Runway adds live moderation for real-time video
Runway has built a moderation layer that checks streamed video frames as they are generated and halts the stream in under half a second, ahead of real-time models it plans to release.
Runway described a new moderation system for streaming video generation. The company says it will need the system for the "increasingly powerful real-time models" it plans to release in the coming months.
Runway's existing pipeline runs its safety classifiers on a prompt and any input media, generates the output, and then checks that output before the user sees it. Streaming breaks that order, because the user watches the video while it is being made. Under the new design, inputs are still screened, but the check has to run faster. Once a stream starts, its frames are sent to moderation in parallel with playback, and the stream stops if a frame is flagged. A final check of the full video runs before it can be downloaded or shared, because a frame that looks fine on its own can be a problem in context.
For the frame checks, Runway uses Zentropi's CoPE-B, a LoRA adapter on Gemma-4-26B-A4B-it. That Mixture-of-Experts model has 25.2B total parameters but only 3.8B active per forward pass. Runway says each check averages under 0.5 seconds.
Runway says plainly that the system has gaps. In testing, the most problematic content was blocked at input review. But edge-case prompts sometimes produced brief "glimpses of prohibited content," including adult female nipples, before the stream was cut.
The real-time layer sits on top of Runway's existing safeguards: training-data filtering and post-training, red teaming focused on real-time risks, CSAM hash matching and classifiers, automatic account suspensions, and C2PA provenance on real-time outputs. Runway did not give release dates for the real-time models or any accuracy figures for the classifier.