Runway partners with Lionsgate on a custom video model

Background

Generative video models produce moving images from text prompts or other inputs, and Runway had built its business around such models. Days before this announcement, Runway launched an invitation-only API for its Gen-3 Alpha Turbo model, which let outside developers call the model through a programming interface rather than only through the company's own tools ^[1]. Lionsgate, a major film studio, held a catalogue of more than 20,000 film and television titles ^[2].

The two sides were therefore approaching the same technology from different directions: one company sold access to a general video model, the other owned a large body of filmed material and the production workflows that use it.

What was announced

On 18 September 2024, Runway and Lionsgate announced a partnership to build a custom AI video model ^[1]. The model was to be trained on Lionsgate's catalogue of more than 20,000 film and TV titles, and was intended for use by Lionsgate-affiliated filmmakers ^[1][2].

The arrangement was exclusive in a specific sense: the custom model was designed only for Lionsgate-affiliated filmmakers and could not be accessed by other Runway users ^[3]. That is different from licensing a general model to the public. A general model is trained once and then offered to anyone who pays for it; a bespoke model is trained on one customer's material and walled off for that customer's use. Here the wall ran in both directions, keeping the model out of the hands of Runway's wider user base.

Key details

Lionsgate planned to use the model first for storyboarding, backgrounds and special effects ^[1]. Those three uses sit at different points in filmmaking: storyboarding is an early planning step in which scenes are sketched out before shooting, backgrounds fill in settings behind the action, and special effects are added in post-production. Starting with storyboarding and backgrounds puts the tool at the visualisation end of the pipeline rather than at the final-image end.

Runway said the Lionsgate library it would train on included titles such as Francis Ford Coppola's Megalopolis ^[1].

VentureBeat described the deal as the first signed directly between a major film studio and an AI video model provider ^[1]. That characterisation is about the structure of the agreement rather than the technology: studios had been discussing generative video publicly, but a direct, bilateral deal to build a model on a studio's own catalogue was presented as a new kind of arrangement.

Why it matters

The announcement put two questions on the table at once. The first is where the training material comes from. A video model learns the look of what it is shown, so a model trained on a large studio catalogue inherits that catalogue's visual conventions, and the studio that owns the material is involved from the start rather than negotiating afterwards. The second is who gets to use the result. By tying the model to Lionsgate-affiliated filmmakers, the partnership framed generative video as an in-house production tool rather than a public service.

Both points matter to people outside the film industry as well. Much of the public debate around generative media has concerned models trained on material collected from the open web, with rights holders objecting after the fact. A deal in which the catalogue owner is a party to the training from the beginning is a different model for how such systems can be built.

The three named uses also say something about expectations. Storyboarding, backgrounds and effects are tasks where rough or intermediate images are useful and where a filmmaker can review and discard output cheaply, which is a more modest target than replacing principal photography.

What the announcement covered

The public information centred on the partnership itself and on the intended uses rather than on a finished product. Runway and Lionsgate announced that they would build the model; Lionsgate described where it expected to apply it first; and Runway described the material it would learn from ^[1][2][3]. The deal was announced in the same week that Runway opened its invitation-only API for Gen-3 Alpha Turbo, so the company was simultaneously widening access to its general model and narrowing access to this one ^[1].