Get better data

[TanguyJeanneau]

The dataset is made of a collection of videos used as content, the StyleNet then process them and loss is computed by the different loss networks.

As of today, data comes from a stock image website, and is therefore not representative of all possible patterns and situations. I suspect that this might lower the end result's quality. Whe should consider using other data source, such as the Sintel dataset described and used in various papers (1).

(1) Ruder et al's use of Sintel: https://arxiv.org/pdf/1604.08610.pdf, Section 5.2
(2) Sintel dataset paper: https://homes.cs.washington.edu/~djbutler/papers/ButlerECCV2012.pdf