Spatial information describes the relative positional or geometric arrangement of objects or scene parts. In text-to-image generation, it concerns whether objects remain coherently arranged across viewpoints, rather than merely matching text with pixel patterns and semantic associations. section 1.5.2
Transformers use positional encodings to represent spatial relationships, but these encodings do not reliably enforce geometric constraints across complex three-dimensional objects. Models also lack robust, consistent representations of depth, perspective, lighting, physics, and the three-dimensional world. section 1.5.2 section 13.2
Geometric inconsistency appears when generated spatial information fails to preserve:
- Consistent three-dimensional structure across objects and viewpoints. section 1.5.2
- Depth and perspective relationships. section 13.2
- Physical relationships governed by lighting and physics. section 1.5.2
Visible failure
A generated object may defy physical laws or appear distorted when viewed from a different angle, revealing inconsistent spatial information. section 1.5.2
