ByteDance drops Bernini 2, an open-source video editor that weighs 180GB

Bernini 2 is an open-source omnimodal video generation and editing pipeline from ByteDance. It allows you to edit existing videos using natural text prompts, like adding two characters to a scene, shifting the overall color temperature, removing specific objects, or even altering camera angles and swapping backgrounds.

You can also feed it reference images, allowing the model to seamlessly insert specific objects into a video, like taking a photo of a helicopter and placing it naturally into your footage.

What it can do, and how it works

Instead of spitting out video frames in a single pass, Bernini 2 splits the workload into two distinct stages: planning and rendering. It uses a Qwen2.5-VL multi-modal LLM as a latent semantic planner to break complex user instructions down into step-by-step layout plans, which are then passed to a Wan2.2-based DiT diffusion renderer to generate the final video.

Compared to the first release in June 2026, version 2 warms up the connector module for thousands of steps before full co-training. ByteDance says this significantly improves reference-guided video editing and OpenS2V performance.

In ByteDance's internal blind human arena evaluations, Bernini reportedly reached the top tier of leading closed-source commercial video editors across six core tasks, including text-to-video, video-to-video, and reference-guided editing.

Hardware requirements and how to run it

The biggest hurdle is pure size. Bernini 2 weighs in at 180GB just for the model weights, and that is before accounting for the T5 text encoder, VAE, tokenizer, and base diffusion modules. Running this locally on consumer hardware is essentially out of the question.

Tooling support is also still rough around the edges. ComfyUI does not support the full Bernini-Diffusers-v2 pipeline yet, because it cannot load the dual MLLM planner architecture. If you want ComfyUI workflows today, you have to use Bernini-R, a standalone renderer release that drops the multi-step planner entirely.

If you have enterprise GPU clusters handy, the model weights are available on Hugging Face under ByteDance/Bernini-Diffusers-v2, and inference scripts supporting single-GPU and multi-GPU runs live in ByteDance's official GitHub repo alongside arXiv paper 2605.22344.

Why it matters

Bernini 2 offers a clever architectural blueprint: putting an explicit reasoning planner in front of a diffusion renderer solves one of the biggest headaches in AI video, which is getting models to follow multi-part creative instructions without hallucinating.

However, for everyday creators and developers, this release might be practically dead on arrival for local workflows. A 180GB footprint is a non-starter for desktop setups, especially when competitive alternatives like MiniMax H3 already handle reference-guided video without demanding server racks. Unless the open-source community releases heavily quantized forks or lightweight ComfyUI adaptations, Bernini 2 will likely remain an enterprise-grade research artifact rather than a daily creative tool.