Mirage by Decart

Mirage is built on Decart's proprietary Live Stream Diffusion (LSD) model — an autoregressive video generation architecture that constructs each output frame individually, conditioned on a window of recent frames and a user-provided text style prompt, rather than generating fixed short clips in batch. This frame-by-frame approach is what enables true real-time transformation: as each frame is generated it is immediately fed back into the next calculation, allowing the model to respond continuously to a live input stream. The underlying MirageLSD model solves two major problems that have constrained earlier AI video tools — slow rendering speed and error accumulation during long sequences. Decart addresses error accumulation through two training techniques: Diffusion Forcing, which adds noise to individual frames teaching the model to clean up artifacts without depending on earlier frames; and History Augmentation, which fine-tunes the model on corrupted input history so it learns to anticipate and correct for drift. The result is a system that generates video infinitely without quality collapse — a first in the AI video category. Mirage currently runs at 20 frames per second at 768×432 resolution with end-to-end latency around 100 milliseconds per frame, making it viable for live streaming, video conferencing, and interactive applications. Shortcut Distillation further reduces diffusion steps from several seconds per clip to under 40 milliseconds per frame by training smaller student models to match the trajectory of larger teacher models. The platform includes preset style themes — including Comic Book, K-POP, Egyptian Pyramids, and Steampunk — as well as support for custom text prompts. Users can upload a video, record directly in the app, or pipe in a live stream and apply transformations that preserve the original motion and structure while replacing the visual world entirely. See related. Mirage is used by live streamers applying branded visual layers or alternate world skins to gameplay in real time, by content creators on TikTok and YouTube generating viral styled video without post-production pipelines, by game studios building dynamic visual overlays and alternate scene aesthetics for interactive experiences, and by advertising teams generating multiple stylized versions of a single video asset on the fly during a live broadcast. Decart, founded in 2023 by Dean Leitersdorf and Moshe Shalev — alumni of Israel's elite military intelligence Unit 8200 — also announced a strategic research partnership with the Technion–Israel Institute of Technology alongside the Mirage launch. The MirageLSD API is available on Crusoe Cloud for enterprise developers, enabling conversion of a one-minute source video into a fully stylized output in one minute — up to 10x faster than alternative video AI solutions. Compared to stream diffusion tools and other live AI video approaches, MirageLSD produces more temporally stable output at faster response times, with previous auto-regressive models running more than 16x slower in response latency. Mirage represents a genuine category breakthrough — the shift from AI video as a rendering queue to AI video as a live, responsive medium. The current practical constraints are resolution (768×432 at launch, with Full HD and 4K on the roadmap), a limited window of past frames that can reduce long-term scene coherence for very extended sessions, and the fact that high-fidelity output benefits from sufficient cloud GPU infrastructure. For consumer apps, live streaming, and short-form social content, the current output quality is sufficient for TikTok-grade production. Teams building production broadcast or cinematic workflows will want to monitor the platform's roadmap for Full HD support and the upcoming facial consistency, voice control, and precise object control features Compare alternatives.