Israeli AI startup Decart has announced that it is being described as the world's first real-time, autoraffic video-to-video model. The new system, named Mirage, is designed to maintain movement, structure and frame rate while converting live video streams into a variety of visual styles.
Unlike previous AI images and video generators that struggled with visual glitches, Mirage generates each video frame in turn, conditioning it previously and adjusts a simple text prompt that describes the desired style. This autoregressive approach aims to keep movement consistent and stabilise vision. This is an area where other diffusion-based models are often lacking.
Decart says Mirage can run at 20 frames per second at standard web resolution. The end-to-end processing delay is approximately 100 milliseconds per frame. This performance allows for live streaming, interactive video calls and other real-time applications to run when hardware and cloud infrastructure are still at their pace. The company plans to expand to full HD and 4K resolution as performance improves.
Mirage is available along with an early version of the user platform. On the user platform, you can upload or record videos, enter style prompts, and watch the model as it immediately applies changes. Unlike traditional video editing, which requires individual production stages, Mirage aims to make live video a more flexible format that can be adapted on the fly.
Dean Leitersdorf, founder and CEO of DeCart, said the company wants to make the video a more responsive media. “The content is no longer modified, but it's adaptive,” he said in a statement, adding that Mirage can give viewers more ways to affect the appearance of the video.
Founded in 2023 by Dr. Dean Leitersdorf and Moshe Shalev, Decart raised $53 million from benchmarks, Sequoia Capital, Zeev Ventures and others.
