What they did
The authors built WithEveryone to generate scenes containing as many as ten specified people. The system injects each reference identity as an individually addressable token, predicts where each person should appear, and renders this structured identity–layout plan as a visual condition for image generation.
For training, Layout-Grounded ID Loss uses annotated face regions to directly associate each planned location with its intended identity, rather than matching multiple predicted faces through potentially unstable embedding-based assignment. A second objective, ID Representation Forcing, requires the model to predict a representation for every requested identity before synthesis. Evaluation used an identity-disjoint benchmark, so test identities did not overlap with training identities.
Key findings
- WithEveryone achieved a target-context face-similarity score of 0.499, compared with 0.462 for GPT-Image-2.
- Its reported copy-paste artifact rate was 0.055, down from 0.169 for GPT-Image-2, suggesting that improved identity fidelity did not come from directly reproducing reference faces.
- The model included 97.3% of requested identities in generated images.
- Only 2.8% of generated identities were duplicates, indicating more reliable one-to-one binding between references and people in the scene.
Why it matters
Group generation requires more than preserving facial resemblance: a model must keep identities distinct, place each person correctly, and avoid omitting or duplicating people. The results indicate that explicitly planning identity-to-location assignments can make this coordination more reliable as the requested group grows.
Caveats
The abstract reports benchmark-level identity, coverage, duplication, and artifact metrics, but does not establish performance beyond groups of ten or describe trade-offs in image quality, prompt adherence, pose diversity, demographic robustness, or computational cost. The practical significance of the score differences also depends on the benchmark design and the reliability of its face-similarity and artifact measures.