NeRFFaceEditing, a mask editing method for facial neural radiation fields, can edit three-dimensional faces without 3D modeling.

Want to personally design a highly realistic three-dimensional face, but find that you are not familiar with professional design software? The 3D face editing method NeRFFaceEditing provides a new solution. Even if you don’t know 3D modeling, you can freely edit highly realistic three-dimensional faces and model personalized digital portraits in the metaverse!

NeRFFaceEditing was completed by researchers from the Institute of Computing Technology, Chinese Academy of Sciences and City University of Hong Kong. Related technical papers were published at ACM SIGGRAPH Asia 2022, the top conference on computer graphics.

##Project homepage: http://geometrylearning.com/NeRFFaceEditing/

NeRFFaceEditing uses two-dimensional semantic masks as a bridge for three-dimensional geometry editing. The semantic editing performed by users in one perspective can be propagated to the entire three-dimensional face geometry while keeping the material unchanged. Furthermore, given an image representing a reference style, the user can easily change the material style of the entire 3D face while keeping the geometry unchanged.

Based on the 3D face editing system based on this method, even users who are not familiar with professional 3D design can easily carry out personalized face design and customize face shape and appearance. Let’s first look at two amazing effects using NeRFFaceEditing!

​Figure 1 Geometric editing effect: the edit on the two-dimensional semantic mask is propagated to the entire three-dimensional geometric space

Figure 2 Style transfer effect: Apply a given style to the entire three-dimensional face space. Keeping geometry unchanged

Part I Background

In recent years, with the development of neural radiation fields[1] and adversarial generative networks[2 ], various high-quality, fast-rendering 3D face generation networks have been proposed, including EG3D [3].

##Figure 3 Generation effects and geometric representation of different perspectives of EG3D

The three-plane representation of this method combines the traditional two-dimensional generative adversarial network and the latest three-dimensional implicit representation, so it inherits the powerful generation ability of StyleGAN [4] and the representation ability of neural radiation fields. However, these generative models cannot provide decoupled control of the geometry and material of the human face, which is an indispensable feature for applications such as 3D character design.

Existing work, such as DeepFaceDrawing [5] and DeepFaceEditing [6], can realize the decoupling control of geometry and material based on line drawing and the generation and editing of two-dimensional face images. DeepFaceVideoEditing [7] applies line drawing editing to face videos, which can generate rich editing effects in time series.

However, image decoupling and editing methods are difficult to directly apply to three-dimensional space. However, existing geometric and material decoupling methods for three-dimensional faces often require retraining network parameters, and the spatial representation methods used have major limitations and lack the good properties of three-plane representation. In order to solve the above problems, NeRFFaceEditing is based on the pre-trained model parameters of the three-dimensional generative adversarial network represented by the three planes, and uses the two-dimensional semantic mask from any perspective as a medium to realize geometric editing of the three-dimensional face and the solution of the material. coupling control.

Part 2 The algorithm principle of NeRFFaceEditing

After the three-plane generator generates the three-plane, it is inspired by AdaIN [8], that is, for the two-dimensional feature map (Feature Map), its Statistics can represent its style, and NeRFFaceEditing decomposes the three planes into mean and standard deviation (a) that express spatially invariant high-level material features, and normalized three planes that express spatially varying geometric features. Combining the standardized three-plane and the decomposed material characteristics (a) can restore the original three-plane. Therefore, given different material characteristics, the same geometry can be given different materials.

Further, in order to achieve decoupled control of geometry and materials, NeRFFaceEditing decomposes the original single decoder into a geometry decoder and a material decoder. The geometry decoder inputs features, predicted densities and semantic labels obtained from normalized three-plane sampling, and is used to express the geometric and semantic mask volume (Volume) of the 3D face. After the geometric features and material features (a) are combined through the controllable material module (CAM) module, the sampled features are then input into the material decoder to predict the color. Finally, through volume rendering, the face image and the corresponding semantic mask from a certain perspective are obtained. Given a different material feature (b), the geometric feature and material feature (b) can be used to obtain another face image with unchanged geometry and changed material through the CAM module and volume rendering. The overall network structure is shown in the figure below:

##Figure 4 NeRFFaceEditing network architecture

In addition, in order to constrain the rendering results of samples with the same material features but different geometries to be similar in material, NeRFFaceEditing uses the generated semantic masks and uses histogram features to represent these respectively. The color distribution of different facial components, such as hair, skin, etc., in samples with the same material characteristics but different geometry. The distance sum of the color distribution of these samples over the individual components is then optimized. As shown in the figure below:

Figure 5 Material similarity constraint training strategy

Part 3 Effect display and experimental comparison

Using NeRFFaceEditing, you can use two-dimensional semantic masks to perform geometric editing on the three-dimensional face space:

Figure 6 Three-dimensional face geometry editing

In addition, you can also edit based on Refer to the picture to perform material style migration in a three-dimensional consistent three-dimensional space:

Figure 7 Three-dimensional face style migration

On this basis, decoupled face interpolation deformation application can be realized. As shown in the figure below, the upper left corner and lower right corner are used as the starting and ending points, and the camera, geometry, and material are Perform linear interpolation:

Figure 8 Decoupling face deformation effect display

Using PTI [9] to back-project real images into the latent space of NeRFFaceEditing, editing and style transfer of real images can also be achieved. Through this, NeRFFaceEditing was also compared with other open source methods for face editing that can control the viewing angle, namely SofGAN [10], proving the superiority of the method.

Figure 9 Example of three-dimensional geometry editing of real images. It can be seen that the authenticity of NeRFFaceEditing is better than SofGAN, and SofGAN has certain changes in identity from other perspectives.

Figure 10 Example of real image style transfer. It can be seen that SofGAN has certain flaws and there are certain changes in identity.

Part 4 Conclusion and Acknowledgments

Digital content generation is widely used in the fields of industrial production and digital media, especially for virtual digital people Generation and editing have received widespread attention recently, and the decoupled editing of 3D face geometry and materials is a possible solution for personalized shaping of real virtual images.

NeRFFaceEditing system, by decoupling the design of the three-dimensional face generation network, can transform the user's modification of the semantic mask from the two-dimensional perspective into the geometry of the entire three-dimensional space Modify and ensure that the material does not change. In addition, with the help of training strategies that enhance the style transfer effect, effective material style transfer in three-dimensional space can be achieved. NeRFFaceEditing's paper has been accepted by ACM SIGGRAPH ASIA 2022, a top computer graphics conference.

The research team of this project includes Jiang Kaiwen (first author), an undergraduate student in the elite class of the Institute of Computing Technology, Chinese Academy of Sciences, Associate Researcher Gao Lin (corresponding author of this article), Dr. Chen Shuyu and City University of Hong Kong Professor Fu Hongbo, etc. For more details about the paper, please visit the project homepage:

​http://geometrylearning.com/NeRFFaceEditing/​

The above is the detailed content of NeRFFaceEditing, a mask editing method for facial neural radiation fields, can edit three-dimensional faces without 3D modeling.. For more information, please follow other related articles on the PHP Chinese website!