Generating Accurate Group Images with Multiple Identities

Summary & Methodology Analysis

WithEveryone addresses the limitations of existing generative models that struggle with identity dilution and incorrect bindings when composing group shots. The architecture begins by loading target reference identities as tokens, which are then processed through an autoregressive mechanism to produce a Layout Chain of Thought. This structured plan explicitly defines person regions, face regions, and pose keypoints, ensuring the spatial arrangement is determined before any pixel generation occurs. To maintain identity fidelity, the system employs ID Representation Forcing, which aligns the model's internal representation with specific target identities before the synthesis stage. The entire process is mediated by a deterministic renderer that transforms these planned specifications into visual conditions for the generator. To ensure robust performance, the system uses a Layout-Grounded ID Loss function, which relies on layout annotations rather than standard embedding-based matching, providing a more reliable way to maintain correspondence between specific people and their planned positions in the scene. The model was evaluated against established systems including GPT-Image 2, the Nano Banana family, and the Seedream family across a benchmark of five to ten person scenarios. The results show a target-context identity similarity of 0.499 and a copy-paste artifact score of 0.055. Despite these gains, the paper notes that evaluating layout quality is challenging because multiple valid configurations can exist for underspecified prompts. Furthermore, the technology introduces significant safety risks, as it can be used to generate images of real people without consent or to impersonate individuals in unauthorized scenes.