Runway releases Gen-4

Background

Video generation models turn written descriptions into moving footage. Most of what they produce comes out as short, separate clips, and one of the practical obstacles to using them in longer pieces is continuity: a character's face, the colour of a jacket or the look of a room can drift from one shot to the next, because each shot is generated on its own. Holding those elements steady is a precondition for anything that resembles a scene rather than a collection of fragments.

What Runway released

On 31 March 2025 Runway released Gen-4, a new video generation model ^[1]. The company centred the release on a single goal: keeping characters, objects and locations consistent across shots ^[1][2]. That is a narrower aim than all-purpose photorealism, and it points at how such models are actually used — not as one-off clips, but as pieces that have to be cut together into a sequence.

How it works

Gen-4 works from visual references combined with text instructions, and does not require fine-tuning ^[1]. Fine-tuning means taking an already trained model and adjusting its parameters further on additional examples so that it specialises in something — a particular face, a particular style, a particular product. It takes data and compute, and it produces a customised version of the model for each job. Gen-4's approach avoids that step: the user supplies reference material showing what should stay the same, and describes the shot in words. The reference carries the appearance; the text carries the action and the camera.

Who could use it

The model rolled out to both individual and enterprise customers ^[1]. That covers the two ends of the user base for this kind of tool — people working alone on short pieces, and organisations that need to fit the output into an existing production pipeline.

Competition and the film business

TechCrunch noted that Runway faced stiff competition in video generation from OpenAI and Google ^[3]. The same report described Runway moving beyond selling a tool: the company had signed a deal with a major Hollywood studio and earmarked millions of dollars to fund films made with AI-generated video ^[3]. Funding finished films is a different kind of bet from shipping a model — it puts the company's own money behind the claim that the output is usable in production.

Why it matters

For video generation, consistency is the line between a demonstration and a production asset. A model that can hold a character and a place steady across shots can be used to build a sequence; one that cannot forces the user to discard most of what it makes. By making cross-shot consistency the headline feature and by letting users steer it with references rather than retraining, Gen-4 aimed at the part of the workflow where a clip becomes a scene ^[1][2]. The competitive context matters too: video generation is contested by large labs with their own models, and the Hollywood arrangement shows one route a smaller company can take — attaching the technology to existing film production rather than competing on the model alone ^[3].