Method, device and storage medium for feature-consistent realistic human face style transfer for digitalized garment sample presentation

Abstract

The application discloses a feature-consistent real face style transfer method and device for digital clothing sample display and a storage medium. The method comprises the following steps: constructing a rendering style face dataset DRFHQ; under the constraint of a face contour and a color tone, a rendering style face generator StyleGAN2-DRFHQ is obtained by fine-tuning a StyleGAN2-FFHQ generator using the DRFHQ; based on the StyleGAN2-DRFHQ, an e4e-DRFHQ encoder is obtained by fine-tuning an e4e-FFHQ encoder using the DRFHQ; a target rendering style face image is obtained from a digital clothing sample display image and is input into the e4e-DRFHQ encoder; a hidden code is obtained by inputting the target rendering style face image into the e4e-DRFHQ encoder; the hidden code is input into the StyleGAN2-FFHQ to generate a real face image; and the real face image is pasted back into the digital clothing sample display image.

Description

Technical Field

[0001] This invention relates to the field of realistic digital clothing display technology, specifically to a method, apparatus, and storage medium for realistic facial style transfer with consistent features for displaying digital clothing samples. Background Technology

[0002] With the rise of online apparel sourcing over the past decade, many apparel companies have been developing and selling clothing based on digital garment samples. Fashion designers use professional apparel design software (Browzwear, CLO, Optitex, etc.) to create digital garment samples and renderings to evaluate and showcase the overall effect of clothing designs. Clothing, the human body, and the face dominate the content of the rendered images. Obtaining photorealistic and convincing digital garment sample images requires a series of laborious and expertise-intensive operations, such as high-quality garment modeling, high-fidelity human body modeling, material creation, and lighting setup. Among these, the facial appearance of the virtual model is particularly crucial to the realism of the images.

[0003] Improving the realism of virtual model faces has long been a challenge. Over the past two decades, the booming entertainment industry, including animation, film, and video games, has driven significant advancements in high-quality facial modeling and rendering techniques. Under passive lighting conditions, methods based on multi-view stereo systems (Thabo Beeler, Bernd Bickel, Paul Beardsley, Bob Sumner, and Markus Gross. 2010. High-Quality Single-Shot Capture of Facial Geometry (SIGGRAPH'10)) can reconstruct high-quality facial geometry. Following the pioneering work of Debevec et al. (Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. 2000. Acquiring the Reflectance Field of a Human Face. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH'00). ACM Press / Addison-Wesley Publishing Co., USA, 145–156.), a series of light-stage-based facial appearance capture methods have been proposed to capture pore-level attributes of human faces. While these methods contribute to the generation of high-quality faces, they are very expensive and time-consuming. Moreover, because humans are inherently highly sensitive to facial features, even the most subtle, unrealistic details are easily observed. Japanese roboticist Masahiro Mori coined the term "uncanny valley" (Masahiro Mori, 1970. Bukimi no tani [the uncanny valley]. Energy 7 (1970), 33–35.) to describe how imperfect humanoid objects, such as robots, 3D animations, and lifelike dolls, can cause a person's reaction to them to suddenly shift from empathy to feeling strange.When these humanoid objects approach but fail to achieve a realistic human appearance, they can evoke feelings of fear, disgust, or "horror," creating the "uncanny valley" effect, which is detrimental to user experience and interaction (Roger K Moore. 2012. A Bayesian explanation of the 'Uncanny Valley' effect and related psychological phenomena. Scientific reports 2,1(2012),1–5.).

