What it is
Runway is a company, and a website of the same name, that makes AI tools for creating and editing video. You can type a description and get back a short video clip, upload a still picture and have it brought to life, or take footage you already shot and change it: remove an object, swap the background, relight the scene, or restyle the whole thing to look like a painting. Its tools run in a web browser and in phone apps, so you don't need an expensive computer or editing software to try them.
The company was founded in New York in 2018 by three graduate students from New York University's Interactive Telecommunications Program (ITP), an art-and-technology course: Cristóbal Valenzuela and Alejandro Matamala from Chile, and Anastasis Germanidis from Greece. Runway began as Valenzuela's thesis project, shown at ITP's spring show in 2018: a desktop app that let artists use machine learning models without writing code. That origin still shapes the company. It has always pitched itself at filmmakers, designers and artists rather than at programmers.
Its early products were editing helpers, such as a tool that cuts a moving person out of their background. Then Runway moved into generating video. Along the way it co-developed the first release of the image generator Stable Diffusion in August 2022, a story told on the Stable Diffusion page. After that, it released a series of its own video models, each named "Gen" plus a number:
- Gen-1 (February 2023) transformed existing video. You gave it a clip and a text prompt or reference image, and it redrew the clip in the new style.
- Gen-2 (announced March 2023, open to the public in June 2023) added text to video: type a description and get a clip of a few seconds, made from scratch. It was one of the first text-to-video tools anyone could use, and the early results were short, blurry and wobbly.
- Gen-3 Alpha (June 2024) was a large jump in quality and motion.
- Gen-4 (March 2025) focused on consistency: keeping the same character, object or place looking the same from shot to shot, guided by a single reference image.
- Aleph (July 2025) was an editing model. It changes video you already have, for example adding or removing objects, showing the scene from a new angle, or changing the lighting.
- Gen-4.5 (December 2025) was its newest general video model at the time of writing, with an updated editing model, Aleph 2.0, following in May 2026.
Runway has not published the full details of how its current models are built or exactly what they were trained on, but, like most AI image and video tools, they are generally understood to use the diffusion model approach: start from random visual static and clean it up, step by step, into frames that match the prompt.
At the time of writing, Runway has grown well beyond single clips. Its app includes an "agent", launched in May 2026, that takes a described idea and assembles a multi-shot video with voiceover and music. It also offers some other companies' models inside its app, such as Kling from Kuaishou and image models from Google, alongside its own. The company increasingly describes its long-term goal as building "world models", AI systems that simulate how the physical world behaves, and it now offers versions of its technology for robotics as well as for creators. Its web address moved to runway.com during this period; the older runwayml.com still redirects there.
Why it matters
It brought video into the same "describe it and get it" era as images. Before tools like Runway, a few seconds of original footage meant a camera, locations, actors or animation skills. Now a small business owner, a teacher or an author can produce a short clip from a sentence and a picture. The clips are short and imperfect, but for social media, presentations and mood-setting they are often good enough.
It is used by professionals, not only hobbyists. Runway's pitch has always been to working filmmakers. Its editing tools were used on an Oscar-winning film (see the examples below), and it has signed deals with the Hollywood studio Lionsgate and the television group AMC Networks. Those deals are as much about marketing, planning and "pre-visualisation" (rough previews of shots before filming) as about putting AI footage on screen.
It sits in the middle of the argument about training data. Video models need enormous amounts of video to learn from, and a 2024 news investigation reported that Runway had collected much of it from YouTube without permission. Runway is also a defendant in the artists' copyright lawsuit Andersen v. Stability AI, covered on the Stable Diffusion page. How those questions are settled will shape what AI video tools are allowed to do and what they cost.
It shows how fast this field moves. Gen-2 clips in 2023 lasted about four seconds and often melted into nonsense. Three years later, the tools can keep a character recognisable across shots and edit real footage convincingly. That speed is exciting for creators and unsettling for anyone who relies on video as proof that something happened.
