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.

P

Phantom

Wan2.1Video GenerationSubject-ConsistentOpen Source

Subject-consistent video generation model by ByteDance Research, built on Wan2.1. Uses cross-modal alignment to preserve subject identity and appearance throughout generated video sequences. Available in 14B and 1.3B parameter variants.

Original RepoHF (Kijai)
DeveloperByteDance Research (Intelligent Creation Team)
ArchitectureWan2.1-based with cross-modal alignment
Model Sizes14B (fp16/fp8), 1.3B (fp16/fp32)
CapabilitiesSubject-to-video generation, identity preservation

Overview

Phantom is a subject-consistent video generation framework from ByteDance Research that adapts the Phantom architecture to the Wan2.1 video generation model. It enables generating videos where a specific subject's identity and appearance are consistently maintained across all frames. The model uses cross-modal alignment techniques to ensure that the generated video faithfully represents the input subject while producing smooth, high-quality motion.

Phantom is available in two sizes: a full 14B parameter model for high-quality output and a lightweight 1.3B parameter variant for faster inference on consumer hardware.

ComfyUI Integration

Phantom-Wan is supported in ComfyUI via the Kijai WanVideoWrapper custom node. Download the appropriate model files from the Kijai repository and place them in your ComfyUI models directory.

Available Sizes

SizePrecisionUse Case
14Bfp16, fp8High-quality subject-consistent video generation
1.3Bfp16, fp32Lightweight inference on consumer GPUs

Resources

ResourceLink
GitHub Repositorygithub.com/Phantom-video/Phantom
Hugging Face (Original)bytedance-research/Phantom
Hugging Face (Kijai)Kijai/WanVideo_comfy

Guides and workflows related to this model series.

No articles found.