[0004] In recent years, the rise of deep learning, especially Generative Adversarial Networks (GANs), has inspired researchers to develop high-quality face generation methods. StyleGAN (Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-Based Generator Architecture for Generative Adversarial Networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR. 4401–4410.) and its inversion techniques have paved the way for semantic editing of photorealistic portraits. Thanks to the high generation quality and diversity of the pre-trained StyleGAN2-FFHQ generator, existing methods for improving face realism all involve projecting the rendered style of the face into the latent space of the aforementioned generator. Garbin et al. (Stephan J. Garbin, Marek Kowalski, Matthew Johnson, and Jamie Shotton. 2020. High Resolution Zero-Shot Domain Adaptation of Synthetically Rendered Face Images. In Computer Vision-ECCV 2020-16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII (Lecture Notes in Computer Science, Vol. 12373). Springer, 220–236.) matched non-realistic rendered portraits with the implicit codes of a pre-trained StyleGAN2 generator while maintaining consistency in pose, expression, hair, and lighting. Although their method attempts to adaptively transform to the real face domain, it still requires significant processing time for each input image. Furthermore, because the input images exceed the domain of the pre-trained model, the output often exhibits artifacts such as distortion and inconsistencies in facial features.Chandran et al. (Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus H. Gross, Paulo Gotardo, and Derek Bradley. 2021. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Trans. Graph. 40, 6(2021), 223:1–223:14.) proposed projecting high-quality but incompletely rendered facial skin into the latent space of StyleGAN2 to generate temporally continuous and realistic facial portraits. However, their method focuses more on repairing missing facial components (such as hair, eyes, and the inside of the mouth). Furthermore, the images output by their method still retain rendering style and lack realism.

[0005] To quickly generate realistic-looking faces from digital clothing sample images, the key lies in how to eliminate false rendering style information in the input rendered portrait image that can easily lead to the "uncanny valley" effect, while maintaining the consistency of facial features. State-of-the-art methods currently available either fail to maintain facial feature consistency, suffer from distortion, or have excessively long processing times (Stephan J. Garbin, Marek Kowalski, Matthew Johnson, and Jamie Shotton. 2020. High Resolution Zero-Shot Domain Adaptation of Synthetically Rendered FaceImages. In Computer Vision-ECCV 2020-16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII (Lecture Notes in Computer Science, Vol. 12373). Springer, 220–236.) or cannot effectively remove rendering styles (Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus H. Gross, Paulo Gotardo, and Derek Bradley. 2021. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM). (Trans.Graph.40,6(2021),223:1–223:14.) and other issues. Summary of the Invention

[0006] To enhance the realism of virtual model faces in digital garment sample displays, this invention provides a method for feature-consistent realistic face style transfer in digital garment sample displays. This method establishes correlations between face images of different styles through transfer learning. The key idea is to develop a fine-tuning method for consistent facial features, thereby obtaining a rendering style face generator that matches the facial features of the real-style StyleGAN2-FFHQ generator. Furthermore, this invention constructs a new high-quality rendering style face dataset, DRFHQ, which, besides being used to fine-tune the real-style StyleGAN2-FFHQ generator to obtain the target rendering style face generator StyleGAN2-DRFHQ, can also be applied to various other downstream tasks. This invention effectively improves the realism of digital garment sample displays and enhances the efficiency of garment design and production.

[0007] A method for feature-consistent, realistic facial style transfer for digital garment sample display includes the following steps:

[0008] (1) Construct the DRFHQ rendering style face dataset;

[0009] (2) Under the constraints of facial contour and tone, the StyleGAN2-FFHQ generator is fine-tuned using the DRFHQ rendering style face dataset to obtain the StyleGAN2-DRFHQ rendering style face generator; the StyleGAN2-FFHQ generator is a generator pre-trained on the FFHQ real face dataset by StyleGAN2-ada.

[0010] (3) Based on the StyleGAN2-DRFHQ rendering style face generator, the e4e-FFHQ encoder is fine-tuned using the DRFHQ rendering style face dataset to obtain the e4e-DRFHQ encoder.

[0011] (4) Input the target rendering style face image obtained from the digital clothing sample display image into the e4e-DRFHQ encoder to obtain the closest hidden code in the rendering style face latent space.

[0012] (5) Input the hidden code obtained in step (4) into the StyleGAN2-FFHQ generator to generate a realistic face image that is consistent with the face features in the input target rendering style face image. This process effectively removes the input rendering style information while maintaining the consistency of gender, facial contour and tone.

[0013] (6) Paste the realistic face image generated in step (5) back into the digital clothing sample display image in step (4).

