Generating images with multiple specified people is unreliable because models struggle to keep identities distinct and correctly placed. WithEveryone solves this by planning the layout of people before generating the image, using addressed identity tokens and a Layout-Grounded ID Loss that supervises each face directly on its predicted region. This avoids the unstable embedding matching that causes identity mixing in existing methods.
On a benchmark with up to ten reference identities, WithEveryone improved face similarity from 0.462 (GPT-Image-2) to 0.499 and reduced copy-paste artifacts from 0.169 to 0.055. It covered 97.3% of requested identities with only 2.8% duplicates. These results show that explicit planning and region-based supervision allow identity-preserving generation to scale to larger groups without resorting to copying reference faces.