This is a short YouTube explainer by Károly Zsolnai-Fehér (2 Minute Papers) summarizing a new AI style-transfer method that learns an artistic style from a collection of images rather than a single example. The presenter contrasts single-image style transfer with collection-trained approaches, shows visual examples (including Van Gogh works) and HD video demos, and reports that human art-history experts judged AI-produced paintings as coming from an artist 39% of the time. The video also notes a new style-aware loss and training differences described in the paper, and mentions the demo runs at about nine images per second for HD video.
Questions this video answers
About nine images (frames) per second.
39% of the time.
Chapters
Dear fellow scholars, this is 2 minute papers with Zsolnai. Style transfer is a mostly AI based technique where we take a photograph, put a painting next to it, and it applies the style of the painting to our photo. A key insight of this new work is that a style is complex and it can only be approximated with one image. One image is just one instance of a style, not the style itself. Have a look here. If we take this content image and use Van Gogh's Road with Cypress and Star Painting as the art style, we get this. However, if we would have used Starry Night instead, it would have resulted in this. This is not learning about the style, this is learning a specific instance of the style. Here you see two previous algorithms that were instead trained on a collection of works from Van Gogh. However, you see that they are a little blurry and lack detail. This new technique is able to address this really well. Also look at how convincingly it stylized the top silhouettes of the bell tower. It can also deal with HD videos at a reasonable speed of nine of these images per second. Very tasty. Love it. And of course, as style transfer is a rapidly growing field, there are ample comparisons in the paper against other competing techniques. The results are very convincing. I feel that in most cases it represents the art style really well and can can decide where to leave the image content similar to the input and where to apply the style. So the overall outlook of the image remains similar. So we can look at these results and discuss who likes which one all day long. But there are also other more objective ways of evaluating such an algorithm. What is really cool is that the technique was tested by human art history experts and they not only found this method to be the most convincing of all the other style transfer methods, but also thought that the AI produced paintings were from an artist 39% of the time. So this means that the algorithm is able to learn the essence of an artistic style from a collection of images. This is a huge leap forward. Make sure to have a look at the paper that also describes a new style aware loss function and differences in the training process of this method as well. And if you enjoyed this episode and would like to see more, please help us exist through Patreon in this website. You can support the series and pick up cool perks like early access to these videos, deciding the order of future episodes and more. You know the drill. A dollar a month is almost nothing, but it keeps the papers coming. We also support cryptocurrencies. You'll find more information about this in the video description. Thanks for watching and for your generous support, and I'll see you next time.