[0014] In one embodiment, step (1) includes three steps: collecting rendered style face images from the Daz3D gallery using keyword search, aligning them according to the StyleGAN2 standard, and manual screening and sorting. All face images used for fine-tuning training and testing need to be aligned according to the StyleGAN2 standard. In this invention, the resolution of all face images after alignment is 1024×1024.

[0015] In one embodiment, step (2) referring to being constrained by facial contours and hue refers to the fine-tuning process introducing the following target loss function L. G :

[0016] L G =λ s L sketch +λ c L color

[0017] Where, λ s , λ c L represents the facial contour constraint loss. sketch Hue constraint loss L color The weights are all greater than 0;

[0018] Target loss function L G It was added to the original loss function term of StyleGAN2-ada to fine-tune and obtain the rendering style face generator StyleGAN2-DRFHQ.

[0019] In one embodiment, the facial contour constraint loss L sketch The following steps are used to calculate:

[0020] S101, input the randomly generated hidden code w+∈W+ into the frozen StyleGAN2-FFHQ generator G respectively. real and rendering style face generator G rendering Realistic face images G with a resolution of 1024×1024 were obtained. real (w+) and rendering style face image G rendering (w+); W+ is the latent space obtained by mapping the original latent space Z of StyleGAN2 to the latent space W through a multilayer perceptron and then expanding it.

[0021] S102, downsample the two face images obtained in step S101 to a resolution of 512×512;

[0022] S103, input the two downsampled face images from step S102 into the line drawing hand-drawn style face extractor S, and output two face line drawings as face outlines.

[0023] S104, Calculate the L1 loss of the two face line drawings obtained in step S103 as the face contour constraint loss L. sketch :

[0024] L sketch =‖S(G real (w+)↓ 512 )-S(G rendering (w+)↓ 512 )‖1

[0025] Among them, G real It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. rendering Using G real StyleGAN2-DRFHQ, a face generator with initialized parameters and requiring fine-tuning, ↓ 512 This indicates the operation of downsampling the face image to 512×512 resolution in step S102.

[0026] In one embodiment, the hue constraint loss L color The following steps are used to calculate:

[0027] S201, input the randomly generated hidden code w+∈W+ into the frozen StyleGAN2-FFHQ generator G respectively. real and rendering style face generator G rendering Realistic face images G with a resolution of 1024×1024 were obtained. real (w+) and rendering style face image G rendering (w+); W+ is the latent space obtained by mapping the original latent space Z of StyleGAN2 to the latent space W through a multilayer perceptron and then expanding it.

[0028] S202, downsample the two face images obtained in step S201 to a resolution of 256×256;

[0029] S203, process the two downsampled face images from step S202 using the Gaussian blur operator to obtain two Gaussian blurred low-resolution face images;

[0030] S204, Calculate the LPIPS loss of the two Gaussian blurred low-resolution face images obtained in step S203 as the tone constraint loss L. color :

[0031] L color =LPIPS(B(G real (w+)↓256 ),B(G rendering (w+)↓ 256 ))

[0032] Among them, G real It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. rendering Using G real The StyleGAN2-DRFHQ face generator, which requires parameter initialization and fine-tuning, where B represents the Gaussian blur operator, ↓ 256 This indicates the operation in step S202 of downsampling the face image to a resolution of 256×256.

[0033] In one embodiment, in step (2), the StyleGAN2-FFHQ generator is fine-tuned using the training parameters provided in the StyleGAN2 configuration file of StyleGAN2-ada. During the fine-tuning process, its ToRGB layer and mapping network are frozen, and only G is updated. rendering And the parameters of the discriminator D, and G real The face extractor S, which features a hand-drawn line drawing style, is frozen.

[0034] The purpose of step (3) is to enable the fine-tuned encoder e4e-DRFHQ to help the input rendered face find the closest latent code in the latent space. The e4e-FFHQ encoder refers to the e4e encoder pre-trained using the FFHQ real face dataset.

[0035] In step (5), the implicit code corresponding to the rendered style face obtained in step (4) is input into the StyleGAN2-FFHQ generator to generate a realistic face image that is consistent with the facial features of the input target rendered style face image. The principle is that during the fine-tuning training process in step (2), for each randomly generated implicit code, it is simultaneously input into G real and G rendering Then, by constraining the facial contours and tones, the generated facial features remained consistent, while only the style shifted from a realistic style to a rendering style.

