Runway staff describe how Worlds 2 streams generated video live
Three people from the company told Latent Space that a distillation step cut the model to four denoising passes, which is what makes it playable.
Runway's GWM Worlds 2 research preview streams 720p video at 24 frames a second with audio at 48,000 Hz, and staff told Latent Space the model reaches that speed through distillation. Anastasis Germanidis, the company's co-founder and co-chief executive, said the work cut diffusion from around 50 denoising steps to four.
Germanidis said the reduction comes "with some quality loss but potentially comparable results", which is the company's own reading of the trade-off. Nobody outside Runway has measured it. Richard MacManus wrote the piece, which went up on Friday and also quotes Kamil Sindi, the chief technology officer, and Robin Kahlow, a principal research scientist.
Runway calls the input format WorldPrompt. Kahlow said it lets a user fix parts of a simulated environment, including the first frame, and then fire timestamped events into it while the video runs. He described the aim as "very detailed control over everything in the scene".
Kahlow put the engineering as two problems: "making the model not generate a whole clip at once", and "making the generation fast, so you can play it in real time". The sequence he described is fine-tuning to the WorldPrompt format, post-training the model to generate autoregressively, then distilling it down for speed.
The limits are in the interview too. Germanidis said error accumulation is the main constraint, because the model feeds generated frames back in and small errors compound, which holds continuous interaction to a few minutes. A character model can run about 30 minutes. Kahlow said movement "works quite reliably" but called this a research preview with flaws that depend on how hard the action is.