Real-world examples
The rock scene in Everything Everywhere All at Once (2022)
Everything Everywhere All at Once won the Academy Award for Best Picture in 2023 with a visual effects team that Variety reported was only eight people. In a February 2023 Variety article, the visual effects artist Evan Halleck said he had used Runway's "green screen" tool, which cuts a moving subject out of its background, on the scene where the main characters are two rocks. The shots were filmed with a rig pushing gravel and sand, which then had to be removed. Tracing around a subject frame by frame by hand is called rotoscoping and is famously slow. Halleck told Variety that the tool cut the characters out in minutes rather than half a day, and that it "was cutting things out better than my human eye was". This is worth being precise about: Runway did not generate any of the film. It was an AI-assisted editing tool that saved a small team time on tedious work, which is still one of the most practical uses of this technology.
The YouTube training data report (July 2024)
In June 2024 Runway released Gen-3 Alpha, saying it was trained on both videos and images. Asked by TechCrunch at the time where the training data came from, co-founder Anastasis Germanidis said only that the company used "curated, internal datasets". On 25 July 2024 the news site 404 Media published what it said was an internal Runway spreadsheet linked to the model's training. It listed thousands of YouTube channels, including those of media companies such as Disney, Netflix and The New Yorker and of well-known creators such as Marques Brownlee and Casey Neistat, along with links to pirated films. A former employee told 404 Media the videos had been downloaded with open-source software. Runway did not respond to requests for comment from 404 Media or from the photography site PetaPixel. The report is not a court finding, and Runway has not publicly confirmed or disputed the spreadsheet. But it put a concrete face on a question every AI video company faces, and the creators named had not been asked or paid.
The Lionsgate deal: big plans, then a reality check (2024 to 2026)
On 18 September 2024, Runway and Lionsgate, the studio behind the John Wick and Hunger Games films, announced that Runway would build a custom AI model trained on Lionsgate's own film and TV library, for use in pre-production and post-production. Because the studio owns the material, this looked like a way round the training-data problem. A year later, on 22 September 2025, the entertainment news site TheWrap reported that the project had run into trouble. People familiar with it, speaking anonymously, said one studio's catalogue was too small to train a strong model, and that actors' rights in their own likenesses raised legal questions. Lionsgate said it was pursuing AI "on several fronts as planned" and that the deal was not exclusive. In June 2026, TheWrap reported that Lionsgate had taken an ownership stake in Runway and that the two planned short AI-made series based on Lionsgate stories. The arc is a useful reminder that studio announcements describe intentions, not finished films.
A film festival for AI films (2023 onward)
Runway started its own AI Film Festival in February 2023, with screenings in New York and San Francisco and ten finalists chosen from several hundred entries. It grew quickly. Runway said the 2025 edition received more than 6,000 submissions, and in August 2025 the finalists were shown in IMAX cinemas in ten US cities. For 2026 the event was renamed the AI Festival and added categories beyond film, such as advertising, design and gaming. Supporters see it as a showcase for a new kind of filmmaker. Critics point out that the festival is run by a company that sells the tools, which is worth remembering when reading any claim about how good AI films have become.
An everyday scenario: a book trailer from a cover image
Picture Priya, who has self-published a fantasy novel and wants a 30-second trailer for social media. She uploads her finished cover art to Runway and asks for slow motion: mist drifting across the castle, a banner stirring in the wind, the camera gently pushing in. The first try makes the castle's towers bend like rubber, so she simplifies the prompt to describe only the camera move and the mist. She makes four more short clips from her concept art, strings them together in a free video editor with her title, a line from a review and a "Buy now" card, and adds music she has the rights to. It takes an evening and uses most of a month's credits on a basic paid plan. She checks the result for anything that looks like someone else's character or artwork before posting.
Try it yourself
Menus, model names and plans change often, so treat these steps as a guide to the process. You need a free account, which asks for an email address.
- Sign up at runway.com. At the time of writing, the free plan gives you 125 credits once, which never expire but are never topped up, and everything you make on it carries a Runway watermark. Runway says the free plan includes "a limited set" of tools that changes over time; when we checked, its help pages listed image generation, text to speech and upscaling. If a video model shows an upgrade prompt, it isn't in the free plan at the moment.
- Understand what a credit buys. At the time of writing, Runway's pricing page says Gen-4.5 video costs 12 credits per second, so a 10-second clip costs 120 credits, almost the whole free allowance. The Standard plan costs 15 US dollars a month (12 dollars a month billed yearly) for 625 credits, about 52 seconds of Gen-4.5 video. Check the current prices yourself.
