Runway Gen-4

What Gen-4 is

Runway Gen-4 is a video generation model that Runway released on 31 March 2025. The company describes it as built for media generation and for what it calls world consistency, meaning the generated world stays stable from one shot to the next ^[2]. The model produces clips between 5 and 10 seconds long ^[3].

Consistency as the central problem

In most video generation, each clip is produced more or less on its own. A character's face, a room's layout or an object's design can drift between shots, which is exactly what makes such footage hard to cut into a continuous scene. Gen-4 is aimed at that problem: it keeps characters, locations and objects consistent across scenes, and it can do so starting from a single reference image, holding the look across different lighting conditions ^[1]. A reference image is a still picture supplied by the user that acts as an anchor, a way of telling the model what a person or place is supposed to look like before it animates anything.

Runway made the same point at launch, saying that Gen-4 kept characters consistent from a single reference image across any lighting condition, location or treatment ^[4].

Working from references, not retraining

The model works from visual references and text instructions without fine-tuning or extra training ^[1]. That matters for how the tool is used in practice: a user does not have to first train a model on a particular person or style in order to reuse them, but supplies material and written direction instead.

Gen-4 can also regenerate a scene from different angles and positions using reference images and a shot description ^[1]. In effect, the same moment can be rendered again from another camera position rather than being frozen at the angle it was first generated at, which is closer to how coverage works on a real shoot.

Release and access

At launch the model was rolled out to individual and enterprise customers through the web app and through the iOS and Android apps ^[1][2]. Runway did not disclose the sources of the data the model was trained on ^[2].

What Runway showed with it

To test the model's narrative capabilities, Runway released a collection of short films and music videos made entirely with Gen-4 ^[4]. The launch also highlighted workflows for consistent objects, full scene coverage, production-ready video physics and GVFX ^[4]. These are the pieces a production needs beyond a single good-looking clip: objects that stay the same when they reappear, enough angles to cover a scene, motion that holds up on screen, and effects work built on top of the generated footage.

Where it fits

Runway described Gen-4 as a step towards models that simulate real-world physics ^[1]. That framing places the model less as a tool for making isolated clips and more as groundwork for systems that maintain a coherent, physically plausible world over time. The consistency features are the part of that ambition that is already visible in the product: the ability to hold a character, a location or an object steady across many short shots is what turns a set of generated clips into something that can be edited as a scene.