amap-ml/peak-end-net

Peak End Net is a research project that teaches a computer model to judge how visually pleasing a video looks, in a way that tries to match how people actually judge things. Instead of averaging every single frame equally, the model is built around a psychology idea called the peak end rule: people tend to remember and judge an experience mostly by its most striking moment and by how it ends, not by everything in between. This paper was accepted at the ACM Multimedia 2026 conference. The model looks at each frame using a pretrained image understanding model, then combines the frame scores across the whole video, giving extra weight to the most attention grabbing moments and to the final part of the video. A separate module then blends this video level judgment with a simpler frame averaged score to produce the final result. The output includes one overall aesthetic score plus ten more detailed scores covering things like lighting, color, composition, and framing. To use the project, a person needs Python along with a matching version of PyTorch for their graphics card, and ideally a capable GPU, although running on a regular processor is possible though slower. A ready to use trained model can be downloaded and used directly to score a video without training anything yourself. Training your own version from scratch is a three step process: first training a base aesthetic judging model on a public image dataset, then training the main video model, and finally training a small module that blends the two scoring methods together. This project builds on existing published research code and is released under the MIT license, so it can be reused and modified freely, including for commercial purposes.