Bernini-R Renderer Now Open: Turn AI Plans Into Photorealistic Video
ByteDance has released Bernini-R, the open-source renderer that powers its Bernini video generation and editing framework. The release provides inference code and model weights, letting users generate and edit videos with the same semantic planning approach used in the full system. Bernini-R takes plans created by a multimodal large language model and turns them into photorealistic images or video frames.
ByteDance developed Bernini as a unified model that divides work between an MLLM-based semantic planner and a DiT-based renderer. By open-sourcing the renderer now, the company gives researchers and developers the ability to run video editing and generation tasks on their own machines. This release concentrates on the rendering stage, while the complete planner stays integrated within the broader Bernini project.
How semantic planning improves video editing
Key features of Bernini-R
- Unifies large language model planning and diffusion.
- Handles text-to-image, video, and editing tasks.
- Preserves source detail with VAE features.
- Reference-image-guided editing and video insertion.
- Chain-of-thought reasoning for better accuracy.
- Segment-Aware 3D RoPE handles multiple visual inputs.
- Ready-to-use diffusers format simplifies loading.
- Interactive Gradio demo for quick testing.
This release is for AI researchers and skilled hobbyists who want to explore advanced video editing on local hardware. Small studios and privacy-conscious professionals can edit footage without sending data to the cloud, since everything runs on-device. The best performance currently demands a Hopper GPU like an H100, though other CUDA cards can still run the code with slower attention backends.
What you need to run Bernini-R
The team recommends a Hopper GPU (H100, H800, or H200) to take advantage of FlashAttention-3, but the pipeline automatically falls back to FlashAttention-2 or PyTorch SDPA on other CUDA GPUs such as A100s. Multi-GPU sequence parallelism is supported for scaling video generation, and a Gradio demo makes the whole pipeline accessible through a browser. On editing benchmarks, Bernini already ranks among the top tier of closed-source commercial models based on blinded human votes collected in the researchers’ own arena.
“We argue that these two families can be unified through a simple division of labor: MLLMs perform semantic planning, while diffusion models render pixels from high-level semantic guidance and low-level visual features.” — Source: Arxiv
Project Links
