Runway AI Release Notes

For most of the internet's history, the interaction model has been the same: you type something, and a result comes back. Web searches. Emails. Product pages. Chatbots. Large language models made the exchange more fluid and conversational, but at its core, it's still text in a box.

We think that era is ending.

The future of online interaction is real-time video, generated on the fly – responsive, personalized, alive.

What Real-Time Video Generation Actually Means

Real-time video generation refers to AI models that synthesize video frame by frame, live, in response to user input – rather than producing a completed output all at once.

Pre-rendered video is static: it was made once, and you watch it. Generative tools like Gen-4.5 let you create video from scratch, but the output is still an artifact you produce and then share. Today's top generative models are limited by their architecture: no matter how complex the prompt, or how sophisticated the output, the model is still predicting and generating the output you want all at once.

Real-time video generation is interactive. The model generates what you see as you see it, responding to what you say and what you do. Every frame is synthesized in the moment, conditioned on the current context of the interaction.

This is only possible because of a fundamental shift in how we think about video models. At sufficient scale, video models go beyond generating plausible-looking footage – they begin to develop an internal representation of how the world works. These micro-interactions—how faces move when people speak, how expressions change when emotions shift, how physics propagates when forces act on objects—drive much of how we experience the world around us.

GWM-1, which we launched last December, is our first general world model family – an autoregressive model that generates frame by frame, runs in real time and can be controlled interactively with actions: camera pose, speech, robot commands. It comes in three variants today: Runway Characters for conversational characters, GWM-Worlds for explorable environments and GWM-Robotics for robotic manipulation. These are distinct post-trained models now; we're working toward unifying them under a single base.

What This Unlocks

The near-term applications of real-time video generation touch almost every domain where digital interaction matters.

Gaming and Interactive Entertainment

NPCs in games today are largely static, with branching dialogue trees, pre-recorded voice lines and scripted behaviors. Real-time video generation makes it possible to build characters that actually listen and respond, holding genuine conversations with players about the world they inhabit. Imagine a guide who can answer any question about lore, or a sports simulation responding live to your choices.

Beyond traditional gaming, real-time video generation opens up new territory for fan platforms, creator experiences and interactive narrative.

Learning and Education

The case for interactive video in education is straightforward: a personalized tutor who reacts to confusion, adjusts explanations in real time and responds to where you actually are in your understanding is categorically more effective than a static lesson. Real-time video generation makes it possible to deploy that kind of experience at scale, across languages, grade levels, subjects and time zones.

There's also an access dimension. A real-time video experience is available at 3am, in any language, with infinite patience. For a student who needs to work through a concept 50 times without embarrassment, or one who's far from any formal support infrastructure, that matters.

Training and Simulation

Some of the most consequential conversations people have in the workplace can't be fully prepared for in a classroom. Real-time video generation enables realistic practice for high-stakes scenarios: an upset customer who escalates, a nervous interviewee who needs to be put at ease, a manager who pushes back on your proposal. For use cases like sales coaching, clinical simulation or law enforcement de-escalation training, real-time video generation is key.

Customer Experience and Brand

The current state of the art for AI customer support is a text chatbot with a company logo on it. Real-time video generation clears that bar significantly by presenting a responsive, expressive presence that engages customers more like a human interaction and less like a form submission. For brands with existing characters or mascots, the opportunity is especially interesting: IP that's existed as a static asset can become genuinely interactive.

Characters: Real-Time Video Generation, Available Now

The most tangible example of real-time video generation we have today is Runway Characters: an audio-driven interactive video generation model, built on GWM-1, that produces fully expressive conversational characters from a single reference image.

The model handles what makes a face feel alive: natural eye movements, lip-sync, facial expressions, gestures during speaking and listening. It sustains quality across extended conversations. And because it ships as an API, developers can create a branded character that can pull from your product catalog, open a support ticket and escalate to a human agent. Companies like BBC, R/GA, Silverside and Supersonik are already building with Runway Characters.

Characters is live now for developers at dev.runwayml.com and available in the Runway web app for anyone who wants to experience it directly.

