Phantom

Phantom is ByteDance's unified framework for subject-consistent video generation, introduced in the paper 'Phantom: Subject-Consistent Video Generation via Cross-Modal Alignment' (arXiv, February 2025, with the follow-up Phantom-Data dataset paper in June 2025). Instead of bolting image conditioning onto a text-to-video model, it redesigns the joint text-image injection module and trains on text-image-video triplet data, so one or several reference images steer the output without the 'copy-paste' problem where the model just replays the reference frame. The project ships two checkpoints, Phantom-Wan-1.3B and Phantom-Wan-14B, both supporting 480p and 720p output.

The framework is aimed at character-consistent storytelling: a product mascot, a game character, or a person can be dropped into new scenes from a text prompt while keeping identity intact, including multi-subject scenes that older reference-driven methods tended to confuse. The authors report results that beat closed commercial subject-consistency solutions on their own evaluations, though no independent benchmark numbers exist yet. Code and weights are on GitHub and Hugging Face under Apache-2.0, and the repo was still being updated in October 2026 with the 14B Pro checkpoint and training code on the roadmap.

Author-reported: outperforms closed-source commercial subject-consistency solutions on subject-consistency and prompt-alignment evals (paper, June 2025); no independent published benchmarks