This benchmark is introduced in the InstanceGen (SIGGRAPH 2025) paper, and is aimed at estimating the performance of image-to-text models in generating complex prompts involving multiple objects, instance level attributes and spatial relationships.
The process used to evalute performance on the benchmark is explained in the paper, but we also plan to release the code that does this in the near future, stay tuned!
