VACE

VACE is an AI video model for creators and developers who want to generate footage or edit existing clips locally. It combines reference-based generation and video editing, including changes limited to masked areas.

• Combined inputs: Guide a task with a text prompt and optional video, reference images and masks, combining generation and editing in the same task.
• Targeted edits: Move or replace objects, erase parts of a scene, animate subjects and extend footage beyond the original frame using references and masks.
• Video restyling: Change a clip's appearance while retaining content, scene structure, subject identity, posture or motion, including style transfer.
• Motion control: Use depth, pose and trajectory guidance to direct movement, or combine reference images with other controls for actions and face replacement.

The Python implementation supports Wan2.1 and LTX-Video and includes a Gradio interface for local inference. Its tested environment uses PyTorch and CUDA on NVIDIA GPU hardware. Generated videos and intermediate images and masks save locally.

The code is open source under Apache 2.0. Model licenses differ: the listed Wan2.1 models use Apache 2.0, while the LTX-Video model uses RAIL-M. Inputs can have any resolution, but each model has a recommended video size range for better results.