Decart has introduced MirageLSD, an AI video model designed for live video transformation rather than short, pre-rendered clips. The launch matters because many AI video systems still run into two practical barriers: slow output and visible quality decline after only brief sequences.
MirageLSD is built around a different workflow. Instead of trying to generate a full video in one pass, it produces the stream one frame at a time, using recent visual context, the current input, and the user’s prompt to decide what comes next.
Why live AI video is difficult
AI video models often work best in short bursts. The source article describes a common limit: many systems typically generate only five- to ten-second clips before image quality starts to break down. That is a serious constraint for livestreaming, video calls, or gaming, where the video does not pause while a model catches up.
The problem is not only speed. A live feed changes constantly. A person moves, the camera shifts, lighting changes, and new objects enter the frame. A model that cannot respond quickly will feel delayed. A model that compounds small errors will gradually lose visual stability.
MirageLSD’s central promise is to make AI video transformation continuous. The model is meant to keep producing transformed frames as the stream unfolds, rather than treating video as a finished sequence to be generated ahead of time.
How MirageLSD builds each frame
MirageLSD works by predicting individual frames. For each step, it uses a small set of recent frames, the active video input, and the user’s prompt. The next frame is then generated and immediately becomes part of the context for the following calculation.
This feedback loop lets the model react to changes in the incoming video. If the live feed changes, the model can incorporate that change into the next frame instead of waiting for a longer rendering process to complete.
According to the source article, this approach enables continuous real-time transformation at 20 frames per second and 768 x 432 resolution. Each frame is processed in under 40 milliseconds, which keeps the delay low enough that most viewers are not expected to notice major lag.
Those figures define the current practical shape of MirageLSD. It is not presented as a maximum-quality offline video generator. It is positioned as an interactive AI video system where responsiveness is the key feature.
How Decart tries to keep quality from falling apart
Longer AI-generated video can suffer when small mistakes persist from one frame to the next. MirageLSD addresses that risk with two training techniques named in the source: diffusion forcing and history augmentation.
Diffusion forcing adds noise separately to each frame during training. The model learns to clean up the image at the frame level, without depending too heavily on earlier frames. The goal is to reduce the chance that errors keep building as the stream continues.
History augmentation tackles a related problem. During training, the model is shown distorted or faulty frames. That teaches it to recognize recurring flaws and correct them, instead of carrying them forward into later frames.
Together, these methods are aimed at making the live stream more durable. In plain terms, Decart is trying to make MirageLSD recover from visual mistakes while it keeps running.
Speed depends on hardware-aware optimization
Decart has tuned MirageLSD for Nvidia Hopper GPUs. The source article says the company uses architecture-aware pruning, which removes less important parts of the model to improve speed and efficiency.
The team also uses shortcut distillation. In this process, smaller models are trained to reproduce the results of larger ones. Decart says this produces a 16x performance boost.
That optimization work is central to the product claim. Real-time AI video is not only a modeling problem; it is also a latency problem. For livestreaming, video calls, and gaming, the model has to finish work quickly enough for the interaction to feel immediate.
MirageLSD’s under 40 milliseconds per-frame processing target is therefore not a minor technical detail. It is what makes the system relevant for use cases where a visible delay would break the experience.
What the Mirage platform offers now
Decart has launched the Mirage platform alongside MirageLSD. A web version is already available, and mobile apps for iOS and Android are on the way.
The platform is aimed at three live-video settings:
- livestreaming
- video calls
- gaming
Decart also plans regular updates throughout the summer. The planned additions include improved facial consistency, voice control, and more precise object control.
Those planned updates point directly to the current limitations. MirageLSD only uses a small window of prior frames, so consistency can still decline over longer videos. The model also has trouble with major style changes and exact control over individual objects.
In other words, MirageLSD appears strongest where fast, continuous transformation matters most. It is less settled when the task requires strict identity preservation, large stylistic shifts, or highly specific object-level edits.
Where MirageLSD fits in AI video
MirageLSD is Decart’s second AI model, after the viral Minecraft project Oasis. The source article says the model took about six months to build.
It also places MirageLSD in a wider field of interactive AI video systems. Other systems such as StreamDiT can reach similar speeds, up to 16 frames per second, and also support interactive capabilities. However, the source notes that such systems still trail top models like Google’s Veo 3 in image quality.
That comparison frames the trade-off clearly. MirageLSD is not simply about producing the best-looking standalone clip. Its focus is live AI video transformation with low latency, continuous response, and a platform aimed at real-time use.
The larger implication is straightforward: AI video is moving from generated clips toward interactive feeds. MirageLSD shows one route to that shift, while also showing the hard parts that remain, including consistency, style control, and precise object handling.