Decart Unveils MirageLSD, a Real-Time AI Engine Reshaping the Future of Video Streaming
Israeli startup Decart has introduced MirageLSD, a groundbreaking Live-Stream Diffusion model capable of transforming video in real time with latency under 40 milliseconds. Unlike traditional video-generation systems that work in slow batch mode, MirageLSD enables continuous, stylistically transformed streaming at 24 FPS without quality degradation.This marks one of the most significant advances in AI video processing to date, positioning Decart as an unexpected challenger to industry giants such as OpenAI, Google DeepMind, and Runway. While existing diffusion-based tools focus on generating short, pre-rendered clips, MirageLSD redefines the entire paradigm by enabling real-time visual transformation of arbitrarily long streams.
A fundamentally different approach to video generation
Most state-of-the-art video models—Sora, Veo, Lumiere and others—operate offline. They ingest a prompt, generate a few seconds of footage, and require substantial GPU compute. This limits their use in applications such as live streaming, gaming, video conferencing, AR pipelines, or interactive tools.MirageLSD, by contrast, works frame by frame, performing diffusion-based transformation in a streaming loop. Each frame is processed individually but coherently, maintaining temporal consistency and avoiding the visual collapse that long-form diffusion models often experience. According to Decart, the system can sustain 24 frames per second at 768×432 resolution, making it suitable for real-time interactive environments.
Under the hood: custom CUDA kernels and extreme optimization
The team at Decart engineered MirageLSD around NVIDIA Hopper GPUs, using custom CUDA kernels hand-optimized for the model’s diffusion steps. These optimizations, combined with architectural pruning and a method the company calls shortcut distillation, reduce computational overhead dramatically.As a result, MirageLSD achieves a sixteenfold speedup over comparable systems. This is not merely a marginal improvement—it is the difference between a theoretical demo and a fully deployable real-time product. Sub-40 ms per-frame latency is low enough to support live human interaction without perceptible delay.
Economics: thousands of times cheaper than Veo
Perhaps the most disruptive factor is cost. Decart reports that generating one hour of transformed video with MirageLSD costs less than $0.25 in compute. By comparison, running Google’s Veo 3 at equivalent quality would cost roughly $1400 per hour. This staggering 5,000× difference could reshape the competitive landscape of AI video tools.If accurate at large scale, MirageLSD’s economics open the door for entire industries—gaming, VTubing, live entertainment, telepresence, education, and even security applications—to adopt real-time diffusion without prohibitive costs.
Applications: streaming, meetings, gaming, and beyond
Decart emphasizes three main use cases for MirageLSD. First is streaming: creators can apply cinematic or stylized transformations in real time while broadcasting. Second is video conferencing, where AI-enhanced identity, background, or stylistic filters can be applied seamlessly without latency spikes. Third is gaming, where real-time visual style transfer could enable dynamic worlds, graphic modes, or interactive effects not possible with traditional engines.Unlike post-processing tools, MirageLSD is designed for continuous pipelines. It supports arbitrary stream length and does not accumulate noise or temporal drift. This makes it compatible with long-duration content such as live shows, virtual events, esports, or multi-hour gameplay sessions.
Positioning against Sora, Veo, and Runway
Although Decart does not directly frame MirageLSD as a competitor to Sora or Veo, the implications are unavoidable. Those models excel in cinematic offline generation, but neither is practical for real-time transformation. Moreover, their high compute requirements make them inaccessible for large-scale continuous workflows.MirageLSD occupies a different niche—one that previously had no real technical solution. It aligns more closely with the needs of creators who require responsiveness, interactivity, and low-latency output. As real-time virtual production grows, this niche could become one of the fastest-expanding segments of the AI video market.
From laboratory to product: Mirage enters the market
MirageLSD is already available through a web platform, with native applications for iOS and Android currently in development. Early demonstrations show smooth, coherent stylistic transformations applied to live webcam feeds and pre-recorded streams. The platform supports a growing library of styles, blending artistic, cinematic, and abstract looks.If mobile performance reaches acceptable levels, MirageLSD could become a mainstream tool for creators, streamers, and even casual users, similar to how real-time filters transformed social media over the past decade—but with vastly more sophisticated capabilities.
Why MirageLSD matters
The debut of MirageLSD marks a turning point in AI video. For the first time, high-quality diffusion models are no longer limited to static workflows or short clips. Continuous generation with true real-time constraints opens new domains of use and sets the stage for AI-native entertainment formats.Decart’s achievement also underscores the importance of software-level innovation. Rather than relying on ever-larger models or escalating compute budgets, MirageLSD demonstrates how architectural efficiency, GPU-level engineering, and targeted distillation can unlock breakthroughs that brute-force scaling cannot achieve.
As the AI video ecosystem accelerates, real-time models like MirageLSD may ultimately define the next decade of digital content creation. For now, Decart has secured a first-mover advantage in a field poised for rapid growth.
Editorial Team - CoinBotLab
Source: Decart
🔵 Bitcoin Mix — Anonymous BTC Mixing Since 2017
🌐 Official Website
🧅 TOR Mirror
✉️ [email protected]
No logs • SegWit/bech32 • Instant payouts • Dynamic fees
TOR access is recommended for maximum anonymity.
🌐 Official Website
🧅 TOR Mirror
✉️ [email protected]
No logs • SegWit/bech32 • Instant payouts • Dynamic fees
TOR access is recommended for maximum anonymity.