Decart’s Mirage is presented as a new class of real-time video-to-video Artificial Intelligence that transforms live streams, including webcams and games, with approximately 40 milliseconds of delay and a target of 16 milliseconds in the next version. The system operates frame-by-frame in what the team calls “next frame prediction,” rather than generating entire clips at once. The team spent six months addressing an error accumulation problem, writing assembly code for GPUs to prevent degradation into static colors.
Early demonstrations and use cases highlight how Mirage can alter live content on the fly. Examples include a Minecraft mod that transforms the entire game on command, live conversions of GTA V into a winter environment, and a consumer-facing selfie app called Delulu for real-time transformations. The article recounts a demo on Dan Shipper’s podcast in which Dean Leitersdorf used Mirage to turn himself into a wizard and convert everyday objects into interactive props, illustrating the platform’s creative potential.
The piece contrasts Decart’s instant-transformation approach with recent work on world models and 3D reconstruction. Tencent released HunyuanWorld-Voyager, an open-source world model that generates camera-controlled 3D environments, exports directly to 3D formats, and ranks first on Stanford’s WorldScore leaderboard. Tencent and Google are focused on 3D reconstruction for VR and simulations, while Decart is betting on instantly transforming existing content. Dean argues for a hybrid approach: deterministic code for game physics and rules and Artificial Intelligence for creative visuals, enabling simple 3D logic to look like AAA titles via real-time skins.
The broader implications suggested include reshaping video calls, content creation, and perceptions of digital reality as real and generated visuals blur every 40 milliseconds. The newsletter also notes related industry moves: CoreWeave acquired OpenPipe, Apple is preparing a new Siri web-search and testing Gemini on Apple servers, and Mistral is reportedly nearing a valuation via a reported €2 billion round (valuation not stated). The author adds a speculative note that major platforms could acquire real-time transformation startups to accelerate metaverse and content initiatives.