ByteDance Open Sources Bernini Framework: Achieving Perfect Unity in Video Generation and Precise Editing

ByteDance's commercialization technology team has officially open-sourced a new video generation and editing framework called Bernini. The framework's core focuses on a "first understand, then generate" collaborative mechanism, aiming to effectively solve industry pain points such as image instability and frame flickering caused by traditional models' inability to accurately understand complex instructions.

Bernini currently ranks among the top in ByteDance's internal testing. Its inference code and the second-stage model Bernini-R have been officially released, and the full version with all features will be fully open in the near future.

Separating Semantics and Rendering

Bernini innovates in its workflow by splitting the entire process into two independent parts: "semantic planning" and "visual rendering." The system first uses a multimodal large model planner to deeply analyze input materials and sketch out a "semantic sketch," which is then converted into stable and continuous video frames by the renderer.

Thanks to this clear division of labor, the framework demonstrates high practical value in controllable editing. Users can not only make natural changes to weather, season, and visual style in the scene through simple commands, but also achieve precise control over camera perspective, focus, and subject actions.

Rich Visual Reference Dimensions

In addition to traditional text control, Bernini also supports using images and videos as visual references, greatly improving the consistency of creation. In video editing scenarios, it can accurately embed specific materials or posters into target areas, ensuring no boundary breakage or distorted perspective.

In new video generation scenarios, the model supports single-image and multi-angle reference generation, and can evolve from keyframes to continuous shots. To solve the problem of models easily getting confused when connecting multiple visual segments, the team introduced a dedicated position encoding mechanism to ensure clear distinction between reference materials and output targets.

Project: https://bernini-ai.github.io/

Understand First, Then Act! ByteDance Open-Sources Unified Framework Bernini to Make AI Video Editing Less About Guesswork

ByteDance open-sources the unified framework for video generation and editing called Bernini. Its core uses a 'understand first, then generate' collaborative mechanism to solve the problems of image instability and frame flickering caused by traditional models' inability to accurately understand complex text instructions. It breaks through bottlenecks such as subject deformation and background drift.

Large Model Agents Say Goodbye to Blind Stacking! Hong Kong Chinese University Team Releases SLIM Framework to Dynamically Manage the Lifecycle of External Skills

The Chinese University of Hong Kong team proposed the SLIM framework to address skill management challenges in evolving large language model agents from conversational to task-oriented capabilities. It enables dynamic skill lifecycle management, breaking the industry's blind skill accumulation, offering a new approach for efficient external ability management.....

OpenRouter Completes $113 Million in Series B Funding: Building a Traffic Hub for the Multi-Model Era

OpenRouter raised $113M in Series B funding, led by Alphabet's CapitalG, with participation from NVIDIA, ServiceNow, MongoDB, Snowflake, Databricks, a16z, and Menlo Ventures. The platform transformed from an experimental field to a productivity cornerstone amid explosive business growth in six months.....

AI Sheriff's Anniversary: Douyin's Large Model Combats Rumors, View Count Drops by 62%

On May 25, Douyin released its annual rumor control report, showing that the 'AI Truth-Seeking Model,' introduced a year ago, has shifted rumor management from 'humans searching for rumors' to 'models proactively tackling them.' The model reduced average rumor views by 62%, significantly shortening the spread window and improving information ecosystem governance efficiency.....