Phantom-Data: Towards a General Subject-Consistent Video Generation Dataset

The paper introduces Phantom-Data, a large-scale cross-pair dataset designed to address the copy-paste problem in subject-to-video generation. It resides in the Cross-Modal Alignment Approaches leaf, which contains three papers total, including the original work. This leaf sits within Single-Subject Identity Preservation, a moderately populated branch with six sub-categories. The taxonomy reveals that cross-modal alignment is one of several competing strategies for identity preservation, alongside motion-explicit methods, training-free enhancement, and optimization-based approaches, suggesting a relatively active but not overcrowded research direction.

The taxonomy shows that neighboring leaves include Motion-Explicit Identity Methods (two papers) and Feature-Guided Consistency (two papers), both addressing identity preservation through different technical mechanisms. The broader Identity-Preserving Generation Methods branch encompasses pose-driven animation (five papers) and spatial-temporal decoupling (two papers), indicating that the field has diversified into multiple architectural paradigms. Phantom-Data's focus on dataset construction distinguishes it from sibling papers that primarily propose model architectures or alignment mechanisms, positioning it at the intersection of data-centric and model-centric approaches within the cross-modal alignment paradigm.

Among twenty candidates examined, both contributions show evidence of prior work. The dataset contribution examined ten candidates with one refutable match, while the three-stage pipeline examined ten candidates with one refutable match. The limited search scope (twenty total candidates from semantic search) means these statistics reflect a focused sample rather than exhaustive coverage. The dataset contribution appears to face more substantial prior work in the form of existing identity-consistent datasets, while the pipeline's novelty depends on the specific combination of detection, retrieval, and verification stages compared to existing data construction methods.

Based on the top-twenty semantic matches examined, the work appears to occupy a relatively sparse niche within cross-modal alignment approaches, specifically targeting dataset construction rather than model architecture. The taxonomy context suggests that while identity preservation is an active area, data-centric solutions remain less explored than model-centric ones. However, the limited search scope and presence of refutable candidates for both contributions indicate that the novelty assessment should be tempered by acknowledgment of existing work in large-scale identity-consistent data curation.