# ReCoNet: Real-time Coherent Video Style Transfer Network

**Authors:** Chang Gao, Derun Gu, Fangjun Zhang, Yizhou Yu  
**arXiv:** [1807.01197](https://arxiv.org/abs/1807.01197)  
**Date:** 2018-07-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1807.01197v2/

## Key points

- Existing image style transfer models based on convolutional neural networks have limitations when applied to videos
- These models often suffer from high temporal inconsistency, meaning that the style transfer is not consistent across consecutive frames in a video
- Some video style transfer models have been proposed to improve temporal consistency, but they sacrifice processing speed or perceptual style quality
- The authors propose a novel real-time video style transfer model called ReCoNet
- ReCoNet aims to generate temporally coherent style transfer videos while maintaining favorable perceptual styles
- Two key innovations are introduced: luminance warping constraint and feature-map-level temporal loss
- The luminance warping constraint captures luminance changes between consecutive frames and increases stylization stability under illumination effects
- The feature-map-level temporal loss enhances temporal consistency on traceable objects in videos by ensuring consistency of distinct features throughout the video sequence
- Experimental results demonstrate that ReCoNet achieves outstanding performance both qualitatively and quantitatively
- ReCoNet provides fast processing speed, nice perceptual style quality, and high temporal consistency simultaneously
- ReCoNet represents an important advancement in real-time video style transfer and can be used for various applications such as video editing and artistic expression.

## Summary

The existing summary discusses the limitations of existing image style transfer models based on convolutional neural networks when applied to videos. These models often suffer from high temporal inconsistency, meaning that the style transfer is not consistent across consecutive frames in a video. While some video style transfer models have been proposed to improve temporal consistency, they often sacrifice either processing speed, perceptual style quality, or both. In response to these challenges, the authors propose a novel real-time video style transfer model called ReCoNet. This model aims to generate temporally coherent style transfer videos while maintaining favorable perceptual styles. To achieve this, the authors introduce two key innovations. Firstly, they add a luminance warping constraint to the temporal loss at the output level. This constraint captures luminance changes between consecutive frames and increases stylization stability under illumination effects. By considering luminance changes, ReCoNet can produce more visually consistent and realistic style transfers. Secondly, the authors propose a feature-map-level temporal loss to further enhance temporal consistency on traceable objects in videos. This loss function ensures that objects with distinct features maintain their consistency throughout the video sequence. Experimental results demonstrate that ReCoNet achieves outstanding performance both qualitatively and quantitatively. The model successfully addresses the limitations of previous approaches by providing fast processing speed, nice perceptual style quality and high temporal consistency simultaneously. Overall, ReCoNet represents an important advancement in real-time video style transfer and its ability to generate temporally coherent and visually appealing stylized videos makes it a valuable tool for various applications such as video editing and artistic expression.