Bernini

Bernini is ByteDance's open-source unified framework for video generation and editing. It pairs an MLLM-based semantic planner with a DiT-based renderer built on Wan2.2, and was released under Apache 2.0 in June 2026.

It handles text-to-image, image editing, text-to-video, instruction-based video editing, reference-guided editing such as garment swaps and video insertion, and subject-to-video that places a person or character into a new scene.

Most video models generate from scratch. Bernini splits the work: an MLLM planner decides the semantics, then the renderer paints pixels. That design gives it strong consistency on edits, untouched parts of a clip stay frozen, and strong identity preservation in subject-to-video.

Yes. The inference code and the renderer weights are public under Apache 2.0, on GitHub and Hugging Face. You can run it on your own hardware; a Hopper-class GPU is recommended, with multiple GPUs for video.

Identity preservation is its standout result. In ByteDance's subject-to-video evaluations it leads comparable systems on face similarity, holding a recognizable face as the subject moves, which makes it a fit for avatars, character work, and serialized content where the same face has to recur.

The default render setting is 480p at 16fps. The release prioritizes editing fidelity and consistency over maximum resolution, and higher settings are possible at greater compute cost.