[0036] The purpose of step (6) is to solve the problems of color difference and misalignment that can easily occur when directly pasting the generated realistic face image back into the original digital clothing sample display image.

[0037] In one embodiment, step (6) specifically includes the following steps:

[0038] S601, the realistic face image x generated in step (5) resPerform facial analysis to obtain segmentation and masking of facial skin, eyebrows, eyes, glasses, ears, nose, mouth, lips, and hair;

[0039] S602, combine all the segmented masks obtained in step S601 into a single mask m;

[0040] S603, the realistic face image x generated in step (5) res By performing the inverse StyleGAN2 alignment operation, the digital clothing sample display image from step (4) is pasted back to obtain Figure x. r ′ es ;

[0041] S604, the mask m obtained in step S602 is pasted onto a blank mask image with the same size as the digital clothing sample display image in step (4) after the same reverse StyleGAN2 alignment operation as in step S603, to obtain mask m';

[0042] S605, perform erosion and Gaussian blur operations on the mask m' to obtain a smoother composite result, thus obtaining the mask.

[0043] S606, the final effect image x is obtained by pasting the realistic face image generated in step (5) back into the digital clothing sample display image in step (4) through the following operation. final :

[0044]

[0045] Where x is the digital garment sample display image from step (4), This indicates element-wise multiplication.

[0046] The present invention also provides a computer device, including a memory and a processor, the memory for storing a computer program, and the processor for executing the computer program stored in the memory, wherein the computer program, when running, causes the processor to perform the method for feature-consistent realistic face style transfer for displaying digital clothing samples.

[0047] The present invention also provides a computer-readable storage medium storing a program or instructions that, when executed by a computer device, cause the computer device to perform the method for feature-consistent realistic face style transfer for displaying digital clothing samples.

[0048] Compared with existing technologies, the beneficial effects of this invention are as follows: The method for feature-consistent realistic face style transfer in digital clothing sample display can effectively improve the realism of digital clothing sample display, significantly improve the efficiency of clothing design and production, and can be used for aesthetic display of digital clothing in online shopping. This invention designs a novel photorealistic portrait generation framework that can effectively remove unnatural rendering styles from input portraits, avoiding user aversion caused by the "uncanny valley" effect. While effectively eliminating unnatural rendering styles, it maintains the consistency of input and output facial features, enabling fully automatic generation of realistic faces. The realistic portrait style transfer framework of this invention uses transfer learning to learn the consistent mapping relationship of facial features from the latent space of rendered style portraits to the latent space of realistic style portraits. This invention constructs a new high-quality rendered style face dataset, which, in addition to its use in this invention, can also be applied to various other downstream tasks. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a method for feature-consistent, realistic facial style transfer for displaying digital clothing samples, according to the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0051] See Figure 1 A method for feature-consistent, realistic facial style transfer for digital garment sample display, comprising the steps of:

[0052] (1) Construct the DRFHQ rendering style face dataset;

[0053] (2) Under the constraints of facial contour and tone, the StyleGAN2-FFHQ generator is fine-tuned using the DRFHQ rendering style face dataset to obtain the StyleGAN2-DRFHQ rendering style face generator; the StyleGAN2-FFHQ generator is a generator pre-trained on the FFHQ real face dataset by StyleGAN2-ada.

[0054] (3) Based on the StyleGAN2-DRFHQ rendering style face generator, the e4e-FFHQ encoder is fine-tuned using the DRFHQ rendering style face dataset to obtain the e4e-DRFHQ encoder.

[0055] (4) Input the target rendering style face image obtained from the digital clothing sample display image into the e4e-DRFHQ encoder to obtain the closest hidden code in the rendering style face latent space.

[0056] (5) Input the hidden code obtained in step (4) into the StyleGAN2-FFHQ generator to generate a realistic face image that is consistent with the face features in the input target rendering style face image. This process effectively removes the input rendering style information while maintaining the consistency of gender, facial contour and tone.

