OpenVDN denoised 14.4 seconds of video in 11.23 seconds on eight B200s

A generator that produces video faster than it plays is a different object. It can respond to input as it runs, which makes possible interactive video, generated game environments, real-time avatars and live translation of visual content. The competition is not visual effects; it is the rendering pipeline.

Eight B200s is a substantial machine — this is not running on a laptop — but it is an amount of hardware a company can rent by the hour, which puts it within reach of product experimentation rather than only research.

The cost side is unresolved

A real-time claim on eight of the most capable accelerators available says little about cost per second of output, and the surrounding evidence suggests it is high.

Vals AI estimated this month that long agentic tasks can carry 10,000 times the energy footprint of a simple query, with some web app builds matching 2.5 hours of household electricity. Continuous video generation is a heavier workload than any of that, running indefinitely rather than for the length of a task.

An application that generates video in real time for many concurrent users is an application with a compute bill proportional to engagement — the opposite of the marginal-cost structure that made streaming media viable.

The wider generative video picture

This lands amid a busy period for the field. DreamX-Creator, published on 1 September, proposes native joint audio-video generation at 2K resolution. UniMate demonstrates text-to-animation that transfers across skeletons without per-rig retraining. Lucida converts video of cluttered rooms into editable 3D assets for robot simulators.

Each attacks a different constraint — audio-video synchronisation, rig generalisation, scene reconstruction, generation speed — and together they describe a field moving from producing clips to producing usable, controllable, interactive output.

The legal environment is moving more slowly. Google has been approaching Disney, Universal and Warner Bros. Discovery about licensing entertainment IP for AI tools, with no agreements confirmed. The EU's transparency rules, in force since 2 August, require disclosure when content is AI-generated — an obligation that content credentials struggle to satisfy in practice, since ordinary processing strips them.

Real-time generation makes that worse in a specific way. Content produced on demand, per viewer, never exists as a file that could carry a durable credential in the first place.

OpenVDN has not published output quality comparisons against slower systems, which is the obvious trade-off and the one the benchmark does not address.