Overview
Bernini 7+14B is a unified image and video generation and editing framework. The full pipeline combines a Qwen2.5-VL-7B-based multimodal semantic planner with a Wan2.2-T2V-A14B diffusion renderer. It is distinct from the renderer-only Bernini-R family. (github.com)
Capabilities and uses
- Text-to-image generation and image editing.
- Text-to-video generation and video editing, including reference-guided video editing and reference-to-video generation.
- Complex generation and editing requests where decomposing instructions and planning semantic changes before rendering may improve instruction following. (github.com)
Release dates
- May 22, 2026: The repository identifies this as the paper release date, not the full model-weight release date.
- June 11, 2026: The repository says the full Bernini inference code and weights were open-sourced. The official model card instead lists June 10, 2026, so the sources differ by one day. (github.com)
Limitations and deployment
- The full pipeline has a heavier checkpoint layout than Bernini-R. The documented setup requires a CUDA GPU and recommends Hopper-class hardware; the supplied full-pipeline video scripts default to 480p, 16 fps and 81 frames. These are script defaults, not verified universal maximums. (huggingface.co)
- The repository and model card state Apache License 2.0. Review the applicable terms and dependencies before use. (github.com)
Pricing
As of November 16, 2026, Bernini-specific official API pricing, including separate input and output charges, could not be verified on either the mainland China Volcengine pricing page or the international BytePlus ModelArk pricing page. No CNY/RMB or USD amount is assumed. Self-hosting and optional external services may incur separate costs. (console.volcengine.com)
Official sources
- Bernini repository and announcement
- Full-pipeline documentation
- Official model card and weights
- Volcengine mainland model pricing — checked November 16, 2026.
- BytePlus international model pricing — checked November 16, 2026.