Peak-End Net is a video aesthetic assessment framework inspired by the
peak-end rule: people tend to judge an experience disproportionately by its
most salient moments and its ending, rather than by uniformly averaging the
entire experience. Peak-End Net translates this insight into a learnable
temporal model that predicts an overall aesthetic score together with ten
fine-grained attribute scores.
✨ Overview
Peak-End Net pipeline
The framework contains five main components:
Frame Aesthetic Perceiver — a frozen CLIP ViT-L/14 encoder and an
AVA-pretrained aesthetic head produce a 10-bin score distribution and an
expected aesthetic score for each frame.
Key Moment Discovery — learnable peak-, valley-, and end-aware signals
are combined into a unified temporal attention distribution.
Peak-End Aggregation — attention-weighted pooling summarizes the frame
features into a video-level representation.
Rhythm Encoder — a multi-scale 1D CNN with kernel sizes 3, 5, and 7
captures local fluctuations and longer-range trends in the frame-score
sequence.
Gated Fusion — a lightweight second-stage module adaptively combines the
learned video-level score with the mean frame-level AVA score.
🚀 Pretrained Model and Inference
The self-contained Peak-End-Net.pth checkpoint includes the CLIP ViT-L/14
encoder, AVA aesthetic head, Peak-End modules, and gated-fusion module. No
separate AVA or Stage 1 checkpoint is required for inference.
Clone the code repository and install dependencies:
The script reports the overall score, ten attribute scores
(composition, shotsize, lighting, visualtone, color, depthoffield,
expression, movement, costume, makeup), the fusion gate, and the two
scores combined by the gate.
🏋️ Training
Full training instructions — pretraining the AVA aesthetic head, Stage 1
(Peak-End Net), and Stage 2 (Gated Fusion) — are provided in the
GitHub repository.
📖 Citation
If you find this work useful, please cite:
bibtex
1@misc{li2026peakendnetpeakendruleinspired,
2 title={Peak-End-Net: A Peak-End Rule Inspired Framework for Generalizable Video Aesthetic Assessment},
3 author={Geng Li and Haiwen Li and Rui Chen and Jing Tang and Lei Sun and Xiangxiang Chu},
4 year={2026},
5 eprint={2607.13941},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2607.13941},
9}
