April 2018 Summaries

Per-title encoding, also known as content-adaptive encoding, is a technique that adjusts video quality based on the type of content. This approach can lead to significant improvements in video quality without increasing bandwidth usage. Mux has developed an instant per-title encoding method using deep learning models, which allows for near Netflix-level results with minimal additional cost and processing time. The process involves training a neural network on millions of videos across various content types and creating feature embeddings for each video ingested into the system. This enables the model to output optimal encoding ladders in just seconds, significantly reducing the time required compared to traditional methods. Initial results have shown improvements in quality as measured by VMAF (Video Multimethod Assessment Fusion), with some clips showing differences of up to 12 VMAF points. As the training set grows over time, the accuracy of the model is expected to improve further, leading to even better video quality for end-users at no additional cost.

The text discusses adaptive bitrate streaming and its importance in modern online video due to varying user bandwidths. It mentions that static encoding ladders are inherently wrong as videos vary widely in complexity and require different encoding approaches. Netflix's Per-Title Encoding Optimization is mentioned, which requires significant engineering time and investment but offers a 20% reduction in bitrates. The text introduces Instant Per-Title Encoding, a machine learning-based solution that analyzes video frames to output the right encoding ladder quickly and accurately. It also mentions that Mux Video provides this optimization at a fraction of the cost compared to running it yourself.