This literature review analyzes text-to-3D generation methods with emphasis on disentangling geometry, appearance, material, and style factors. It traces evolution from single-object SDS-based methods to multi-object compositional approaches, proposing a taxonomy along controllability and visual guidance dimensions. The review identifies critical gaps in factor-level disentanglement and synthesizes findings from leading computer vision venues including CVPR and NeurIPS.
Key findings
Geometry-appearance separation has been achieved by methods like Fantasia3D, but fine-grained factor disentanglement remains largely unsolved.
Visual guidance has evolved from text prompts to incorporate reference images, sketches, layout constraints, and 3D self-priors.
Standardized evaluation benchmarks for quantifying disentanglement quality are currently lacking in the field.
Object-level disentanglement via Neural Category Fields represents recent progress toward compositional scene generation.
Limitations & open questions
Limited discussion of computational efficiency constraints and real-time generation capabilities.
Future research directions, while promising, lack concrete technical implementation pathways.
Evaluation of disentanglement relies primarily on qualitative assessment without standardized quantitative metrics.