What is it about?
AI image models make convincing pictures, but do they get the lighting right? We tested 16 models, released over the past four years, on more than 23 000 indoor photographs. In each, we hid the real grey sphere, a light probe long used in the film industry, and asked the model to generate a new one. Comparing the light's direction and colour on the real and generated spheres, we found no clear improvement over time.
Featured Image
Photo by Sean Sinclair on Unsplash
Why is it important?
As AI image models start to simulate the world for robots and self-driving cars, their lighting must be correct, not just convincing. Our test is a physical, large-scale measure of lighting accuracy that needs no human judges and can be run on each new model. Our findings show no clear improvement over four years; size, commercial status and type of tool do not predict accuracy; and all models favour light from the front. Our code and data are open at https://lvsn.github.io/SheddingLight/.
Perspectives
My earlier research on evaluating lighting estimation taught me that convincing lighting is not always physically correct, and this project grew out of that lesson. I love how simple the final test is: a plain grey sphere is enough to measure whether lighting is accurate, which the human brain is not very good at judging. I honestly expected newer models to light scenes more accurately as their images improved, and they did not. I hope this benchmark serves as a ruler rather than a ranking: if you work with image models, please test them and tell us what you find.
Justine Giroux
Universite Laval
Read the Original
This page is a summary of: Shedding Light: A Benchmark for Evaluating Lighting Understanding in Generative Image Models, ACM Transactions on Graphics, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3842579.
You can read the full text:
Contributors
The following have contributed to this page







