Runway ML has introduced a new foundation model, Gen-3 Alpha, which the company says can generate high-quality, realistic scenes of up to 10 seconds long from text prompts, still images or a video sample. Offering a variety of camera movements, Gen-3 Alpha will initially roll out to Runway’s paid subscribers, but the company plans to add a free version in the future. Runway says Gen-3 Alpha is the first of a new series of models trained on the company’s new large-scale multimodal infrastructure, which offers improvements “in fidelity, consistency, and motion over Gen-2,” released last year.

Runway says Gen-3 Alpha is a step toward something it calls “general world models” — systems that understand the visual world and its dynamics, around which the company has focused a research effort.

Runway CTO Anastasis Germanidis posted on X that in addition to the aforementioned generative prompts, Gen-3 Alpha will use new prompts that “are only now possible with a more capable base model,” but did not specify what those might be (though the platform site says they’ll offer “more fine-grained control over structure, style, and motion”).

VentureBeat reports that Runway has been “collaborating and partnering with leading entertainment and media organizations to create custom versions of Gen-3,” which “allows for more stylistically controlled and consistent characters, and targets specific artistic and narrative requirements, among other features.” Runway is inviting inquiries for custom Gen-3 models.

Runway explains in its blog post that Gen 3-Alpha is “trained jointly on videos and images,” and its training was “a collaborative effort from a cross-disciplinary team of research scientists, engineers, and artists” focused on interpreting “a wide range of styles and cinematic terminology.”

VentureBeat notes that “specific data sets have not yet been disclosed — following a trend of most other leading AI media generators who also don’t disclose precisely what data their models were trained on, and if any was procured through paid licensing deals or just scraped on the web.” Germanidis expands on Gen-3 in an interview with VentureBeat.

TechCrunch writes that “a major unsolved problem with video-generating models is control — that is, getting a model to generate consistent video aligned with a creator’s artistic intentions.”

This is due to the fact that things taken for granted in traditional filmmaking, “like choosing a color in a character’s clothing, require workarounds with generative models because each shot is created independently of the others.” Sometimes, TechCrunch adds, “not even workarounds do the trick — leaving extensive manual work for editors.”

So, while generative video still has a ways to go, Tom’s Guide calls Gen-3 Alpha “a huge step forward” and says it “could be one of the best AI video generators yet.”