[0057] (6) Paste the realistic face image generated in step (5) back into the digital clothing sample display image in step (4).

[0058] In this embodiment, step (1) includes three steps: collecting rendered style face images from the Daz3D gallery using keyword search, aligning them according to the StyleGAN2 standard, and manually selecting and organizing them to obtain a dataset DRFHQ containing 11,399 rendered style face images. All face images used for fine-tuning training and testing need to be aligned according to the StyleGAN2 standard. In this invention, the resolution of all face images after alignment is 1024×1024.

[0059] In this embodiment, the step (2) mentioned under the constraints of facial contour and tone refers to the fine-tuning process introducing the following target loss function L. G :

[0060] L G =λ s L sketch +λ c L color

[0061] Where, λ s , λ c L represents the facial contour constraint loss. sketch Hue constraint loss L color The weights and λ s =5×10 -6 , λ c =3.75×10 3 ;

[0062] Target loss function L G It was added to the original loss function term of StyleGAN2-ada to fine-tune and obtain the rendering style face generator StyleGAN2-DRFHQ.

[0063] In this embodiment, the facial contour constraint loss L sketch The following steps are used to calculate:

[0064] S101, input the randomly generated hidden code w+∈W+ into the frozen StyleGAN2-FFHQ generator G respectively.real and rendering style face generator G rendering Realistic face images G with a resolution of 1024×1024 were obtained. real (w+) and rendering style face image G rendering (w+); W+ is the latent space obtained by mapping the original latent space Z of StyleGAN2 to the latent space W through a multilayer perceptron and then expanding it.

[0065] S102, downsample the two face images obtained in step S101 to a resolution of 512×512;

[0066] S103, input the two downsampled face images from step S102 into the line drawing hand-drawn style face extractor S, and output two face line drawings as face outlines.

[0067] S104, Calculate the L1 loss of the two face line drawings obtained in step S103 as the face contour constraint loss L. sketch :

[0068] L sketch =‖S(G real (w+)↓ 512 )-S(G rendering (w+)↓ 512 )‖1

[0069] Among them, G real It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. rendering Using G real StyleGAN2-DRFHQ, a face generator with initialized parameters and requiring fine-tuning, ↓ 512 This indicates the operation of downsampling the face image to 512×512 resolution in step S102.

[0070] In this embodiment, the hue constraint loss L color The following steps are used to calculate:

[0071] S201, input the randomly generated hidden code w+∈W+ into the frozen StyleGAN2-FFHQ generator G respectively. real and rendering style face generator G rendering Realistic face images G with a resolution of 1024×1024 were obtained. real (w+) and rendering style face image G rendering(w+); W+ is the latent space obtained by mapping the original latent space Z of StyleGAN2 to the latent space W through a multilayer perceptron and then expanding it.

[0072] S202, downsample the two face images obtained in step S201 to a resolution of 256×256;

[0073] S203, process the two downsampled face images from step S202 using the Gaussian blur operator to obtain two Gaussian blurred low-resolution face images;

[0074] S204, Calculate the LPIPS loss of the two Gaussian blurred low-resolution face images obtained in step S203 as the tone constraint loss L. color :

[0075] L color =LPIPS(B(G real (w+)↓ 256 ),B(G rendering (w+)↓ 256 ))

[0076] Among them, G real It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. rendering Using G real The parameter-initialized, fine-tuned rendering style face generator StyleGAN2-DRFHQ, where B represents kernel=13 and σ=10 Gaussian blur operator, ↓ 256 This indicates the operation in step S202 of downsampling the face image to a resolution of 256×256.

[0077] In this embodiment, in step (2), the training parameters provided in the StyleGAN2 configuration file of StyleGAN2-ada are used to fine-tune the StyleGAN2-FFHQ generator. During the fine-tuning process, its ToRGB layer and mapping network are frozen, and only G is updated. rendering And the parameters of the discriminator D, and G real The face extractor S, which features a hand-drawn line drawing style, is frozen.

[0078] The purpose of step (3) is to enable the fine-tuned encoder e4e-DRFHQ to help the input rendered face find the closest latent code in the latent space. The e4e-FFHQ encoder refers to the e4e encoder pre-trained using the FFHQ real face dataset.

