Geometry describes the three-dimensional shape and spatial arrangement of objects. In text-to-image generation, geometric consistency means preserving those relationships across objects and viewpoints rather than merely matching 2D pixel and text patterns. section 1.5.2 section 13.2
Text-to-image models commonly struggle with geometry because they:
- Train primarily on diverse 2D images and lack an internal representation of the three-dimensional world. section 1.5.2
- Use positional encodings that help represent spatial relationships but do not reliably enforce complex 3D constraints. section 1.5.2
- Represent depth, perspective, lighting, and physical rules inconsistently, producing distortions or objects that defy physical laws. section 13.2
Improving geometric consistency requires mechanisms that explicitly model 3D structure, simulate physics, or otherwise enforce spatial coherence. These additions address consistency but are presented as ongoing challenges rather than a completed implementation. section 1.5.2 section 13.2
