Description

Bernini Video is ByteDance's open-source AI video creation platform that combines seven task types—text-to-video, image-to-video, video-to-video editing, reference-guided editing, reference-to-video, content insertion, and text-to-image—into one unified architecture. Its two-stage pipeline uses a Qwen2.5-VL MLLM planner to reason about composition, motion timing, and object relationships before a Wan 2.2 DiT renderer generates the final frames, enabling "think before you draw" intelligence that produces more coherent and instruction-faithful results. The model supports over 12 distinct edit types including background replacement, object addition or removal, style transformation, weather adjustment, expression changes, and camera angle shifts, with configurable output up to 720p at 24fps. Designed to run entirely on local hardware, it ensures that sensitive footage never leaves the user's machine, making it suitable for privacy-conscious professionals, AI researchers who need an open baseline (published on arXiv 2605.22344 with state-of-the-art benchmark results), open-source developers and indie hackers integrating AI video into their products at zero per-use cost, and ComfyUI enthusiasts who can chain video edits with other models in node-based workflows. The model weights are openly available on Hugging Face, the code is public on GitHub under Apache 2.0, and all generated content carries no watermark and comes with full commercial usage rights.