MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing

Researchers introduced MPIE-Bench and MPIE-Eval to test whether image editors can create physically plausible multi-person interactions.

MPIE-Bench contains 2,500 video-derived editing triplets spanning 405 scenes, 14 interaction types, and four contact-density levels.

MPIE-Eval uses mesh-based reconstruction metrics to assess anatomical correctness and interaction plausibility.

Across ten editors, the new metrics exposed failures such as limb fusion, missing limbs, and body interpenetration that VLM judges often missed.

Human ratings aligned more closely with the proposed evaluation, suggesting a better benchmark for real-world multi-person editing quality.