MOGAR: Multi-view optical and geometry adaptive refinement for high-fidelity 3D asset generation

Multi-modal 3D asset generation from sparse-view inputs remains a core challenge in both computer vision and graphics due to the inherent difficulties in modelling multi-view geometric consistency and recovering high-frequency appearance details. While Convolutional Neural Networks (CNNs) and Transformers have demonstrated impressive capabilities in 3D generation, they suffer from significant limitations–CNNs struggle with capturing long-range dependencies and global multi-view coherence, whereas Transformers incur quadratic computational complexity and often yield view-inconsistent or structurally ambiguous outputs. Fortunately, recent advancements in state space models, particularly the Mamba architecture, have shown remarkable potential by combining long-range dependency modelling with linear computational efficiency. However, the original Mamba is inherently constrained to unidirectional causal sequence modelling, making it suboptimal for high-dimensional visual scenarios. To address this, we propose MOGAR (Multi-View Optical and Geometry Adaptive Refinement), a novel and efficient multi-modal framework for 3D asset generation. MOGAR introduces the Multi-view Guided Selective Mamba (MvGSM) module as its core, enabling cross-directional and cross-scale alignment and integration of geometric and optical features. By synergistically combining feed-forward coarse asset generation, multi-view structural optimisation, optical attribute prediction, and cross-modal detail refinement via a UNet architecture, MOGAR achieves a tightly coupled reasoning pipeline from global structure to fine-grained details. We conduct extensive evaluations on several standard benchmarks, demonstrating that MOGAR consistently outperforms existing approaches in terms of geometric accuracy, rendering fidelity, and cross-view consistency, establishing a new paradigm for efficient and high-quality 3D asset generation under sparse input settings.