How Nvidia researchers generated pictures of faces that appear to be real but aren't by analyzing photos of celebrities and detecting patterns
To create the final image in this set, the system generated 10 million revisions over 18 days. — The woman in the photo seems familiar. Tweets: @alexsteffen , @pnhoward , @rafael_parente , @cademetz , @nvidiaai , and @nvidiaaidev Tweets: Alex Steffen / @alexsteffen : Humanity isn't even vaguely ready for the new technologies enabling easy, seamless fabrication of imagery (especially video). Combined with a fact-check-free 24-hour news cycle, this is going to unleash a serious force for societal destabilization. http://www.nytimes.com/... Phil Howard / @pnhoward : How an A.I. ‘Cat-and-Mouse Game’ Generates Believable Fake Photos http://www.nytimes.com/... fake headshots to go with fake news Rafael Parente / @rafael_parente : Researchers in Finland have developed artificial intelligence that can generate images of celebrity look-alikes — and another system that tests how believable they are https://nyti.ms/2EDgt6G Cade Metz / @cademetz : Nvidia boss Jen-Hsun Huang recently asked me why the @nytimes was doing a piece on “Progressive GANs,” new Nvidia tech that generates realistic images of fake people. Because it shows how A.I. will change how we view the world, for better or worse: https://www.nytimes.com/... Nvidia Ai / @nvidiaai : Recognize this celebrity? Unlikely — because she was created via #AI with a system that analyzes thousands of photos & recognizes common patterns to create new, believable photos, even of celebrities. More on this @NVIDIA research via @nytimes' @cademetz: http://nvda.ws/2CDo1ZR pic.twitter.com/fLYTmayO0h Nvidia Ai / @nvidiaaidev : Thanks @nytimes for writing about one of our recent @NVIDIA Research #AI projects! Using a single Tesla P100 GPU and #GANs, the team generated photorealistic pictures of fake celebrities. http://nvda.ws/2CztMYD pic.twitter.com/BojNrtHQc6
Context & Ripple Effects
This early-2018 piece captured Nvidia researchers demonstrating a GAN that could generate photorealistic faces of people who don't exist — 10 million revisions over 18 days to produce one image — trained on celebrity photos. Commentators quoted at the time, including Alex Steffen and Phil Howard, flagged the obvious risk: seamless fabricated imagery meeting a fact-check-free news cycle.
The arc since then splits in two directions. On one side, capability kept compounding — Nvidia's own follow-up showed how far AI image generation advanced within the year in creating realistic and customizable faces, and Microsoft's VASA-1 extended the trick from stills to talking-head video driven by a single portrait and an audio file from a portrait photo. On the other side, the same technique was commercialized into consented uses: Synthesia built multilingual training videos on it corporate-friendly uses for deepfakes, and Hour One began paying people to license their likenesses for AI-voiced characters paying people to use their likenesses.
First-order effects
- The immediate effect of the demo was evidentiary: photographic realism stopped being proof of authenticity, putting journalists, platforms, and fact-checkers on notice exactly as Steffen and Howard warned in the article.
Second-order effects
- Startups converted the research into a business model — Synthesia and Hour One turned synthetic faces into paid products, creating a licensed-likeness market where individuals are compensated rather than impersonated.
Third-order effects
- If the pattern holds, the binding constraint shifts from generation quality to verification: provenance standards and likeness-rights frameworks become the infrastructure layer that determines whether synthetic media is trusted at all.
The trend: Synthetic media is maturing from research demonstrations of fake faces into a commercial likeness economy, where trust and consent — not rendering quality — become the scarce resources.