Mirage Stores Video World Model Memory in Latent Space, Skipping Pixel Reconstruction

Mirage replaces RGB-space point cloud memory in video world models with a persistent 3D cache operating directly in diffusion latent space, using depth-guided back-projection to lift latent tokens into 3D. This avoids repeated VAE encoding/decoding cycles, reducing compute and preserving richer learned features compared to pixel-space round-trips.