LucidLyra is a new framework that lets one input image turn into a full 3‑dimensional scene where geometry, appearance, language‑aligned labels, and object identities are all encoded in the same structure. The method fuses a frozen video‑generation model that produces multi‑view latent codes with pretrained 2‑D vision models to supply semantic and instance supervision. A feed‑forward Gaussian decoder maps these latents, camera rays, and image features into a jointly‑learned 3‑D Gaussian splatting scene. Evaluation on synthetic and real‑image benchmarks shows that LucidLyra outperforms all two‑stage baselines that first reconstruct and then segment. Its semantic mean intersection‑over‑union reaches 64.1 % on synthetic data and 49.8 % on real images, while instance tracking scores rise to 21.5 % and 18.5 %, respectively. Reconstruction quality (PSNR, LPIPS) stays on par with state‑of‑the‑art 3‑D reconstruction. The dynamic, monocular‑video extension retains good visual fidelity, though PSNR drops slightly compared to a reference 3‑D reconstruction, indicating a small reconstruction trade‑off. LucidLyra therefore offers a fast, annotation‑free pipeline to build queryable 3‑D worlds that can be rendered, segmented, or edited on demand.