paper web signal

WithEveryone Reports Face Fidelity Gains Over GPT-Image-2 Baseline

TL;DR

  • The paper reports 97.3% identity coverage with a 2.8% duplication rate when placing up to ten specified people in a single image.
  • Face similarity rises from 0.462 for a GPT-Image-2 baseline to 0.499, with copy-paste artifacts falling from 0.169 to 0.055.
  • The framework injects each identity as an addressed token, predicts a structured identity-layout plan, then renders that plan as a visual condition.

A framework called WithEveryone can place up to ten specified people in a single generated image, its authors report, with identity fidelity metrics running ahead of a GPT-Image-2 baseline. Hengyuan Xu and seven co-authors posted the preprint to arXiv on August 20, 2026.

The reported metrics: 97.3% coverage of requested identities with a 2.8% duplication rate, face similarity rising from 0.462 to 0.499, and copy-paste artifacts falling from 0.169 to 0.055. The method "injects each selected identity as an addressed token, predicts a structured identity–layout plan, and renders the plan as a visual condition," the abstract states.

The published comparison runs only against GPT-Image-2, and the abstract reports aggregate scores without per-identity or per-image breakdowns.