ByteDance Releases Lynx for Identity-Preserving Video

ByteDance has released Lynx, a new open-source model focused on generating personalized video clips. Published under a permissive Apache 2.0 license, Lynx is designed to create short videos that preserve the identity of a subject from a single reference photo.

The model works by combining a detailed reference image of a person with a text prompt describing a desired action or scene. Built upon the Wan2.1-T2V-14B text-to-video foundation, Lynx uses this information to animate the subject, aiming for consistent facial features and identity throughout the generated clip.

While many text-to-video models can generate generic human figures, Lynx's specific focus on subject consistency is notable. This release provides researchers and developers with an open tool for exploring applications in personalized content creation and digital avatars. The model and code are available on Hugging Face for experimentation.