Generative video is improving along an efficiency axis, with some models producing continuous video in real time at a fraction of the cost of slower generation.
2
Real-time video makes applications such as webcam-based magic mirrors, visual AI interfaces, and live content steering possible.
3
The difficult engineering work is serving these models globally, synchronizing controls frame by frame, and connecting multiple models into streaming pipelines.
Summary
Keegan McCallum argues that generative video progress should be measured by efficiency and generation length as well as visual quality. His example is Helios, a model served by uRun that can generate continuous video in real time at roughly the quality of frontier models from the previous year. He says ten dollars can now buy about three hours of continuous generated video, while fifty dollars can buy fifteen hours. That price and speed change what people can build. A webcam could show someone different clothes, a car, or a haircut. Visual interfaces could make AI easier to use for people who do not think primarily in text. Creators could steer a shot while it is being generated instead of repeatedly paying for prompts and keyframes. McCallum says the remaining challenge is infrastructure: globally placed GPUs, WebRTC networking, synchronized controls, and pipelines that combine several models into one stream.
Generative video is improving along an efficiency axis
McCallum shifts attention from the familiar quality race to efficiency and long-horizon generation. He shows Helios, which uRun serves, generating a long continuous video while another stream produces clips faster than a viewer can consume them. He says Helios is a distillation of Wan 2.1 14B. In a comparison, the real-time result is close to the slower result in quality, while the slower video has better motion and costs about one hundred times more to produce. The point is that a slightly weaker clip can have a different practical value when it arrives continuously and cheaply.
Many models now support real-time or long-horizon generation
McCallum describes several model techniques behind the recent progress. World models can maintain consistency over longer horizons and allow fine-grained control of the camera and viewport. Avatar models can generate interactive characters, while video-to-video models transform a live view almost like a magic mirror. He says at least 40 models with real-time capabilities or long-horizon generation capabilities were released during the year. This makes Helios part of a wider movement rather than an isolated result.
Low generation costs make continuous visual interaction practical
McCallum compares the cost of video generation with the audience's spending on coding tokens. He says ten dollars can provide three hours of continuously generated video with most of these models, while fifty dollars can provide fifteen hours. That means a person could interact with an AI through a visual medium for much longer than the short clips associated with earlier generation systems. The change is economic as well as technical. Continuous use no longer requires treating every generated second as an expensive one-off production.
Real-time video can turn a webcam into a visual interface
One example is a magic mirror built from a webcam and an open-ended video model. A user could ask to see themselves in a particular outfit, inside a car they like, or with a haircut they are considering. Because the model transforms the live webcam view, the interaction can respond to changing requests rather than producing a fixed image. McCallum also connects this to accessibility. AI tools rely heavily on reading and text, while some people think visually or learn better through visual interaction. Live video could give those users another way to work with AI.
Creators can steer a shot while it is being generated
McCallum contrasts real-time steering with the current content-creation workflow, which he describes as a slot machine. A creator sets a prompt and perhaps some keyframes, spends about ten dollars per minute, and hopes to get the desired shot. A real-time model can be steered in under a second during generation, letting the creator control the content more granularly. The same pattern could let someone watch an agent work and see what it is producing as it happens. McCallum says Google Gemini Omni can render these results into a more full-fidelity clip.
Serving interactive video requires synchronized streaming infrastructure
Building these applications requires more than choosing a model. A global audience may require GPUs distributed around the world, along with decisions about where users connect and which GPUs serve them. McCallum names WebRTC, ICE, and TURN as part of the networking setup. More advanced applications may connect several models in continuous streaming workflows. The system has to keep user controls synchronized with every frame while maintaining a smooth stream.
uRun is building programmable tools for model pipelines
McCallum describes uRun's proposed developer experience as a React component that can be dropped into an application to provide interactive video with any model. Behind it is a programmable Python runtime for building asynchronous pipelines. Developers could use it for avatar models, video-to-video transformations, and other combinations they design. He also describes a CLI and an MCP server for agents and says that, in 2026, developers need software factories and ways for agents to interact with these systems. His conclusion is that the models exist, and the frontier is now in how they are served.
"And so, our idea is what if there was just a React component that you could drop into your application to make it easy to provide video interactively inside your applications with any model."06:45
Who should watch
You are building a video or webcam product and need to understand what real-time generation changes in the interaction model.
Your AI interface relies heavily on text, and you are exploring visual interaction for users who think or learn more easily through images and video.
You are operating generative models in production and need to plan for GPU placement, WebRTC connectivity, model orchestration, and frame-level synchronization.
