Mira-Scene: Pixel-Aligned Layouts for Generative 3D Scene Reconstruction

arXiv cs.GR details Mira-Scene, a compositional 3D scene reconstruction framework aimed at fixing one of the messier parts of generative 3D: putting objects into a scene coherently. Instead of regressing sparse, unbounded pose variables, it recovers dense correspondences between visible pixels and canonical object space, then aligns those to scene-space geometry.

The core pieces are the Canonical Coordinate Map, or CCM, and a scene-space Point Cloud Map, or PCM. CCM gives each visible object pixel a bounded surface coordinate target, which makes training more stable and lets the system learn from object-level 3D data without needing scene-level layout annotations. PCM comes from monocular geometry estimation and helps recover object transforms through geometric alignment.

Mira-Scene also uses a multimodal diffusion transformer that jointly generates object geometry and CCMs, with separate expert streams but shared attention and positional encoding to keep layout and shape in sync. In experiments across indoor, outdoor, synthetic, and in-the-wild scenes, it reportedly beat strong baselines on layout accuracy, with relative gains of 39.8% in 3D-IoU and 16.5% in 2D-IoU over SAM3D.

For developers, the practical angle is clear: better scene composition from weaker supervision could make AI-assisted level dressing, asset placement, and reconstruction tools less brittle. The exact production cost and runtime profile still matter, but the method points toward more reliable object-level scene generation without hand-authored layout labels.