At TwitchCon San Diego, the Israeli startup Decart staged a live demonstration of something I’ve not seen before: real-time generative AI video. Partnering with Los Angeles–based Geenee AR, the company installed its system inside Geenee’s Magic Mirrors. Attendees could see themselves transformed on screen into different clothes, settings, or entire worlds, all rendered live, tracking them as they moved. The system’s latency measured under forty milliseconds, fast enough to maintain full 1080p motion without visible delay.
TwitchCon draws tens of thousands of streamers and fans who already work and play in real time. Their broadcasts unfold without editing or retakes. This audience, CEO Dean Leitersdorf said, is Decart’s proving ground. “They don’t storyboard,” he told me. “They improvise. Our AI needs to be just as fast.” Decart’s adds an AI layer to the camera software so streamers can prompt as they go.
DecartStream, as the streamer app is called, drops into OBS as a browser source, the same software most Twitch streamers use. It adds a new control layer where a creator types a prompt like “make the room pink,” “add a dog,” “change my outfit,” and the video adjusts midstream. Leitersdorf says this differs fundamentally from tools like Snap Camera, which rely on prebuilt filters. “We don’t apply overlays,” he said. “We regenerate the video itself.” The model reimagines the frame each instant while preserving motion and coherence, something Snap’s static effects cannot do.
Two other real-time applications of Decart technology were demoed as well: a mixed reality experience for the Quest 3, available on Sidequest, and a mobile version, available for free in the app stores. Each allows you to transform live images you see with real time AI voice prompts. A person walking past you could have have a cartoon head, or be walking on the moon.
The technology behind the system, Live Stream Diffusion, is a custom autoregressive model trained for continuous inference. Leitersdorf described it as “AI with a reflex.” Decart’s system is both lighter and cheap enough to do this because it was built from the ground up for speed. “We write below CUDA (Nvidia’s proprietary software layer), down to GPU assembly,” Leitersdorf said. “That’s how we keep latency under forty milliseconds.”
Founded in 2023, Decart has raised about $150 million from Sequoia, Benchmark, and other venture firms. The company employs roughly eighty people in Tel Aviv, San Francisco, New York, and Switzerland. Leitersdorf, who completed his PhD in distributed computing at the Technion at 23, says Decart’s focus is still research. “We’re building infrastructure for live creation,” he said. “Revenue comes later.”
Decart’s first year of income came from licensing its GPU optimization layer to labs and chip providers, work that funds its new products. In addition, they provide an API to partners. By driving down compute costs, Leitersdorf said, the company can afford to distribute its creative tools free to Twitch streamers while it grows a developer and streamer base. He estimates only a small fraction of its $150 million has been spent so far. “We burn less than ten percent,” he said.
That strategy mirrors the Twitch ecosystem itself, which grew on free access and community participation. At TwitchCon, Leitersdorf observed how tightly knit that culture remains. “People here build relationships through their streams,” he said. “It’s a conversation, not a newsfeed.” He described the Twitch environment as an economy of attention where users, creators, and moderators co-create the experience. By giving them generative control of their own feeds, Decart wants to make the medium more responsive to both sides of the screen.
Twitch streamers who tested DecartStream before the event began posting clips that quickly spread on X and TikTok. The company says thousands of developers worldwide are now experimenting with its API for games, virtual production, and interactive shows. Leitersdorf believes the same infrastructure could support multiplayer worlds where the environment reacts to dialogue or viewer input.
At TwitchCon, surrounded by creators who already live inside their broadcasts, Decart’s system fit naturally. Its AI doesn’t wait for an edit. It works inside the moment, adjusting faster than the human performing for the camera. In a community built on constant live exchange, Decart’s pitch was simple: this is AI that moves at the speed of Twitch.
