System and method for generating specified portrait picture
By extracting face and text features, performing multimodal information fusion and inverse diffusion processing, combined with super-resolution technology, the randomness and insufficient details of generating specified face images in Stable Diffusion technology are solved, achieving high quality and stability The specified portrait image generation.
Patent Information
- Application Number
- CN202411308474.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2024-12-13
AI Technical Summary
The existing Stable Diffusion technology cannot accurately generate specified face images, and there are problems of randomness and insufficient details, and it is impossible to effectively use face information and text description to generate realistic images.
By preprocessing face pictures and text, feature vectors are extracted and multimodal information fusion is performed, denoised image feature vectors are generated using the inverse diffusion process, and combined with super-resolution operations, a designated portrait picture is generated. Specific steps include face detection, text encoding, noise injection, multimodal information fusion and image generation, and finally image recovery and super-resolution processing.
It realizes accurate acquisition of face information, reduces the difference between generated images and provided image image elements, improves image details and resolution, reduces noise, and generates specified portrait images more realistic and stable.
Smart Images
Figure CN119131868A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to a system and method for generating a specified portrait picture. Background Art
[0002] In recent years, with the rapid development of science and technology and the continuous progress of society, generative technologies based on large models have made significant progress in the field of image generation. These technologies not only occupy an important position in academic research, but also show great potential and value in practical applications. Among them, Stable Diffusion, as an emerging image generation method, has quickly attracted widespread attention and application with its high efficiency, stability and flexibility, as well as the advantage of continuously updated open source code.
[0003] Stable Diffusion is a generative model based on the diffusion process, which generates high-quality images by gradually denoising and reconstructing the input data. Compared with traditional generative adversarial networks (GANs) and variational autoencoders (VAEs), Stable Diffusion has stronger generation capabilities and higher stability. This enables it to more accurately capture the details and textures of the image when dealing with complex image generation tasks, thereby generating more realistic and delicate images.
[0004] However, existing Stable Diffusion technologies all have the problem of random generation, and cannot specify the generated facial information. Therefore, they can only randomly generate facial images, and cannot generate pictures of specified faces (such as full-body photos, landscape photos, etc.). Since the generation process is implicit, it is impossible to accurately express the desired image by describing it in general terms through text. At the same time, if you want to generate the same person with different styles, you need a lot of text to locate the style. This method is not only time-consuming and labor-intensive, but also the generated results are unstable.
[0005] At the same time, the existing technology does not constrain the details of portrait generation, and there is no way to specify the generated results. This randomness is contained in the generation method. Only by injecting given information into the generation process and extracting details such as portraits can the effect of specifying portrait generation be achieved.
[0006] Therefore, those skilled in the art are committed to developing a system and method for generating a specified portrait picture. Summary of the invention
[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to accurately obtain the information provided by the face image to avoid the image elements of the generated image being too different from those of the provided image and the generated image being insufficient in details.
[0008] Stable Diffusion is a generative model based on the inverse diffusion process, which generates high-quality images by gradually denoising and reconstructing the input data. Specifically, the face image and text are first preprocessed to extract the face feature vector, text feature vector, noise feature vector and reference image feature vector, and then fused to obtain the multimodal information feature vector. The denoised image feature vector is then obtained through the inverse diffusion process, and the denoised image feature vector is decoded and then super-resolution operation is performed to obtain the specified portrait image.
[0009] In one embodiment of the present invention, a system for generating a specified portrait image is provided, comprising:
[0010] The information preprocessing module preprocesses the face image and text, extracts the face feature vector, text feature vector, noise feature vector and reference image feature vector, and sends them to the multimodal information fusion module;
[0011] The multimodal information fusion module receives and guides the face feature vector, the text feature vector, the noise feature vector and the reference image feature vector to fuse with each other, obtains the multimodal information feature vector, and sends it to the image generation module;
[0012] The image generation module receives the multimodal information feature vector, obtains the denoised image feature vector through the inverse diffusion process, and sends it to the post-processing module;
[0013] The post-processing module receives the denoised image feature vector, performs image restoration and super-resolution operations, and obtains a specified portrait image;
[0014] The information preprocessing module, the multimodal information fusion module, the image generation module and the post-processing module are sequentially communicatively connected;
[0015] The information preprocessing module preprocesses the face image and text, extracts the face feature vector, text feature vector, noise feature vector and reference image feature vector, and sends them to the multimodal information fusion module for fusion to obtain the multimodal information feature vector, which is sent to the image generation module to obtain the denoised image feature vector through the inverse diffusion process, and is sent to the post-processing module for super-resolution operation to obtain the specified portrait image.
