BMF (Babit Multimedia Framework) is a cross-platform, customizable multimedia processing framework developed by ByteDance. With over 4 years of testing and improvements, BMF has been tailored to adeptly tackle challenges in our real-world production environments. Now it's widely used in ByteDance's video streaming, live transcoding, cloud editing and mobile pre/post processing scenarios. More than 2 billion videos are processed by the framework everyday.
Here are some key features:
-
Cross-Platform Support: Native compatibility with Linux, Windows, and Mac OS, as well as optimization for both x86 and ARM CPUs.
-
Easy to use: BMF provides Python, Go, and C++ APIs, allowing developers the flexibility to code in their favourite languages.
-
Customizability: Developers can enhance the framework's features by adding their own modules, thanks to its decoupled architecture.
-
High performance: BMF has a powerful scheduler and strong support for heterogeneous acceleration hardware. Moreover, NVIDIA has been cooperating with us to develop a highly optimized GPU pipeline for video transcoding and AI inference.
-
Efficient data conversion: BMF offers seamless data format conversions across popular frameworks (PyTorch/OpenCV/TensorRT) and between hardware devices (CPU/GPU).
Dive deeper into BMF's capabilities on our website for more details.
In this section, we will directly showcase the capabilities of the BMF framework around five dimensions: Transcode, Edit, Meeting/Broadcaster, GPU acceleration, and AI Inference. For all the demos provided below, corresponding implementations and documentation are available on Google Colab, allowing you to experience them intuitively.
This demo describes step-by-step how to use BMF to develop a transcoding program, including video transcoding, audio transcoding, and image transcoding. In it, you can familiarize yourself with how to use BMF and how to use FFmpeg-compatible options to achieve the capabilities you need.
If you want to have a quick experiment, you can try it on
The Edit Demo will show you how to implement a high-complexity audio and video editing pipeline through the BMF framework. We have implemented two Python modules, video_concat and video_overlay, and combined various atomic capabilities to construct a complex BMF Graph.
If you want to have a quick experiment, you can try it on
This demo uses BMF framework to construct a simple broadcast service. The service provides an API that enables dynamic video source pulling, video layout control, audio mixing, and ultimately streaming the output to an RTMP server. This demo showcases the modularity of BMF, multi-language development, and the ability of dynamically adjusting the pipeline.
Below is a screen recording demonstrating the operation of broadcaster:
The video frame extraction acceleration demo shows:
-
BMF flexible capability of:
- Multi-language programming，we can see multi-language module work together in the demo
- Ability extend easily, there are new C++, Python modules added simply
- FFmpeg ability fully compatible
-
Hardware acceleration quickly enablement and CPU/GPU pipeline support
- Heterogeneous pipeline is supported in BMF, such as process between CPU and GPU
- Usefull hardware color space convertion in BMF
If you want to have a quick experiment, you can try it on
The GPU transcoding and filter module demo shows:
- Common video/image filters in BMF accelerated by GPU
- How to write GPU modules in BMF
The demo builds a transcoding pipeline which fully runs on GPU:
decode->scale->flip->rotate->crop->blur->encode
If you want to have a quick experiment, you can try it on
This demo shows the how to integrate the state of art AI algorithms into the BMF video processing pipeline. The famous open source colorization algorithm DeOldify is wrapped as an BMF pyhton module in less than 100 lines of codes. The final effect is illustrated below, with the original video on the left side and the colored video on the right.
If you want to have a quick experiment, you can try it on
This demo implements the super-resolution inference process of Real-ESRGAN as a BMF module, showcasing a BMF pipeline that combines decoding, super-resolution inference and encoding.
If you want to have a quick experiment, you can try it on
This demo shows how to invoke our aesthetic assessment model using bmf. Our deep learning model Aesmode has achieved a binary classification accuracy of 83.8% on AVA dataset, reaching the level of academic SOTA, and can be directly used to evaluate the aesthetic degree of videos by means of frame extraction processing.
If you want to have a quick experiment, you can try it on
This Demo shows a full-link face detect pipeline based on TensorRT acceleration, which internally uses the TensorRT-accelerated Onnx model to process the input video, and uses the NMS algorithm to filter repeated candidate boxes to form an output, which can be used to efficiently process a Face Detection Task.
If you want to have a quick experiment, you can try it on
-
- Install
- Create a Graph
- one of transcode example with 3 languages
- Use Module Directly
-
sync mode with 3 languages. You can try it on:
Python:
C++:
-
- Create a Module
- customize module with python, C++ and Go. You can try it on
- customize module with python, C++ and Go. You can try it on
-
APIs
The project has an Apache 2.0 License.
Contributions are welcomed. Please follow the guidelines.
We use GitHub issues to track and resolve problems. If you have any questions, please feel free to join the discussion and work with us to find a solution.
The decoder, encoder and filter reference ffmpeg cmdline tool, and are wrapped as BMF's built-in modules under a LGPL license.
The project also draws inspiration from other popular frameworks, such as ffmpeg-python and mediapipe. Our website is using the project from docsy based on hugo.
Here, we'd like to express our sincerest thanks to the developers of the above projects!