AI Finally Conquers Live Video: New Model Transforms Streams Infinitely

A new artificial intelligence model promises to revolutionize live video by transforming streams in real time, a feat that has long been a significant hurdle for the industry. GenAI startup Decart has launched MirageLSD, a system designed to overcome two of the most persistent problems in AI video generation: slow rendering speeds and the degradation of image quality over extended periods.[1] This development could unlock a new wave of interactive and customizable content for live streaming, gaming, video conferencing, and beyond.

The core challenge for real-time AI video has been the immense computational power required and the tendency for errors to accumulate in sequentially generated frames.[2] Traditional video diffusion models often take several seconds or even minutes to process short clips, and the visual quality can start to break down after just a few seconds.[3][1] MirageLSD, however, employs a different strategy. Instead of generating entire video sequences at once, it creates each frame individually and sequentially.[1] This "causal, autoregressive" approach allows the model to react instantly to changes in the live feed by using a window of recent frames, the current video input, and a user's text prompt to predict the next frame.[1][2] This process enables the system to transform video streams continuously at 20 frames per second with a resolution of 768x432, achieving a latency of under 40 milliseconds, a speed that is virtually imperceptible to the human eye.[2][3][4]

To maintain visual integrity and prevent the compounding errors that plague other models, Decart has implemented innovative training techniques.[2] One key method is "diffusion forcing," which trains the model to clean up noisy frames independently, without over-relying on the context of previous frames.[1][2] Another technique, called "history augmentation," involves intentionally fine-tuning the model on corrupted input frames, effectively teaching it to anticipate and correct artifacts.[4][5] This dual approach makes the model robust against the "drift" that typically causes quality to degrade in autoregressive generation, allowing for stable, long-duration video transformations.[4] The result is a system that can theoretically generate infinitely long videos without a drop in quality, a significant leap forward from the 20-30 second limit of many previous models.[4]

The potential applications for this technology are vast and varied. In the realm of entertainment and social media, users could transform their live streams on platforms like TikTok or Discord into different visual styles, such as anime or cyberpunk, simply by using a text prompt.[2] Video calls could become more immersive, turning a standard conversation into a shared adventure in a fantasy world.[3] The gaming industry also stands to benefit significantly; developers could use MirageLSD to rapidly create game worlds and handle all graphic effects, potentially building a functional game in as little as 30 minutes.[3] The technology could also be applied to pre-recorded media, allowing viewers to change the visual style of movies and TV shows as they watch.[2][6] Decart has already demonstrated the model's capabilities by transforming gameplay from titles like Minecraft and Call of Duty into entirely new visual environments.[2]

Founded in 2023, the Israeli startup Decart, led by CEO Dean Leitersdorf and Moshe Shalev, has rapidly secured significant funding, raising $53 million from prominent investors including Sequoia Capital and Benchmark.[7][6] MirageLSD is the company's second major model, following the viral "AI Minecraft" project, Oasis.[4] The company's ambition extends beyond just video, with plans to incorporate audio, emotions, and music to create a comprehensive platform for transforming all sensory inputs.[4] While the current model operates at a standard web resolution, Decart aims to support full HD and 4K resolutions in the future as hardware and optimization improve.[2][8] The launch of MirageLSD marks a pivotal moment in the evolution of AI video, shifting the paradigm from static, pre-rendered content to a dynamic and adaptive medium that can be shaped in real time by the user's imagination.[9][6]