GPU inference, randomness, or uniqueness would go beyond the evidence.

StyleGAN is not StyleGAN2

The research timeline establishes an important version boundary. NVIDIA first submitted the original StyleGAN paper in December 2018 and the distinct StyleGAN2 paper in December 2019. These dates show that StyleGAN and StyleGAN2 are separate research releases. They do not, by themselves, prove which revision ran on the historical website or whether its implementation ever changed.

The current website provides no implementation history that resolves that question. Describing the historical service as StyleGAN2 therefore requires site-specific evidence, not a date-based inference. The same applies to a StyleGAN2 FID score of 2.84. A score from a research paper is measured under that paper's dataset, implementation, sample, and metric conditions. It is not a benchmark for an independent website merely because both relate to synthetic faces.

The original StyleGAN repository reports its own metrics for its pretrained FFHQ network, but even those should not be turned into a service score. The website may have used different weights, truncation settings, sampling code, image processing, or infrastructure. None of those conditions is available from the current domain. Treat the architecture as historical context and do not assign laboratory numbers to a service that was not measured under the same conditions.

This version discipline also changes how readers evaluate tutorials. A guide that says "StyleGAN2" may be describing a later local implementation, not the historical site. Before following it, check the repository, checkpoint, licence, environment, and date. Similar names do not make two workflows interchangeable.

What the FFHQ source establishes

The official FFHQ repository says the dataset contains 70,000 aligned PNG images at 1024 by 1024 resolution. It includes variation in age, ethnicity, background, and accessories. The images originated on Flickr, so NVIDIA also warns that the dataset inherits biases from that source. This evidence supports a description of the training dataset used in the StyleGAN paper. It does not prove balanced demographic performance or "unlimited diversity" in generated results.

The repository draws another boundary that is easy to miss. FFHQ was created as a GAN benchmark and is not intended for developing or improving facial-recognition technology. A synthetic portrait page should not be promoted as a privacy-safe research dataset without defining the task, risks, licence, and review process. Synthetic data can preserve patterns and biases from its training source. It can also create false confidence if researchers assume that no direct photograph of a participant means no privacy or ethics question remains.

Licensing is layered. Individual Flickr images in FFHQ carry several possible licences, with metadata identifying the author and terms. The dataset package is released under CC BY-NC-SA 4.0. NVIDIA's original StyleGAN materials, excluding FFHQ, are released under CC BY-NC 4.0 according to the official repository. Both named licences contain non-commercial restrictions in this context. They should not be summarized as unrestricted open source.

None of those licences automatically answers whether a JPEG once served by This Person Does Not Exist can be reused in an advertisement, interface, game, or training set. The code licence governs code and related materials. The dataset licence governs the dataset package. An output served by an independently operated website needs terms tied to that service and output. The current site did not expose those terms during this review.

Four rights layers you must separate

Before adopting a synthetic face, identify four distinct objects: the service, the generator code or weights, the training dataset, and the exact output. A permissive answer for one object cannot be copied to the others.

The service layer answers who operates the site, what account or usage agreement applies, and whether the service is still available. The code and weights layer answers whether you may run, modify, or redistribute the implementation. The dataset layer answers how source images were licensed and what purposes the dataset publisher excludes. The output layer answers what you may do with the image you actually received. This final layer may also involve privacy, publicity, impersonation, and platform-policy concerns that a software licence does not settle.

For This Person Does Not Exist, the first and fourth layers are presently unresolved. The domain's sale message makes operator continuity uncertain. No current output licence was found. That uncertainty blocks a recommendation to use the images commercially, even though the architecture and dataset have public repositories.

A designer may be tempted to call a generated face a privacy-preserving placeholder because no consenting model sat for that exact portrait. That description promises too much. The generator learned from photographs of real people, and a generated face is not guaranteed to avoid resemblance to a real person. More importantly, a placeholder can become deceptive when it is placed beside a testimonial, employee name, customer story, medical case, or profile. Context determines the representation a reader will infer.

The safest practical rule is simple: if the exact output has no documented provenance and rights, do not treat it as a reusable stock asset. Use an asset source that records the licence, creation method, and intended public use, or create and review an output under terms you can archive.

A five-gate image-use decision

This source-based worksheet turns the uncertainty into a reproducible decision. It is not legal advice, and completing it does not replace counsel for a high-risk use.

