Few-Shot Adversarial Learning of Realistic Neural Talking Head Models

By Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, Victor S. Lempitsky. 2019

692 citations recorded by the source corpus.

Several recent works have shown how highly realistic human head images can be obtained by training convolutional neural networks to generate them. In order to create a personalized talking head model, these works require training on a large dataset of images of a single person. However, in many practical scenarios, such personalized talking head models need to be learned from a few image views of a person, potentially even a single image. Here, we present a system with such few-shot capability. It performs lengthy meta-learning on a large dataset of videos, and after that is able to frame few- and one-shot learning of neural talking head models of previously unseen people as adversarial training problems with high capacity generators and discriminators. Crucially, the system is able to initialize the parameters of both the generator and the discriminator in a person-specific way, so that training can be based on just a few images and done quickly, despite the need to tune tens of millions of parameters. We show that such an approach is able to learn highly realistic and personalized talking head models of new people and even portrait paintings.

Subfields: Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

ORBIT: A Real-World Few-Shot Dataset for Teachable Object Recognition (2021, background): “[52] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and”
StyleFlow: Attribute-conditioned Exploration of StyleGAN-Generated Images using Conditional Continuous Normalizing Flows (2020, background): “Other notable works (Nie et al. 2020; Siarohin et al. 2020; Zakharov et al. 2019) produce high quality image animations and manipulations.”
Meta-Learning in Neural Networks: A Survey (2020, neutral): “Image and Video Generation In [45] an amortized probabilistic meta-learner is used to generate multiple views of an object from just a single image, generative query networks [213] render scenes from novel views, and talking faces are generated from little data by learning the initialization of an a”
Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs (2020, methodology): “…e.g ., transferring knowledge of a well-trained source model [45, 32, 44], learning meta-knowledge for quick adaptation to a target domain [24, 47, 42], using an auxiliary task to facilitate training [7, 26, 48, 49], improving an inference procedure of suboptimal models [2, 39, 29, 38], using…”
DeepLandscape: Adversarial Modeling of Landscape Video (2020, methodology): “Here, we borrow ideas of using an encoder into the latent space followed by gradient descent [25], the latent space expansion for StyleGAN [26], and generator fine-tuning [27, 28].”
FakeSpotter: A Simple Baseline for Spotting AI-Synthesized Fake Faces (2019, background): “A recent work [5] could swap faces with only several photos in high quality.”
Deepfake Video Detection Using Convolutional Vision Transformer (2021, background): “With the rapid advancement of the CNNs [4, 20], Generative Adversarial Networks (GANs) [18], and its variants [22], it is now possible to create hyper-realistic images [32], videos [61] and audio signals [53, 15] that are much harder to detect and distinguish from real untampered audiovisuals.”
PIE (2020, methodology): “Few-shot editing techniques [Wang et al. 2019a; Wiles et al. 2018; Zakharov et al. 2019] require only a small set of images of the target person as input.”
The DeepFake Detection Challenge Dataset (2020, background): “Like other GAN methods, NTH did not produce good results in darker settings.”
What comprises a good talking-head video generation?: A Survey and Benchmark (2020, methodology): “4 shows the t-SNE plot of different ESD features on CREMA-D testing set, from where we observe that the group boundaries of ESD feature extracted from baseline method, Wiles et al. [18], and Zakharov et al. [20] are more clear than the ESD feature extracted from synthesized videos produced by Vougio”
Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction (2020, methodology): “Instead of this dense conditioning input or rendered feature maps, there are also methods that work on rendered facial landmarks [39, 8, 38].”
Nerfies: Deformable Neural Radiance Fields (2020, methodology): “A common strategy is to use re-rendering networks to produce more realistic results from proxy renders [18, 25, 32, 35, 59].”
Memory Efficient Meta-Learning with Large Images (2021, background): “Societal impact Few-shot learning systems hold much positive potential – from personalizing object recognizers for people who are blind [14] to rendering personalized avatars [30] (see [23] for a full review).”
LOHO: Latent Optimization of Hairstyles via Orthogonalization (2021, background): “GAN based alignment networks [38, 5] may be used to transfer pose, or alignment of hair across difﬁcult samples.”
Deepfakes generation and detection: state-of-the-art, open challenges, countermeasures, and way forward (2021, methodology): “[90] presented a meta transfer learning approach where the network was first trained on multiple identities and then fine-tuned on the target identity.”
First Order Motion Model for Image Animation (2020, background): “Recently, deep generative models have emerged as effective techniques for image animation and video retargeting [2, 41, 3, 42, 27, 28, 37, 40, 31, 21].”
Cross-Domain and Disentangled Face Manipulation With 3D Guidance (2021, methodology): “Facial landmarks are also popular to represent facial geometry, widely used in conditional face translation [Ha et al. 2020; Natsume et al. 2018; Nirkin et al. 2019; Xiang et al. 2020; Zakharov et al. 2019].”
The Creation and Detection of Deepfakes (2020, background): “Ladv: y, kt , ൗ xg xt (i) LFM: xg ∙ vst, xt (i) ∙ vst [183] Neural TalkingHeads:”