Phantom

Keeping the same face across every frame is the problem Phantom tackles: ByteDance's subject-to-video framework generates clips in which reference people and objects stay consistent, trained through cross-modal alignment on text-image-video triplet data. It builds on the Wan2.1 architecture, shipping Phantom-Wan checkpoints at 1.3B and 14B parameters that render at 480p or 720p, with the base Wan2.1-T2V-1.3B weights downloaded separately. Single and multi-subject generation are both supported, so a person and a product from separate reference images can appear together in one clip. Inference runs on a single GPU or scales out with FSDP and xDiT parallelism, and installation is a git clone plus requirements.txt with PyTorch 2.4 or newer. The work was accepted to ICCV 2025, the team released the companion Phantom-Data training dataset in June 2025, and a 14B Pro variant is planned. Code is Apache-2.0 with about 1,500 GitHub stars, and the paper is arXiv 2502.11079. Advertising, e-commerce, and character-driven content are the obvious use cases.