D-ID

The name stands for de-identification, and that is no accident

D-ID was founded in 2017 by Gil Perry, Sella Blondheim and Eliran Kuta, and it started out in exactly the opposite direction from the one it is known for today. Its first product was meant to protect people from automatic face recognition. It made tiny changes to a photo that the human eye cannot notice at all, but that stopped face recognition systems from matching the photo to the person in it. The idea came from a personal need of the founders, who had served in sensitive roles and could not post photos of themselves online even after their discharge. In 2017, when face recognition systems were starting to become commercially available at large scale, this was a real and new problem. The solution was technically impressive, because it required understanding the structure of the face precisely enough to disrupt the very features the systems rely on, rather than the general appearance. The irony that runs through the whole story is worth noticing: the same company that was born to make faces unrecognizable is the one that taught computers to animate faces convincingly.

From protection to creation: the same model, the opposite direction

The path from protecting faces to creating faces looks like a contradiction, but it is actually a natural result. Both tasks need exactly the same thing: a model that understands how human faces are built and how they change. If you know how to change facial features in a controlled way so that a computer will not recognize them, you also know how to change them in a controlled way so they look as if they are moving, talking or smiling. The only difference is the direction in which the capability is applied. That is how the company rolled into a completely different product: animating a still photo into a talking figure, with head movements, blinks and lip sync to a voice. Open models such as SadTalker, published in 2022, demonstrate the field that research calls talking head generation: they produce such a video from a single face image and a speech clip. This is a pattern that repeats across the whole field. Tools built for recognition and protection and tools built for creation are often exactly the same technology, and only the question of what is done with it separates them. This point is worth returning to in any discussion of regulating this technology. You cannot ban the capability itself without also harming uses no one disputes, such as accessibility, translation and historical restoration. The real debate is about use, consent and labeling, not about whether the model exists.

February 2021: Deep Nostalgia

The moment that brought the technology into global awareness came in February 2021, when the Israeli company MyHeritage launched Deep Nostalgia on top of it: a service that animates old photos of relatives who have died. A person uploads a photo of their grandmother from the 1950s, and seconds later sees her turn her head, blink and smile at them. The service went viral within days, the Reuters news agency described it as a spooky tool, and according to the company it alone produced about one hundred million animations by March 2022. It is important to tell the two companies apart here, because the confusion is common: MyHeritage ran the service for the public, and the technology behind it was D-ID's. The version from that time also did not make the figure talk, but only moved its face. That very distinction later became central to the debate. The number also shows something about how fast tools like this are adopted: what had until then required a studio, a budget and professionals became something a person does on their phone while sitting in the living room, without knowing anything about the technology behind it. Deep Nostalgia is the prominent Israeli case of a wider phenomenon, which today has providers and price lists in several countries and is explained in the Hebrew entry on the digital resurrection of the dead.

The public debate it opened

Deep Nostalgia was a key moment less because of the technology and more because of who used it. Until then, the debate about video manipulation had revolved around politicians, celebrities and scams, that is, around scenarios of harm. Here, for the first time, millions of ordinary people applied such technology to their own families, of their own free will and out of emotion. Suddenly the question of what may be done with the likeness of the dead stopped being theoretical. Reactions were clearly split, and the split did not follow technical knowledge. Some described the experience as moving and comforting, and some described it as violating something: memory, respect for the dead, the line between a photograph and an invention. Both sides were right in a sense, and that is exactly why this debate continues today and has not been settled. What made this moment significant is that it did not stay theoretical. Millions of people made a small, personal decision, to upload a photo of a relative who had died and watch them move. Each of these decisions is in fact a position on a question that experts argued about for years without resolution.

The move to the business market

After the wave of attention, the company turned to the market where the money is: talking avatars for business use, which can be produced from one photo and written text, for customer service, employee training and marketing. In March 2022 it raised $25 million in a Series B round led by Macquarie Capital, bringing its total funding at the time to $48 million. Customers reported at the time included MyHeritage, Warner Bros. studios, the food company Mondelez and advertising agencies. The business logic here is simple. Producing a training video with an actor, a camera operator and a studio costs thousands of dollars and takes weeks, while an avatar that speaks typed text costs little and can be updated in minutes. The price, of course, is that the viewer knows, or does not know, that they are talking to a figure that does not exist. That is exactly the point where the question of labeling and transparency becomes practical rather than philosophical. It is also worth noting what this does to the production industry itself. When the cost of producing a video falls by an order of magnitude, it is not only that more videos are made: videos are made that would never have been produced at all, such as a short training video for an internal company process. This is an expansion of the market, and alongside it, clear pressure on those who make a living from traditional production.

The uncanny valley, and the deepfake fault line

Technically, the main difficulty in animating a portrait is not the lip movement but everything else. A person also speaks with their eyes, their eyebrows, small head movements and the rhythm of their breathing. When these details are missing or out of sync with each other, the result falls into what is called the uncanny valley: it is similar enough to a person that the brain expects a person, and different enough that it feels something is wrong. Much of the progress in the field since 2021 concerns exactly these details and how they sync with the voice. Publicly, the technology sits right on the fault line of deepfakes. The same capability that produces a legitimate virtual presenter for a training video can produce a video of a real person saying things they never said. The company publishes usage restrictions and verification mechanisms, but the line between commercial and harmful use remains an open issue across the whole field, not only at D-ID. As of today there is no complete technical answer to the problem. Digital watermarking, provenance verification and automatic detection tools all exist, and all are partial. In recent years the debate in the field has shifted from the question of how to detect a fake to the question of how to prove authenticity.