1991–2026 · Software

From morphing to Sora: a short history of AI video

Last reviewed 1 November 2026

In 1991 a few seconds of faces melting into each other in a Michael Jackson video took a specialist studio weeks. Thirty-odd years later a sentence typed into a browser produces a moving clip. This is how computers learned to make video, from hand-placed morph lines to GANs, deepfakes and text-to-video models such as Sora.

Before AI: the morph

Computer-made faces were a novelty long before computers could make them. Max Headroom, the “computer-generated” TV host of 1985, was the actor Matt Frewer in prosthetic make-up, because no computer of the time could render a convincing talking head.

The first effect that really looked like a computer changing one thing into another was the morph. Industrial Light & Magic used its MORF software in the 1988 film Willow to turn a sorceress through a series of animals. Three years later the technique reached a far bigger audience. In the closing section of Michael Jackson’s “Black or White” video, released in November 1991, the faces of a series of men and women of different ages and backgrounds dissolve smoothly into one another.

That sequence was made by Pacific Data Images. Its engineers Thaddeus Beier and Shawn Neely published the method a year later at the SIGGRAPH graphics conference as “feature-based image metamorphosis”. Animators drew matching lines on the two faces (eyes to eyes, jaw to jaw), and the software warped and cross-faded the images so those features stayed lined up. It was clever geometry, not learning. The computer knew nothing about faces except the lines a person had placed on every key frame, which is why a few seconds of footage took weeks of skilled work.

Teaching computers to move a face

The next step was getting software to change what a real person appeared to do. In 1997 researchers at Interval Research, Christoph Bregler, Michele Covell and Malcolm Slaney, presented Video Rewrite. It took existing footage of a speaker and new audio, then rearranged and blended mouth images so the lips matched words the person had never said. It is usually cited as the first system to automate this kind of facial reanimation.

For almost twenty years such work stayed in research labs. Then it sped up. Face2Face, shown in 2016 by a team from Erlangen, the Max Planck Institute and Stanford, transferred one person’s expressions onto a video of another in real time using an ordinary webcam. In 2017 researchers at the University of Washington published “Synthesizing Obama”, which generated realistic lip movements for Barack Obama from audio alone, learned from hours of his weekly addresses.

GANs: two networks that compete

The idea that made generated images look real came in June 2014, when Ian Goodfellow and colleagues at the University of Montreal described generative adversarial networks, or GANs. A GAN pairs two neural networks. One, the generator, makes images. The other, the discriminator, tries to tell them apart from real ones. Each improves by competing with the other, and over many rounds the generator learns to produce pictures the discriminator cannot reject.

Early GAN images were small and blurry. Within four years they were not. Nvidia’s StyleGAN, published in December 2018, produced high-resolution portraits of people who do not exist, and in February 2019 the website thispersondoesnotexist.com began showing a new one on every page load. Generating a convincing still face had become a solved problem. Moving ones were next.

The deepfakes era

The word “deepfake” comes from a Reddit user called deepfakes who, in late 2017, began posting pornographic videos with celebrities’ faces swapped onto performers’ bodies. The method used autoencoders, neural networks that learn to compress and rebuild a face, trained on photos of the two people. Motherboard reported on it in December 2017, and in January 2018 a desktop program called FakeApp made the technique available to anyone with a decent graphics card.

The reaction was quick. Reddit banned the deepfakes community on 7 February 2018, and Twitter and Pornhub announced bans around the same time. In April 2018 BuzzFeed and the director Jordan Peele released a public service video in which a deepfaked Barack Obama warns viewers not to trust everything they see online.

Deepfakes were the first time machine-made video caused real harm at scale, mostly to women whose faces were used without consent. They later became a fraud tool too. Our article on AI and security describes the 2024 case in which an employee of the engineering firm Arup transferred about US$25 million after a video call with deepfaked colleagues. Face-swapping, though, still started from real footage. Making a scene from nothing needed a different approach.

From words to pictures to moving pictures

That approach was the diffusion model. A diffusion model is trained by taking real images, adding random noise step by step until nothing is left, and learning to reverse the process. To generate, it starts from pure noise and removes it a little at a time, steered by a text description. A 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel made the method practical, and in 2022 it went public. OpenAI’s DALL-E 2 arrived in April and the open Stable Diffusion model in August, and typing a sentence to get a picture became an everyday activity.

Video followed within months. Tsinghua University released CogVideo in May 2022. Meta showed Make-A-Video in September, and Google announced Imagen Video and Phenaki in October. The clips were a few seconds long, low resolution and often strange, but they were made from text alone. In 2023 the New York start-up Runway released Gen-1, which restyled existing video, and then Gen-2, one of the first text-to-video tools the public could use.

The best-known clip of that period is a failure. In March 2023 an AI-generated video of Will Smith eating spaghetti, made with an open-source ModelScope model, spread widely because it was so obviously wrong: melting faces, hands merging with forks, noodles appearing from nowhere. It became an informal benchmark, and people still recreate it with each new model to show how far the technology has come.

Sora and the race that followed

On 15 February 2024 OpenAI showed Sora, a model that made clips of up to a minute with a steady camera, consistent characters and plausible physics, at least most of the time. Its demonstration videos, including a woman walking down a rain-soaked Tokyo street, were a clear step up from anything shown before. OpenAI described Sora as a diffusion transformer that breaks video into small “spacetime patches”, much as a language model breaks text into tokens, which let it train on footage of many lengths and shapes.

Competitors followed within months. Google announced Veo in May 2024, and the Chinese video platform Kuaishou launched Kling in June. Sora itself became available to paying ChatGPT users in December 2024. In May 2025 Google’s Veo 3 generated synchronised sound, including dialogue, alongside the picture, removing one of the clearest signs that a clip was synthetic. OpenAI answered on 30 September 2025 with Sora 2, which also added sound and came with a social app built around generated videos.

Each release also restarted the arguments that had followed deepfakes: what the models were trained on, who owns the style of a film-maker or actor, and how viewers can tell what is real. Labelling schemes such as the C2PA content credentials standard, and visible or hidden watermarks on generated clips, are the industry’s main answer so far. None of them is foolproof.

Where it is now

The biggest change is not quality but cost. The “Black or White” morph needed a specialist studio, custom software and weeks of artists’ time for a few seconds of footage. A text-to-video model in a research lab in 2022 needed a data centre. Today an AI video generator such as HighReach can make a short clip from a sentence or a still photo in a browser, something that needed a visual effects team when the Black or White morph aired.

Turning a still image into motion has become one of the most common uses. A family photo, a product shot or a drawing can be given a slow camera move, drifting clouds or a turning head. That is the same job the morph did in 1991, done by a model that has learned from millions of videos rather than from lines an animator drew.

The limits are clearer than the hype suggests. Clips are still short, usually seconds rather than minutes, and keeping a character or object identical from shot to shot remains hard. Hands, text and fast motion still go wrong. Most services are paid, run in the cloud and have rules about real people and copyrighted characters.

The history also explains why sceptics are uneasy. Every step that made video easier to create, from FakeApp to Sora, also made it easier to fake. Making a clip now takes a prompt, not a studio, and checking where a video came from matters more than ever.