Nothing quite like this has existed before, which means the questions around responsible deployment are ones we're actively working through. We've written about our approach to identity, consent and transparency—and what responsible deployment looks like—here.

What Comes Next

We wrote last year that we expect to achieve human-scale world simulation within half a decade. Within a decade, we expect to simulate physics and biology accurately enough to meaningfully address a significant percentage of today's scientific challenges.

That's a long arc, but the near-term steps are already visible, and real-time generation will continue to improve. The consistency across extended interactions will deepen. The action spaces these models can respond to will expand.

Enterprises can build with real-time video generation today. To learn more, visit runwayml.com/enterprise or contact our sales team.

AI video generation models have improved exponentially since we released Gen-2, the first publicly available text-to-video model, in early 2023. Two years ago, these models took several minutes to generate choppy, pixelated clips that were a few seconds long. Today, leading video generation models can reliably produce outputs that are virtually indistinguishable from real video.

This week, we released image-to-video capabilities for Gen-4.5, our latest base model. Today, we're publishing new research, evaluating people's ability to determine if a five second video is real, or was generated by our model. We're also launching a new site where anyone can try for themselves.

For this research study, we recruited a random sampling of 1,043 participants. Each participant viewed 20 videos (10 real, 10 generated) in randomized order and judged whether each was real or AI-generated. Each video was generated only once — the outputs were not edited, and no video was regenerated to improve quality or skew results.

Results

Over 90% of participants could not reliably distinguish Gen-4.5 outputs from real video.

Only 99 of 1,043 participants (9.5%) achieved statistically significant accuracy (>=15/20 correct, p < 0.05, binomial test). Overall detection accuracy was 57.1%, only slightly above chance. Performance was similar on real (58.0%) and generated (56.1%) videos, indicating no systematic detection strategy.

Detection accuracy varied by content category. Human-related videos (faces, hands, actions) were easier to detect (58-65%), while animals and architecture fell below chance (45-47%) — participants were more likely to mistake generated videos for real than vice versa.

These findings represent a fundamental shift in how we should think about video authenticity. For years, we've been building toward General World Models. Realistic simulation is a prerequisite for solving hard problems in the physical world. Gen-4.5 is the most capable simulator we've built yet. But that capability comes with responsibility. When 90% of people cannot reliably distinguish synthetic from real footage—and when generated content in certain categories is more convincing than reality—detection is an inadequate strategy for trust and verification.

Conclusions

Video generation models will continue their exponential improvement, assuming we continue to scale training data and compute. The AI industry and society at large have reached a tipping point, where the average person cannot determine if a video is generated by AI or not.

From photography to photoshop to traditional CGI, technology has consistently shifted public opinion on what makes a piece of content "real." As AI models continue to improve, we expect another, similar shift. We believe that foundational model developers, including Runway, have a responsibility to drive public conversation around the quality of model outputs, and explore how we can mitigate the societal challenges this technology will introduce while continuing to push the boundaries of AI research and innovation.

All Runway-generated outputs include C2PA metadata, allowing us to certify the origin and provenance of the content our models produce. This open technical standard is embraced by a wide variety of media companies and news organizations, but it is not infallible. We need to build new, more capable standards that preserve trust while enabling creative possibility. That requires technical solutions like C2PA, but also new literacies, updated editorial standards and ongoing dialogue about authenticity.

Moving forward, we're committed to three principles: maintaining transparency about our models' capabilities, collaborating with industry partners on verification standards and engaging directly with creators, enterprises and policymakers to establish new norms for synthetic media.

Methodology

Source videos were sampled from Filmpac across five content categories: faces, full-body human motion, animals, nature scenes and urban environments. For each category, we selected examples representative of content people often aim to generate. The first frame of each video was extracted and used as input to Gen-4.5 with default settings. Each video was generated once, with no regeneration or post-processing. Real and generated clips were trimmed to five seconds and matched in resolution. Participants could view each video for up to 10 seconds before making their judgment. Participants who achieved greater than 75% accuracy (>=15/20 correct, p < 0.05, binomial test) were classified as successful detectors.