A new experiment shows what happens when you chain open-source AI models into a continuous broadcast. The AI WebTV project pipes text-to-video generations from Zeroscope V2 through a post-processing stack that includes FILM frame interpolation and MusicGen-generated audio, then pushes the result to a live RTMP stream. The entire pipeline runs on NodeJS and TypeScript, with model inference handled by duplicated Hugging Face Spaces using the @gradio/client npm package.
The architecture is straightforward but clever. A human writes a base theme, ChatGPT expands it into individual shot prompts, and those prompts feed Zeroscope’s two-stage generation: a first pass at 576x320 resolution, then an optional upscale to 1024x576 using Zeroscope XL. The article is explicit about one gotcha — you must use the same prompt for both generation and upscaling, or the chain breaks.
The post-processing layer is where this gets interesting. Each upscaled clip runs through FILM (Frame Interpolation for Large Motion) to smooth out the choppiness that plagues most text-to-video output. MusicGen then layers in a soundtrack. The result is a playlist of mp4 files that FFmpeg concatenates and streams as FLV to an RTMP server — node-media-server in this case, though the author notes you could point it at Twitch’s RTMP entrypoints instead.
The creator is refreshingly honest about the limitations. The video sequences are deliberately short, and the WebTV is positioned as a tech demo or showreel rather than an actual show with art direction. That’s the right framing. Zeroscope’s output is impressive for open-source, but it’s not ready for narrative television. What it does prove is that the infrastructure for fully AI-generated video broadcasting is now something a single developer can assemble from public models and a weekend of TypeScript.
The fragility of relying on public Hugging Face Spaces is worth noting — the author flags that Spaces can become overcrowded and paused at any time, recommending you duplicate them under your own account. That’s a real constraint for anyone thinking about production use. But as a proof of concept, the AI WebTV demonstrates the full arc: prompt generation, video synthesis, upscaling, interpolation, audio generation, and live streaming — all open-source, all automated.
