Phantom: Subject-Consistent Video Generation via Cross-Modal Alignment

Phantom by ByteDance Research is a subject-consistent video generation model on Wan2.1. Preserves subject identity across frames using cross-modal alignment in 14B and 1.3B.