The most important news of this week: I generated fake cucumbers using StyleGAN. Do we need thisfruitdoesnotexist.com?..
This week is rich for generated stuff, have a look:
Creating realistic photos from arts (portraits, sculptures, etc.) online from neural.love
Styling faces to anime or portraits by AgileGAN
Styling text to any font from Facebook AI
Cartooned face-filters from Snapchat
Generating Instagram, models in, em, underwear aka thismodeldoesnotexist
Changing hairstyle by Barbershop GAN
Amazing faceswap (or deepfake?) that works with videos given a single(!) image to swap
ICLR Keynote on Revolution in Data Science
Course from Huggingface
To the details!
Portraits with love from neural.love
What: Guys from neural.love released their service to transform artistic or photographic sources into HD portraits using their own neural network pipelines. It is in beta version, and if you want to get rid of the watermark you can pay.
Play around here:https://neural.love/portraits
It looks great, especially on some nice art portraits. I uploaded the ”Girl with a Pearl Earring” by Vermeer, but the queue now is like 600 people, so it will take time to get the results. But, I’ve also uploaded the photo of my daughters doll, hah. I think that one was tricky for the network, just look at this creepy cutie.
Styling faces to anime or portraits by AgileGAN
What: This algorithm is transforming faces into 3 styles. They claim, that ~1 hour is enough to train this model, which is impressive. This face styling works best if the face is on white background in my experience)
Play around: http://www.agilegan.com/
Project page: https://guoxiansong.github.io/homepage/agilegan.html
TextStyleBrush - styling text to any font from Facebook AI
TextStyleBrush, an AI research project that can copy the style of text in a photo using just a single word. With this AI model, you can edit and replace text in images.
Project page: https://ai.facebook.com/blog/ai-can-now-emulate-text-style-in-images-in-one-shot-using-just-a-single-word
Unfortunately, no demo to play around available yet, but looks good!
New cartooned face-filter from Snapchat
What: Yet another cartoon face filter, but now in 3D. I tried it on my dog (black labrador) but it didn’t work, so, I tried it on myself instead, sorry. It did quite a good job in real-time, although my beard has disappeared (but tried to come back, you can see it below the filter, lol)
Where: https://lens.snapchat.com/bad1396b0f8d494d845b1d6438c115f2
Thismodeldoesnotexist
What: Alright, someone is trying to generate the whole person's body and not only the face. Looks good, but faces sometimes are creepy. Step closer for full human body generation. Future datasets need wider diversity though - the current task is too specific and model oriented.
Where: https://www.instagram.com/thismodeldoesnotexist/
Barbershop GAN
What: A model to change the hairstyle of a person in the photo. Works as the combination of segmentation and GAN-based style transfer.
Project page: https://zpdesu.github.io/Barbershop/
Faceswap (or deepfake?) that works with videos given a single(!) image
What: Yet another face swap, but looks good. Especially on the videos. Impressive, that it can work with a single image. There are high-quality videos you can find on the project repo.
Where: description and the testing code are here, training code is coming soon.
Non-GAN related stuff
ICLR Keynote on Revolution in Data Science
What: Great talk from ICLR keynote about Geometric foundations of Deep Learning.
Geometric Deep Learning is an attempt for geometric unification of a broad class of ML problems from the perspectives of symmetry and invariance. These principles not only underlie the breakthrough performance of convolutional neural networks and the recent success of graph neural networks but also provide a principled way to construct new types of problem-specific inductive biases.
Blog post here.
Huggingface course
What: One of the best AI communities releases its course. Can’t give feedback yet, but should be good to get NLP knowledge.
This course will teach you about natural language processing (NLP) using libraries from the Hugging Face ecosystem — 🤗 Transformers, 🤗 Datasets, 🤗 Tokenizers, and 🤗 Accelerate — as well as the Hugging Face Hub. It’s completely free and without ads.
Where: https://huggingface.co/course/
That’s all for this week!