🚀 Microsoft Research has introduced Mirage — a new Latent Spatial Memory architecture for world video models.
This technology allows 3D scene information to be stored directly within the diffusion latent representation space, bypassing the heavy rendering cycle into RGB pixels. This increases generation speed by 10.57x and reduces memory consumption for the 3D cache by 55x.
🌍 Mirage addresses a major bottleneck in video models: the complexity of maintaining spatial consistency. Moving to latent token management makes creating stable video worlds significantly cheaper and faster.
👤 This is a step toward creating fast AI video generators capable of building complex 3D spaces without "hallucinations," which is critical for simulations and VR.
Source 1: https://microsoft.github.io/LatentSpatialMemory/ Source 2: https://github.com/microsoft/LatentSpatialMemory