Modern image generators can turn out pictures that look right at a glance while quietly getting the identity wrong. A new arXiv paper, "LEVI: Benchmarking Visual Identity Fidelity with Verifiable Low-Entropy Constructs," takes on that problem by proposing a benchmark that treats exact identity as a testable property rather than a matter of visual similarity.

LEVI is built around verifiable low-entropy constructs—test items designed so that the correct identity can be checked unambiguously. The authors argue that evaluating identity failures requires separating violations of the requested identity from permissible variation, a distinction that is easy to blur when judging images by eye.

The abstract is brief, so the full methodology and results are not yet available from the announcement. The contribution, as described, is a framework for measuring whether a model has actually delivered the identity it was asked to produce, not just an image that looks like it might have.