Runway

Contribution to AI

Latent diffusion — running the denoising process in a compressed latent space rather than on pixels, which is what made image generation cheap enough for a consumer graphics card — was written by researchers at LMU Munich and Runway, and it was Runway that published the Stable Diffusion 1.5 weights in October 2022, at a moment when its collaborators would rather it had waited. The dispute settled little legally and a great deal practically: the checkpoint that most of the open image-generation ecosystem built on for the next two years came from a small company in New York. The video work followed the same logic. Gen-1, described in early 2023, conditioned a video diffusion model on the structure of an existing clip and the content of a prompt or reference image, so footage could be restyled without retraining for every input. Gen-2 dropped the structure input and generated clips from text alone, and was among the first such systems anyone could simply pay to use rather than read about. What changed was less the picture quality, which was poor, than the constituency: video synthesis acquired ordinary paying customers. Rotoscoping and background-removal tools had already put the company inside real post-production — a seven-person effects team used them on Everything Everywhere All at Once — and the 2024 Lionsgate agreement to train a model on a studio's own catalogue proposed licensing as the alternative to scraping.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).