Consider a concrete but hypothetical design task: a team needs eight avatars for an internal wireframe that will never leave a controlled workshop. The team still records where each image came from, keeps it out of public claims, and replaces it before release. If the same faces are later moved into a public customer directory, the decision changes. The representation, audience, and rights requirement changed even though the pixels did not.

The worksheet also prevents a common shortcut. "AI-generated" is a method label, not a licence and not proof that no person can be harmed. A defensible yes decision needs evidence at every relevant gate. One unresolved gate is enough to choose a different asset.

Failure modes before image reuse

The first failure mode is stale identity. A familiar domain can change owner or purpose while old articles continue to describe the original service. Check the rendered page and important endpoints close to the decision date. Record what they returned, not merely the HTTP status. A 200 response can be a sale page, login screen, or unrelated replacement.

The second is architecture drift. If a page calls the service StyleGAN2 because StyleGAN2 later became better known, compare publication dates and require migration evidence. Correcting the label matters because it affects benchmarks, code examples, licences, and expected artifacts.

The third is rights inference. Do not reason from public code to commercial output permission. Locate the output clause, confirm which operator issued it, and save the terms and date connected to the generation event. If the clause cannot be found, the observable result is not "probably permitted." It is an unresolved right, so the workflow stops.

The fourth is false identity assurance. A random sample does not prove uniqueness or zero resemblance. Never use a synthetic face to imply a real employee, customer, voter, patient, witness, or reviewer. If a public design needs a fictional character, label the context appropriately and conduct a risk review that matches its reach.

The fifth is unsupported detection advice. Earrings, hair, backgrounds, and reflections were useful clues for some older GAN outputs, but artifacts change across models. A checklist of visual quirks cannot authenticate an image reliably. If authenticity matters, use provenance, source records, trusted capture context, and appropriate forensic review rather than promising that a viewer can spot every synthetic face.

When another workflow makes more sense

Choose a rights-documented stock or commissioned portrait when the image represents a person in advertising, editorial work, or a product interface. Choose clearly fictional illustration when realism is not necessary. Choose a current synthetic-media service only after its operator, terms, output rights, controls, and disclosure requirements have been reviewed for the intended use.

For research into StyleGAN itself, work from NVIDIA's official repository and paper rather than scraping an independent website. Record the exact checkpoint, environment, seed, settings, licence, and output. That route provides reproducibility, but the original StyleGAN materials and FFHQ package still have non-commercial licence constraints. A newer model may have a different licence and behavior, so it requires its own source pack.

For media-literacy education, the historical site can be discussed without distributing its images as reusable assets. A lesson can compare the 2018 paper, the 2019 public demonstration, and the current domain state. The useful question is not whether students can memorize old artifacts. It is whether they can ask who created an image, what evidence supports its origin, and what uncertainty remains.

Evaluate any current alternative against the actual job, rights, controls, and disclosure needs. A matching search phrase or free button does not establish that it is suitable.

Method, dates and remaining unknowns

This evaluation used primary NVIDIA research, code, dataset, and licence material, plus direct HTTP inspection of the current domain on September 20, 2026. The original StyleGAN paper was first submitted on December 12, 2018. The distinct StyleGAN2 paper was first submitted on December 3, 2019. Those dates distinguish the research releases but do not establish the website's server history.

The site observation is volatile. The home page, background response, domain status, or operator may change after this review. Recheck the root page and critical routes close to any decision that relies on the description. A changing image response should remain described only as a changing response unless server behavior or operator documentation proves more.

The HTTP observation was repeated on September 20, 2026: one GET each for the home page and /image, followed by three consecutive GET requests to the same /random-person.jpeg URL. Each response was recorded with UTC time, status, content type, byte length and the SHA-256 of the received body. Requests used Cache-Control: no-cache without a changing URL parameter. All three JPEG responses had different hashes. This establishes changing responses, not image generation. The check did not include browser interaction or model execution. It refreshed the domain observation only; the NVIDIA research, dataset and license discussion retains its earlier source review.

We did not run the historical generator, download an output for publication, test demographic behavior, measure inference time, compare faces with real identities, or obtain legal advice. We also did not locate a first-party Philip Wang archive. Those limits prevent claims about current generation, uniqueness, fairness, speed, output rights, or hands-on quality.

The final decision is therefore evidence-based but intentionally conservative. Treat the site as a historically important demonstration and use it to understand the provenance questions synthetic faces raise. Do not use the current domain as the foundation of a new synthetic-face workflow until the service identity and exact-output terms are verifiable. If you need a current, source-documented system instead of a historical case study, begin with the AI model directory and apply the same rights and provenance checks.