Identify the effect of prompt wording on lighting understanding

Determine whether the wording of text prompts affects generated-probe quality or the lighting understanding of generative image models.

Background

The authors compare four prompts of varying complexity, from “two grey spheres” to a detailed instruction specifying sphere geometry, material, placement, and consistency with scene lighting. Although no statistically significant differences in light-direction accuracy are found, usable-probe yield varies across prompts. The available experiment therefore does not resolve whether prompt wording changes probe generation or the models’ underlying lighting understanding.

References

Therefore, it was not possible to determine whether the prompt affected the probe generation or the model's understanding of lighting.

— Shedding Light: A Benchmark for Evaluating Lighting Understanding in Generative Image Models  (2609.10787 - Giroux et al., 9 Sep 2026) in Appendix, Section “Further evaluation,” subsection “Prompt adherence”