Report
PersistBench tests whether 4D models remember objects after they leave the camera's view
The researchers use 360° recordings to score object permanence, motion continuity and appearance preservation beyond the input view.
TLDR
The PersistBench team introduced a dataset and evaluation suite for testing visual memory in models that generate or reconstruct 3D scenes over time. Its 2,000 paired sequences use 360° recordings to show what happens to objects after they leave the model’s input view. The researchers say every evaluated model with comparable visible and invisible outputs performed worse once the target left view.
Combined views
36
1 Source, first seen ago
