Phantom
ByteDance framework that keeps reference people and objects consistent in generated video.
About
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.
Should you use Phantom?
Pick it when
Pick Phantom when the same person or product from reference images must stay recognizable through a clip, such as ad or e-commerce shots combining a model and an item, and you want Apache-2.0 code with a 1.3B option for one GPU.
Look elsewhere when
Skip it if you also need audio-driven avatars or footage extension, which SkyReels-V3 bundles with reference-to-video. For prompt-only generation the base Wan 2.1 is simpler, and the 14B Pro variant is still only planned.
Alternatives to Phantom
- SkyReels-V3
Keeps 1 to 4 references stable at 720p and adds extension and talking avatars, but models start at 14B and the Skywork license replaces Apache-2.0.
- VACE
Also built on Wan2.1 and covers reference-to-video plus masked editing, inpainting, and pose transfer in one model, broader but less focused on identity.
- MiniMax H3
Reference-to-video from identity images with native stereo audio, but it is a 33B model with a four GPU reference deployment and a community license.
- Wan 2.1
The base model Phantom builds on; simpler for prompt or image driven clips, without the cross-modal training that holds a specific face across frames.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Video Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Added
- Jul 29, 2026