Decart Unveils MirageLSD: Real-Time AI Model for Transforming Video Streams

**Decart**, a startup in the realm of generative AI, has unveiled **MirageLSD**—an AI-based video model that transforms video in real time. This system aims to tackle two significant challenges faced by existing AI-driven video tools: slow rendering speeds and rapid degradation of image quality over time.

AI video models often operate slowly and typically can only produce short clips lasting from five to ten seconds before the image quality begins to wane. In contrast, **MirageLSD** adopts a different methodology. Rather than generating complete video sequences, the model creates each frame individually.

The system utilizes a window of recent frames, the current video input, and user requests to predict the next frame as it streams. Each new frame is immediately applied in the subsequent computational step, enabling the model to swiftly respond to changes in the live stream. This configuration allows continuous real-time video transformation at a rate of 20 frames per second and a resolution of 768 x 432, while maintaining low latency for interactive applications.

To ensure the video quality remains consistent during extended sessions, Decart employs two training methods. The first, known as **»forced diffusion,»** adds noise to each frame individually, training the model to clean the images without relying on prior frames. This approach helps prevent the accumulation of errors over time.

The second technique, called **»history extension,»** assumes that during training, the model encounters distorted or erroneous frames, and therefore it learns to identify and correct recurring mistakes instead of simply skipping them.

Decart has tailored MirageLSD specifically for Nvidia Hopper GPUs, optimizing the architecture to eliminate less essential parts of the model and enhance speed and efficiency. The team also implements **»label distillation,»** training smaller models to replicate the results of larger ones. According to their statements, this process boosts performance by a factor of 16, allowing each frame to be processed in under 40 milliseconds while keeping latency low enough that most viewers won’t notice any significant delays.

However, MirageLSD does have limitations. Currently, it only processes a limited number of previous frames, which may lead to inconsistencies when handling longer videos. Moreover, the model struggles with significant style changes and precise control over individual objects.

Alongside MirageLSD, Decart launched the **Mirage** platform, with the **[web version now available](https://mirage.decart.ai/)** and mobile applications for iOS and Android in development. This platform is designed for live streaming, video calls, and gaming. Decart plans to release regular updates throughout the summer, introducing features such as enhanced facial consistency, voice control, and more precise object manipulation.

This marks Decart’s second AI model, following their viral **[Oasis project in Minecraft](https://the-decoder.com/ai-generated-game-oasis-now-turns-images-into-playable-3d-worlds/)**. The development of MirageLSD took approximately six months. Other systems, like **[StreamDiT](https://the-decoder.com/ai-system-streamdit-generates-livestream-videos-from-text-at-16-fps-512p/)**, can achieve similar speeds—up to 16 frames per second—and also offer interactive features, but still lag behind leading models, such as **[Google’s Veo 3](https://blog.google/technology/ai/google-veo-generative-video-model/)**, in terms of image quality.

**Delegate some of your routine tasks with [BotHub](https://bothub.chat/?utm_source=contentmarketing&utm_medium=habr&utm_campaign=news&utm_content=DECARTS_LAUNCHES_MIRAGELSD_A_REAL-TIME_VIDEO_TRANSLATION_AI_MODEL)!** The service does not require a VPN for access, and you can use a Russian card. Follow this [link](https://bothub.chat/?invitedBy=m_aGCkuyTgqllHCK0dUc7) to receive 100,000 free tokens for your initial tasks and start working with neural networks right away!

Translation source of the news can be found [here](https://the-decoder.com/decart-launches-miragelsd-an-ai-model-that-transforms-live-video-feeds-in-real-time/).