An Evaluation Framework for Generating Multi-View Images of a Person in a Scene

In response to the bottleneck in spatial consistent camera angle changes of existing generative image editing models and the lack of dedicated training data, researchers proposed the Head Scene Rotation Difference (HSRD) metric to quantitatively evaluate camera movements around characters. The study attempted to use synthetic data generated by various advanced image editing models and found that hallucinations where character heads turned in inconsistent ways with the background often occurred in the output. The proposed HSRD metric solves this problem by decoupling camera movement from local head pose operations. Experiments show that HSRD provides the necessary evaluation pipeline for constructing high-quality multi-character scene multi-view synthesis datasets.