Anastasis Germanidis

Anastasis Germanidis is the co-founder and co-CEO of Runway, where he has helped develop generative video models and creative tools. Formerly its CTO, he became co-CEO alongside Cristóbal Valenzuela in February 2026. His work connects image and video generation with a longer-running interest in interactive art and simulation.[1][2]

Germanidis describes the first version of Runway as a way to put open machine learning models into artists’ hands. The company then built research and product capabilities of its own, including Green Screen for separating subjects from a scene and the Gen-2 video model. Filmmakers’ requests for camera control helped shape the company’s emphasis on steerable generation.[1][3]

In his account of Gen-2, Germanidis traces Runway’s work from latent diffusion to generating motion. He later introduced Gen-3 Alpha, a model trained on both images and video with temporally detailed captions for finer control.[3][4]

Germanidis’s current research direction extends video generation into interactive world models. In his Latent Space interview, he explains how a model can generate a few frames at a time in response to actions, and how reducing diffusion steps can lower latency—potentially at a cost to quality.[1]

The approach has a difficult failure mode: each generated frame becomes input for the next, allowing small errors to accumulate. For Germanidis, another test is whether a model produces realistic outcomes for different possible actions, not merely a convincing-looking continuation. He uses football footage as an example: training data may contain far more successful goals than missed attempts.[1]

He also sees opportunities to combine a reasoning model that plans a scene with a diffusion model that renders it. These are research directions, with simulation fidelity and real-time response still central challenges.[1]

Before Runway, Germanidis worked across backend and machine learning engineering while maintaining an interactive art practice. In one gallery project, software assigned visitors invented identities and coordinated encounters between them; an early collaboration with a Runway co-founder used an image-to-image model to turn semantic scene maps into surreal street imagery.[1]