MirageLSD (Live-Stream Diffusion) is a real-time AI video model from Decart that transforms a live video feed — a webcam, a game, a screen — into a different visual world continuously, guided by a text prompt, with latency low enough to feel instant.
Most AI video generators work like a rendering farm: you submit a prompt, wait, and download a finished clip. MirageLSD belongs to a newer category. Instead of producing a fixed-length clip, it ingests a live video stream and re-generates every frame on the fly, so the output keeps running for as long as the input does. Decart introduced Mirage publicly in mid-2025 and described it as the first live-stream diffusion model — a diffusion system engineered specifically for infinite-length, real-time video-to-video transformation rather than offline clip generation.
That combination — video in, transformed video out, continuously, now — is why the term took off. "MirageLSD" is both a specific model name and shorthand for the broader shift toward real-time video diffusion: generative video that responds to you while it happens.
This is an independent explainer. LiveGen is not affiliated with Decart; product details below are based on public reporting and may change — check Decart's official materials for the current state.
Break the name apart and the concept is right there:
The hard part is doing both at once. A live model can't look ahead, can't re-render a bad frame, and can't pause to think. Public coverage of MirageLSD highlighted two engineering claims that matter: per-frame generation fast enough for interactive frame rates (latency reported on the order of tens of milliseconds per frame), and stability over unlimited duration — avoiding the drift and melt that older auto-regressive video models suffered after a few seconds of self-feeding.
MirageLSD is significant less as a single product and more as an existence proof. It demonstrated, publicly and at consumer scale, that diffusion-quality video transformation can run on a live feed — which unlocks a product category that batch generators structurally cannot touch:
If you're tracking the space, MirageLSD sits in the same real-time wave as the engine behind LiveGen: both treat video as something generated in the moment you're living it, not rendered after the fact. The umbrella concept is real-time interactive video.
You don't need a research demo to feel what live-stream diffusion means. LiveGen runs a real-time generative engine directly in a browser tab: open your camera, pick a mode, and watch the whole frame transform as you move — no download, no render queue.
Live-Stream Diffusion — a diffusion model architected to transform a continuous live video stream rather than generate fixed-length clips.
Decart, an AI research company that introduced the Mirage system publicly in 2025 and has continued to develop it. This page is an independent explainer and isn't affiliated with Decart.
Text-to-video models create a clip from scratch, offline, over seconds or minutes. MirageLSD transforms an existing live feed continuously, with per-frame latency low enough for interactive use — different input, different constraint, different product category.
Infinite-length generation is one of its headline claims: by conditioning on its own recent output with corrective training, it aims to avoid the quality drift that limited earlier models to a few seconds.
No. LiveGen runs its own real-time engine. Both belong to the same technology wave — live, generative, prompt-steerable video — which is exactly why the comparison is worth understanding.
In your browser: LiveGen's Style Morph and Freestyle apply a real-time generative transform to your live camera, free to start.
Open your camera and become any character — free to start.
Start generating free