DeepFaceLive/doc/developer_faq/developer_faq.md

Why Python for such high load software and not C++? All CPU-intensive tasks are done by native libraries compiled for Python, such as OpenCV, NumPy, PyQt, onnxruntime. Python is more like a command processor that tells what to do. All you need is to organize such commands properly. What does the internal architecture look like? It consists of backend modules that work in separate processes. The modules work like a conveyor belt. CameraSource(FileSource) generates the frame and sends it to the next module for processing. The final output module outputs the stream to the screen with the desired delay. Backend modules manage the abstract controls that are implemented in the UI. Thus, the Model-View-Controller pattern is implemented. To reduce latency, some custom interprocess communication elements are implemented. What are the current problems for implementation for AMD ? Very slow inference in DirectML build of onnxruntime. no alternative for CuPy. Without CuPy FaceMerger will only work on the CPU, which is only applicable for frames less than 720p. What are the current problems for implementation for Linux ? No problems. Technically, you only need to write an installer, and check the work of all the modules. You may have to make some adjustments somewhere. I do not use linux, so I do not have time to support development on it. How many people were involved in the development? Just me. It took eight months until the first release.