LARRUZ explores neural networks that help digital artists generate and manipulate images. The video covers neuron visualizations, compositional pattern producing networks, and 3D texture and style transfer tools, with source code available for viewers to experiment with.
Dear fellow scholars, this is two minute papers with in this series we have seen many times how good neural network based solutions are at image classification. This means that the network looks at an image and successfully identifies its contents. However, neural network based solutions are also capable of empowering art projects by generating new interesting images. This beautifully written paper is explores how a slight tweak to a problem definition can drastically change the output of such a neural network. It shows how many of these research works can be seen as the manifestation.
Of the same overarching idea.
For instance, we can try to visualize what groups of neurons within these networks are looking for and we get something like this. The reason for this is that important visual features like the eyes can appear at any part of the image and different groups of neurons look for it elsewhere. With a small modification, we can put these individual visualizations within a shared space and create a much more consistent and readable output. In a different experiment it is shown how a similar idea can be used with compositional pattern producing networks or CPPNs. In short, these networks are able to.
Take spatial positions as an input and.
Produce colors on the output, thereby creating interesting images of arbitrary resolution. Depending on the structure of this network, it can create beautiful images that are reminiscent of light paintings. And here you can see how the output of these networks change during the training process. They can also be used for image morphing as well. A similar idea can be used to create images that are beyond the classical 2D RGB images and create semi transparent images instead. And there is much, much more in the paper. For instance, there is an interactive demo.
That shows how we can seamlessly put.
This texture on a 3D object. It is also possible to perform neural style transfer on a 3D model.
This means that we have an image.
For style and the target 3D model.
And you can see the results over here.
This paper is a goldmine of knowledge and contains a lot of insights on how neural networks can further empower artists.
Working in the industry.
If you read only one paper today, it should definitely be this one. And this is not just about reading.
You can also play with these visualizations and as the source code is also available for all of these, you can.
Also build something amazing on top of them.
Let the experiments begin.
So this was a paper from the.
Amazing Distill Journal and just so you.
Know, they may be branching out to.
Different areas of expertise, which is amazing news. However, they are looking for a few.
Helping hands to accomplish that.
So make sure to click the link to this editorial update in the video.
Description to see how you can contribute.
I would personally love to see more.
Of these interactive articles.
Thanks for watching and for your generous.
Support and I'll see you next time.
