WithEveryone is a framework for generating images containing up to ten specified people, combining addressed identity tokens with an explicit identity–layout plan rendered as a visual condition. The key training objective, Layout-Grounded ID Loss, supervises intended identities directly using annotated face regions, while ID Representation Forcing trains a prediction for each identity before image synthesis. This design is positioned as a way to scale identity-preserving generation to larger groups without relying on direct reference-face copying.1

Context

Identity-preserving generation typically degrades as the number of specified persons grows because the model must bind every reference to a distinct face and location while training losses rely on noisy embedding-based face matching. WithEveryone addresses this by predicting a structured plan of identities and locations, rendering the plan as a visual condition, and forcing each identity representation before synthesis. This is a shift from implicit binding via global embeddings to explicit layout grounding.1

Assessment

The reported gains are meaningful: face similarity increases from 0.462 to 0.499, copy-paste artifact rate drops from 0.169 to 0.055, and identity coverage reaches 97.3% with a duplicate rate of only 2.8%. These numbers suggest the method reduces the common failure of repeating or merging identities in group scenes. However, the evaluation is based on a single identity-disjoint benchmark, and the artifact rate is automatic, so definitive assessment requires full experimental details.1

Caveats

The document is an arXiv abstract and has not been peer-reviewed. The identity-disjoint benchmark's construction, dataset size, and exact protocol are not described. Without code or comparison statistics, the robustness of the improvements and the generalizability to other domains remain uncertain. Independent reproduction would increase confidence; evidence that the identity-layout plan is effective beyond the specific benchmark would also strengthen the claim.1