MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing
Multi-person image editors still fail at body contact in ways current VLM judging can miss.
MPIE-Bench tests 2,500 video-mined editing cases across 405 scenes, 14 interaction types, and four contact-density levels. The paper adds MPIE-Eval, using multi-person mesh reconstruction to score whether bodies are anatomically complete and whether contact geometry matches the prompt. Across ten editors, the best mesh Anatomy score is 0.65 and the best mesh Interaction score is 0.72, while VLM checklist scores exceed 0.95 on the same images. A five-rater study found the mesh-based axes tracked human judgment better than a zero-shot VLM judge. HF Daily Papers' note
MPIE-Bench tests 2,500 video-mined editing cases across 405 scenes, 14 interaction types, and four contact-density levels. The paper adds MPIE-Eval, using multi-person mesh reconstruction to score whether bodies are anatomically complete and whether contact geometry matches the prompt. Across ten editors, the best mesh Anatomy score is 0.65 and the best mesh Interaction score is 0.72, while VLM checklist scores exceed 0.95 on the same images. A five-rater study found the mesh-based axes tracked human judgment better than a zero-shot VLM judge. HF Daily Papers' note
score 4