Geometric consistency is the preservation of coherent spatial structure, proportions, and object geometry across objects, scenes, and viewpoints. Text-to-image models often fail because training on diverse 2D images does not provide an internal representation of 3D structure or the physical rules governing spatial relationships. section 1.5.2
The problem appears through several linked limitations:
- Positional encodings capture some spatial relationships but do not reliably enforce strict geometric constraints across complex 3D objects. section 1.5.2
- Models lack robust, consistent representations of depth and perspective and do not understand lighting or physics as humans do. section 13.2
- Failures can produce objects that defy physical laws or become distorted when viewed from particular angles. section 1.5.2
Improving geometric consistency requires mechanisms that explicitly model 3D structure, simulate physics, or otherwise enforce spatial coherence. These approaches address a limitation shared by transformer- and diffusion-based text-to-image systems, but their computational cost remains an important concern. section 13.2
A viewpoint failure
A generated object may look plausible from one angle yet appear distorted from another, revealing inconsistent geometry rather than merely imperfect visual detail. section 1.5.2