- Start from a still image, not a blank prompt. Image to video gives you much more control. Use a picture you own: your book cover, a photo you took, or an image you made with an image tool.
- Describe only the motion. Runway's own guidance for image to video is to describe what moves, not what is already in the picture. Try something like this, adapted to your image:
Slow push-in toward the castle. Mist drifts from left to right across the hills. The flag on the tower moves gently in the wind. Calm, steady camera. - Pick a short duration. At the time of writing, Gen-4.5 makes clips between 2 and 10 seconds long. Five seconds is plenty for a first test and costs half as much.
- Watch the result three times. First for the overall feel, then for anything that bends, melts or appears from nowhere, then at the edges of the frame. Faces, hands, text and logos are the usual weak spots.
- Change one thing and try again. If the result is too busy, remove instructions rather than adding them. One simple movement usually works better than three.
- Build longer videos from short clips. Real trailers are a sequence of short shots. Make several and join them in a video editor, where you also add your title text, which AI video still tends to garble.
- Check the rules before you publish. Runway's help pages say, at the time of writing, that you own what you make on every plan and may use it commercially without crediting Runway. Check whether the platform you post to requires you to label AI-generated video.
Common mistakes and misunderstandings
Expecting a finished film from one prompt. Individual clips are seconds long. Anything longer is made by planning shots, generating many takes, throwing most away and editing the rest together. Runway's newer agent can assemble a multi-shot video for you, but you still need to review every second.
Burning through credits on vague prompts. Every attempt costs credits, and video is expensive. Test ideas with short, cheap clips or with a still image first, and save longer, higher-quality generations for ideas that already work.
Assuming "I own it" means "it's protected". Runway's terms say it does not claim ownership of what you make and does not restrict commercial use. That is not the same as copyright. In the United States, the Copyright Office's January 2025 report said material generated by AI from prompts alone is not protected, although your own editing, arrangement and added material can be. Rules differ between countries. For anything valuable, ask a lawyer where you live.
Forgetting the licence you give Runway. At the time of writing, Runway's terms also say your inputs and outputs may be used to train and improve its models, and give Runway a broad, permanent licence to use them for that. Don't upload unpublished work, client footage or private photos unless you are comfortable with that, and check whether your plan offers any opt-out.
Using other people's faces, characters or footage. Animating a photo of a real person without their permission, or generating a famous character, can break Runway's rules and the law, and can hurt the person involved. The Lionsgate story shows that even a studio has to think about actors' rights in their own likenesses. Stick to material you own or have clear permission to use.
Believing video is proof. Tools like Runway can now change what happens in real footage, not only create new footage. Treat surprising video online with the same caution you would give a surprising photo. Runway says it attaches C2PA "content credentials", a kind of digital label, to what it generates, but labels can be stripped when a video is re-saved or screen-recorded.
Thinking it replaces the whole craft. Runway can give you moving images. It doesn't decide what story the trailer should tell, which shot comes next, or when to cut. Editing, pacing, sound and knowing what to leave out still make the difference.
Picking a tool by reputation alone. Runway is one of several AI video tools, alongside Sora, Veo and Kling, and each company publishes rankings that favour itself. Runway even offers some rival models inside its own app. Try the same prompt in two or three tools and judge the results for your own project.
Related elements
- Diffusion Model: the technique behind Runway's video models and most AI image tools.
- Stable Diffusion: the open image model Runway co-developed in 2022, and home of the artists' lawsuit that names Runway.
- Sora: OpenAI's text-to-video model, one of Runway's best-known rivals.
- Veo: Google's video generation model.
- Kling: Kuaishou's video model, also available inside Runway's app at the time of writing.
- Midjourney: a popular way to make the still images you then animate in a video tool.
- Adobe Firefly: Adobe's creative AI app, which added Runway's Gen-4.5 model after a partnership announced in December 2025.
- ElevenLabs: AI voices for narrating the trailers and clips you make.
- World Models: the AI systems that simulate how the world behaves, which Runway now describes as its long-term goal.
- Datasets: why the question of what an AI was trained on matters so much.
- Prompt: how to describe what you want clearly, which matters even more when every attempt costs credits.