Re:Cognize - A Framework for Open-Set Sequential Character Re-Identification

Character re-identification (Re-ID) is the foundation of every downstream task on long-form visual narratives such as manga and comics: without consistent character identity across pages, archives spanning hundreds of millions of pages stay opaque to search and reasoning. Existing benchmarks, however, evaluate Re-ID under a closed-set assumption, retrieval against a fixed gallery of known identities. That assumption collapses on a fresh volume, where new characters appear over time, the gallery is built online, and predictions must stay consistent across long temporal gaps. We introduce $\textbf{Re:Cognize}$, a framework that exposes this regime through four evaluation protocols spanning closed-set retrieval, few-shot retrieval with a fixed gallery, unsupervised online clustering, and pre-seeded online gallery growth, instantiated on large comic datasets (PopCharacters, Manga109, Re:Verse) across five Re-ID backbones from person, manga-native, and multimodal pre-training. The decomposition surfaces a striking property of the regime: across every backbone, a single seed image per identity recovers nearly the entire closed-set retrieval ceiling, while online accumulation overshoots that ceiling within a handful of seeds. Identity $\textit{maintenance}$, not gallery construction, is the practical deployment bottleneck, and identity $\textit{emergence}$ is the harder, separable sub-problem that closed-set evaluation has never made visible. To complete the suite, we further introduce $\textbf{MeCha}$, a memory-augmented baseline that confirms the protocols are sensitive enough to discriminate sequential-context-aware models from purely embedding-based ones.