VToonify: A real-time AI model for generating artistic portrait videos

News Report Technology

In Brief

A revolutionary VToonify framework was developed by developers to provide controlled, high-resolution portrait video style transfers.

To produce stunning artistic portraits, the framework makes use of StyleGAN’s mid- and high-res layers.

It allows the extension of existing StyleGAN-based image toonification models to video.


The Trust Project is a worldwide group of news organizations working to establish transparency standards.

Researchers from Nanyang Technological University have introduced a novel VToonify framework to generate controllable high-resolution portrait video style transfer. VToonify leverages the mid- and high-resolution layers of StyleGAN to render high-quality artistic portraits based on the multi-scale content features extracted by an encoder to better preserve frame details. Experimental results show that our framework can generate videos with consistently high quality and desired facial expressions without the need for face alignment or frame-size restrictions.

As a result, a fully convolutional architecture that accepts non-aligned faces in videos of various sizes produces complete faces with organic motions. VToonify framework inherits appealing features of these models for flexible style control on color and intensity. It is compatible with existing StyleGAN-based image toonification models to extend them to video toonification. This work introduces two instantiations of VToonify for collection-based and exemplar-based portrait video style transfer, respectively, built upon Toonify and DualStyleGAN.

Vtoonify: A real-time AI model for generating artistic portrait videos

Extensive experimental findings show that proposed VToonify framework outperforms competing approaches in producing artistic portrait films with adjustable style controls that are of excellent quality and temporally consistent. Check GitHub for more details.

Related article: OpenAI is working on creating an AI model for video

In order to provide a controllable high-resolution portrait video style transfer, VToonify combines the advantages of the image translation framework and the StyleGAN-based framework.

(A) To support variable input size, an image translation system uses fully convolutional networks. It is challenging to impart high-resolution and controlled style, nevertheless, when teaching from scratch.

(B) StyleGAN-based framework, which only supports fixed picture size and detail losses, uses the pre-trained StyleGAN model for high-resolution and controllable style transfer.

(C) In order to create a completely convolutional encoder-generator architecture resembling that of the image translation framework, our hybrid system extends StyleGAN by deleting its fixed-sized input feature and low-resolution layers.

In order to preserve frame details, developers train an encoder to extract multi-scale content features from the input frame as an additional content condition. VToonify inherits the StyleGAN model’s style control flexibility by putting it into the generator to distill both its data and model.

Vtoonify: A real-time AI model for generating artistic portrait videos
Related article: Lambda Labs announced an AI image mixer that can combine up to five images

VToonify framework inherits the appealing characteristics for flexible style control from the current StyleGAN-based image toonification models and is compatible with them to expand them to video toonification. Our VToonify offers the following using the DualStyleGAN model as the StyleGAN foundation:

  • Transfer of style from exemplar-based structures;
  • Modification of style degree;
  • Transfer of color style based on exemplars.
Vtoonify: A real-time AI model for generating artistic portrait videos
For StyleGAN distillation, developers compare two backbones Toonify and DualStyleGAN, as well as the high-resolution image-to-image translation baseline Pix2pixHD. VToonify-T and VToonify-D outperform their comparable backbones, Toonify and DualStyleGAN, in terms of stylizing the entire video while keeping the same high quality and visual elements as the backbones for each individual frame. VToonify-T, for example, follows Toonify to impose a strong style effect, such as violet hair in the Arcane style. VToonify-D, on the other hand, does a better job of maintaining facial features. Pix2pixHD has flickers and artifacts when compared to VToonify-D.

Read more about AI:

Disclaimer

Any data, text, or other content on this page is provided as general market information and not as investment advice. Past performance is not necessarily an indicator of future results.

Damir Yalalov

Damir is the Editor/SEO/Product Lead at mpost.io. He is most interested in SecureTech, Blockchain, and FinTech startups. Damir earned a bachelor's degree in physics.

Follow Author

More Articles
🗞 Metaverse Newsletter
👾 Follow us
  YouTube Icon     YouTube Icon     YouTube Icon     YouTube Icon