[0079] In step (5), the implicit code corresponding to the rendered style face obtained in step (4) is input into the StyleGAN2-FFHQ generator to generate a realistic face image that is consistent with the facial features of the input target rendered style face image. The principle is that during the fine-tuning training process in step (2), for each randomly generated implicit code, it is simultaneously input into G real and G rendering Then, by constraining the facial contours and tones, the generated facial features remained consistent, while only the style shifted from a realistic style to a rendering style.

[0080] The purpose of step (6) is to solve the problems of color difference and misalignment that can easily occur when directly pasting the generated realistic face image back into the original digital clothing sample display image.

[0081] In this embodiment, step (6) specifically includes the following steps:

[0082] S601, the realistic face image x generated in step (5) res Perform facial analysis to obtain segmentation and masking of facial skin, eyebrows, eyes, glasses, ears, nose, mouth, lips, and hair;

[0083] S602, combine all the segmented masks obtained in step S601 into a single mask m;

[0084] S603, the realistic face image x generated in step (5) res By performing the inverse StyleGAN2 alignment operation, the digital clothing sample display image from step (4) is pasted back to obtain Figure x. r ′ es ;

[0085] S604, the mask m obtained in step S602 is pasted into a blank mask image (i.e., all black) with the same size as the digital clothing sample display image in step (4) after the same reverse StyleGAN2 alignment operation as in step S603, to obtain mask m'.

[0086] S605, perform erosion and Gaussian blur operations on the mask m' to obtain a smoother composite result, thus obtaining the mask.

[0087] S606, the final effect image x is obtained by pasting the realistic face image generated in step (5) back into the digital clothing sample display image in step (4) through the following operation. final :

[0088]

[0089] Where x is the digital garment sample display image from step (4), This indicates element-wise multiplication.

[0090] The present invention also provides a computer device, including a memory and a processor, the memory for storing a computer program, and the processor for executing the computer program stored in the memory, wherein the computer program, when running, causes the processor to perform the method for feature-consistent realistic face style transfer for displaying digital clothing samples.

[0091] The present invention also provides a computer-readable storage medium storing a program or instructions that, when executed by a computer device, cause the computer device to perform the method for feature-consistent realistic face style transfer for displaying digital clothing samples.

[0092] Furthermore, it should be understood that after reading the above description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for feature-consistent photorealistic face style transfer for digitalized apparel sample presentation, characterized in that, Including the following steps: (1) Construct the DRFHQ rendering style face dataset; (2) Under the constraints of facial contour and tone, the StyleGAN2-FFHQ generator is fine-tuned using the DRFHQ rendering style face dataset to obtain the StyleGAN2-DRFHQ rendering style face generator; the StyleGAN2-FFHQ generator is a generator pre-trained on the FFHQ real face dataset by StyleGAN2-ada. (3) Based on the StyleGAN2-DRFHQ rendering style face generator, the e4e-FFHQ encoder is fine-tuned using the DRFHQ rendering style face dataset to obtain the e4e-DRFHQ encoder. (4) Input the target rendering style face image obtained from the digital clothing sample display image into the e4e-DRFHQ encoder to obtain the closest hidden code in the rendering style face latent space. (5) Input the hidden code obtained in step (4) into the StyleGAN2-FFHQ generator to generate a realistic face image that is consistent with the face features in the input target rendering style face image. This process effectively removes the input rendering style information while maintaining the consistency of gender, facial contour and tone. (6) Paste the realistic face image generated in step (5) back into the digital clothing sample display image in step (4); The constraint of face contour and color tone in step (2) refers to the fine-tuning process to introduce the following target loss function : wherein, , respectively represent the weight of the face contour constraint loss , the hue constraint loss and are both greater than 0. Target loss function is added to the original loss function term of StyleGAN2-ada to fine-tune the rendering style face generator StyleGAN2-DRFHQ; Face contour constraint loss Computed by the following steps: S101, the randomly generated hidden code The inputs are respectively fed into the frozen StyleGAN2-FFHQ generator. and rendering style face generator Realistic face images with a resolution of 1024×1024 were obtained. and rendering style face image ; It is the original latent space of StyleGAN2 Mapped to the latent space by a multi-layer perceptron Then, the hidden space is expanded to obtain the hidden space; S102, downsample the two face images obtained in step S101 to a resolution of 512×512; S103, input the two downsampled face images from step S102 into the line art hand-drawn style face extractor respectively. In the process, output two line drawings of the human face as the facial outline; S104, calculate the L1 loss of the two face sketches obtained in step S103 as the face contour constraint loss : in, It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. It is StyleGAN2-DRFHQ, a face generator with initialized parameters and requiring fine-tuning, is used for rendering styles. This indicates the operation of downsampling the face image to 512×512 resolution in step S102.

