Generación de video

Bernini

de ByteDance

ByteDance's open‑source model for instruction‑based video edits, identity held.

Funciones clave

Instruction-based video editing

Tell Bernini what to add, remove, or alter in a clip in plain language, and it makes just that change. The planner works out what the edit means before the renderer paints it.

Locked scene consistency

Everything you don't name stays frozen. Untouched regions hold across the whole clip with no flicker or drift, which is the model's strongest editing trait.

Reference-guided swaps and insertion

Feed a reference image or a second clip to steer the edit. Swap a garment onto a moving subject, or insert a product or screen-content video so it tracks the source.

Subject-to-video identity

Place a person or character from reference images into a new scene and the face stays recognizable as they move. Identity preservation is the model's standout result in ByteDance's subject-to-video evaluations.

Motion redirection

Change what a subject is doing inside an existing clip, a person crouches instead of bending over, while pose continuity, lighting, and background stay put.

Open-weights model family

An MLLM planner paired with a 14B DiT renderer built on Wan2.2, released under Apache 2.0. The weights are self-hostable on your own hardware.

Add a snowman beside the dog and keep the rest of the clip unchangedPrueba ahora

Especificaciones técnicas

Planner + DiT

Qwen2.5-VL planner, 14B Wan2.2 renderer

Edit, Generate, R2V

Editing, generation, subject-to-video

480p / 16fps

Default render setting

Apache 2.0

Open weights, self-hostable

Casos de uso

Augment real footage

Add or remove props, fix a detail, or restyle an element in a clip without re-shooting. The consistency lock keeps the rest of the shot identical, so edits read as native.

Recurring characters and avatars

Keep the same face across episodes, ads, or an avatar series. Subject-to-video holds a person's identity from a few reference images as they move through new scenes.

Virtual try-on and product placement

Swap clothing onto a moving model from a reference, or drop a product or on-screen video into a shot, for fashion and ad work that needs the source clip kept intact.

Re-block an action

Change what someone is doing in a take, a stand becomes a crouch, without re-filming. Motion editing alters the action while identity, framing, and lighting stay fixed.

Ejemplos de prompts

Consistency edit

Add a snowman beside the dog on the snowy path, and keep the dog, the road, and the trees unchanged

Edit prompt

Identity-locked subject

Place this person on a neon city rooftop at night, gently turning to camera, keeping their face and jacket

Edit prompt

Reference swap

Replace the outer shirt with the one in the reference image, keep the pose, lighting, and motion exactly

Edit prompt

Precios simples

Comienza gratis hoy, con la opción de mejorar tu plan o cancelar cuando quieras.

Basic

/ mes

facturado como $0 por año

900 mensual créditos

1 usuario

Todos los modelos

Workflows

Standard

/ mes

facturado como $0 por año

3200 mensual créditos

1 usuario

Todos los modelos

Workflows

Pro

/ mes

facturado como $0 por año

6200 compartidos mensual créditos

1 usuario

+ hasta 4 más con costo adicional

Todos los modelos

Workflows

Pro Max

/ mes

facturado como $0 por año

24000 compartidos mensual créditos

1 usuario

+ hasta 9 más con costo adicional

Todos los modelos

Workflows

Enterprise

Para límites más altos

Personalizado

términos de precios y facturación

High-volume credits

Límites de asientos personalizados

Todos los modelos

Workflows

Free

Para experimentar

gratis para siempre

Hasta 20 créditos

Solo 1 usuario

Modelos limitados

Workflows

Comparar detalles de los planes

Preguntas frecuentes

What is 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.

What can Bernini do?

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.

How is Bernini different from a standard video model?

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.

Is Bernini open source?

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.

How well does Bernini preserve a subject's identity?

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.

What resolution does Bernini output?

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.

Bernini

Funciones clave

Instruction-based video editing

Locked scene consistency

Reference-guided swaps and insertion

Subject-to-video identity

Motion redirection

Open-weights model family

Especificaciones técnicas

Casos de uso

Augment real footage

Recurring characters and avatars

Virtual try-on and product placement

Re-block an action

Ejemplos de prompts

Consistency edit

Identity-locked subject

Reference swap

Precios simples

Preguntas frecuentes

Otros modelos