ByteDance Open-Sources Bernini, an Apache-2.0 Video Editor That Plans the Edit Before It Renders
- ByteDance open-sourced Bernini under Apache-2.0, shipping in pieces across June 2026: renderer weights on June 1, a lightweight 1.3B variant on June 9, and the full pipeline on June 11.
- Bernini splits editing into two brains: a Qwen2.5-VL-7B semantic planner predicts a target semantic embedding from the text instruction, source video, and reference frames, then a DiT renderer built on Alibaba's Wan 2.2 backbone runs flow-matching denoising conditioned on that plan.
- A mechanism called Segment-Aware 3D Rotary Positional Embedding (SA-3D RoPE) tags every token by which input it came from, so the model can reason across a source clip, a reference photo, and an inserted object without mixing them.
- The same model covers six tasks through one prompt box: text-to-image, image editing, text-to-video, video editing, reference-to-video, and reference-guided video editing, plus content insertion.
- The claimed payoff is that a named change is applied while everything not named stays frozen, avoiding the flicker and drift that come from editing pipelines that mutate the whole clip.