2. The method according to claim 1, characterized in that, Step (1) includes three steps: collecting rendered style face images from the Daz3D gallery using keyword search, aligning them according to the StyleGAN2 standard, and manually screening and sorting them.

3. The method of claim 1, wherein, Hue constraint loss Computed by the following steps: S201, the randomly generated hidden code The inputs are respectively fed into the frozen StyleGAN2-FFHQ generator. and rendering style face generator Realistic face images with a resolution of 1024×1024 were obtained. and rendering style face image ; It is the original latent space of StyleGAN2 Mapped to the latent space by a multi-layer perceptron Then, the hidden space is expanded to obtain the hidden space; S202, downsample the two face images obtained in step S201 to a resolution of 256×256; S203, process the two downsampled face images from step S202 using the Gaussian blur operator to obtain two Gaussian blurred low-resolution face images; S204, Calculate the LPIPS loss of the two Gaussian blurred low-resolution face images obtained in step S203 as the tone constraint loss. : in, It is a frozen StyleGAN2 generator, StyleGAN2-FFHQ, pre-trained on the FFHQ real face dataset, with the StyleGAN2-ada structure as its backbone. It is StyleGAN2-DRFHQ, a face generator with initialized parameters and requiring fine-tuning, is used for rendering styles. This represents the Gaussian blur operator. This indicates the operation in step S202 of downsampling the face image to a resolution of 256×256.

4. The method according to claim 1 or 3, characterized in that, In step (2), the training parameters provided in the StyleGAN2 configuration file of StyleGAN2-ada are used to fine-tune the StyleGAN2-FFHQ generator. During the fine-tuning process, its ToRGB layer and mapping network are frozen, and only the following are updated: and discriminator The parameters, at the same time Face extractor with line drawing / hand-drawn style It's frozen.

5. The method of claim 1, wherein, Step (6) specifically includes the following steps: S601, the realistic face image generated in step (5) is processed Face analysis is performed to obtain segmentation masks of face skin, eyebrows, eyes, glasses, ears, nose, mouth, lips, and hair. S602, synthesize all the segmentation masks obtained in step S601 into one mask ; S603, the realistic face image generated in step (5) is used to replace the face in the image generated in step (4) to obtain a realistic face image The realistic face image is pasted back to the digitalized clothing sample display image in step (4) through the inverse StyleGAN2 alignment operation to obtain an image ; S604, apply the mask obtained in step S602. After performing the same reverse StyleGAN2 alignment operation as in step S603, it is pasted onto a blank mask image of the same size as the digital clothing sample display image in step (4), thus obtaining the mask. ; S605, performing erosion and Gaussian blur operations to get a smoother compositing result to get the mask ;​ S606, the final effect of pasting the realistic face image generated in step (5) back to the digitalized garment sample display image in step (4) is obtained by the following operation : wherein, is the digitized garment example display of step (4), denotes an element-wise multiplication. 6.A computer device, comprising a memory and a processor, the memory is configured to store a computer program, and the processor is configured to execute the computer program stored in the memory, characterized in that, When the computer program is executed, the processor performs the method for feature-consistent, realistic face style transfer for displaying digital clothing samples as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that, when executed by a computer device, cause the computer device to perform the method for feature-consistent realistic face style transfer for displaying digital clothing samples as described in any one of claims 1-5.

Images