ByteDance's open-weights video generation and editing model pairing a Qwen2.5-VL planner with a Wan2.2 renderer.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Bernini v2 (Bernini-Diffusers-v2) is a unified video generation and editing framework from ByteDance that combines an MLLM-based semantic planner (Qwen2.5-VL-7B-Instruct) with a DiT-based renderer built on Wan2.2-T2V-A14B. The checkpoint packages the planner, Bernini planning weights and Wan2.2 diffusion components in one self-contained diffusers-format directory. It is described in the paper "Bernini: Latent Semantic Planning for Video Diffusion".
It supports text-to-image, image-to-image, text-to-video, video-to-video, reference video-to-video and reference-to-video tasks. Per the model card, the planner decomposes complex instructions and plans semantic changes before rendering, which suits editing requests such as adding or removing objects, camera changes and reference insertion. Version 2 changes the training recipe by warming up the connector for thousands of steps before co-training, which the authors say improves reference-guided editing and OpenS2V results. Reported scores include EditVerse 8.02, OpenVE 3.96, OpenS2V 63.83 and VBench 84.46; these are the authors' own figures.
Weights are downloadable from Hugging Face under Apache-2.0, and inference code is in the Bernini GitHub repository (recommended environment: Python 3.11, PyTorch 2.7.1 with CUDA 12.6). The checkpoint is heavy (a YouTube reviewer cited roughly 180 GB), so it needs substantial GPU and storage resources for local use. The model card notes the full pipeline is heavier than the renderer-only Bernini-R releases.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source omnimodal video editor that edits videos via natural language, including adding or removing objects, camera changes and reference insertion. The model is ~180 GB, so the creator thinks it may be dead on arrival.
Closest neighbours by meaning, not just by category.
Model · ByteDance
ByteDance unified video generation and editing framework pairing an MLLM semantic planner with a DiT renderer.
Model · ByteDance
ByteDance Seed's video generation model supporting multi-reference inputs and one-take creation.
Model · ByteDance
ByteDance's open-weights one-step diffusion transformer for restoring and upscaling video and images.
Research · ByteDance
ByteDance research framework that distills multiple specialist image models into one via on-policy velocity matching.
Model · ByteDance
ByteDance Seed's multimodal video generation model, producing video from text and reference inputs.
Model · Sand AI
Open-weights 114B MoE model from Sand AI that generates synchronized video and audio from text or image prompts.
