Meta releases Composer-2: a 500B multimodal model with 3D object understanding

AI · 6 min read

Composer-2 represents Meta's next step in multimodal AI, with explicit architecture components for 3D pose estimation and mesh completion. The model inputs can include images, sparse point clouds, and textual instructions; outputs range from completed meshes to mapping-friendly texture atlases.

Meta demonstrated Composer-2 filling missing geometry in partial scans, generating UV-preserved textures, and suggesting retopology-aware simplifications for complex assets. These features are targeted at AR/VR creators and game artists who need rapid iterative content generation.

Composer-2 will be available through Meta's developer platform with tiered access for researchers, studios, and enterprise customers. Meta emphasized safety and IP safeguards, including provenance metadata embedded into generated assets to help track the origin and license constraints.