[0016] Optionally, in the system for generating a designated portrait image in the above embodiment, the information preprocessing module includes:
[0017] The face detection and extraction unit uses the face detection model to determine the face position information, cuts out the individual face, and uses a multi-layer convolutional neural network to extract the face feature vector;
[0018] The text encoder unit uses encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture the semantic and contextual information of the text, and extract text feature vectors;
[0019] The noise injection unit injects random noise into the reference image using a noise injection method to obtain a diffused noise image, which is used as the starting point of the reverse diffusion process to extract the noise feature vector;
[0020] The image encoder unit uses a multi-layer convolutional neural network to extract useful features from the image, converts it into a latent vector space representation, and extracts the reference image feature vector.
[0021] Furthermore, in the system for generating a designated portrait image in the above embodiment, the face detection model is based on a convolutional neural network.
[0022] Furthermore, in the system for generating a designated portrait image in the above embodiment, the face feature vector contains rich face semantic information, including facial details, gender, hairstyle, and retains the identity information of the target person.
[0023] Furthermore, in the system for generating a designated portrait image in the above embodiment, the encoding technology uses the CLIP (Contrastive Language-Image Pre-training) model to perform unsupervised learning on a large-scale text corpus to obtain rich language representation capabilities.
[0024] Furthermore, in the system for generating a specified portrait image in the above embodiment, the CLIP model uses a multi-layer Transformer encoder to encode the embedding vector to generate a text feature vector of a fixed length. The encoding layer captures the relationship between words and context dependencies in the sentence through a self-attention mechanism.
[0025] Furthermore, in the system for generating a designated portrait image in the above embodiment, the random noise adopts standard Gaussian noise, which is injected into the initial image by generating a random vector of Gaussian distribution.
[0026] Furthermore, in the system for generating a designated portrait image in the above embodiment, the noise injection method is to directly add the noise vector to the image pixel value.
[0027] Furthermore, in the system for generating a designated portrait image in the above embodiment, the latent vector space representation will be used to guide the process of generating a new image.
[0028] Optionally, in the system for generating a specified portrait image in any of the above embodiments, the multimodal information fusion module includes:
[0029] The cross attention unit receives and guides the fusion of face feature vector, text feature vector, noise feature vector and reference image feature vector.
[0030] Optionally, in the system for generating a specified portrait image in any of the above embodiments, the image generation module includes:
[0031] The conditional generative network unit uses the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, and obtains a noise image that integrates the text feature vector and the face feature vector;
[0032] The inverse diffusion process unit gradually denoises the noisy image by fusing the text feature vector and the face feature vector to generate a denoised image feature vector.
[0033] Furthermore, in the system for generating a specified portrait picture in the above embodiment, the conditional generation network unit includes a downsampling part and an upsampling part, wherein the downsampling part fuses the text feature vector and the noise image to guide the generation process, and the upsampling part fuses the face feature vector and the noise image processed by the downsampling part to guide the generation process, thereby obtaining a noise image that fuses the text feature vector and the face feature vector.
[0034] Optionally, in the system for generating a specified portrait image in any of the above embodiments, the post-processing module includes:
[0035] The image decoder unit uses a multi-layer convolutional neural network to convert the denoised image feature vector into a real image at the pixel level;
[0036] The super-resolution unit performs super-resolution operations on real images at the pixel level to improve the resolution and clarity, generate high-resolution images from low-resolution images, and obtain specified portrait images.
[0037] Based on any of the above embodiments, in another embodiment of the present invention, a method for generating a specified portrait image is provided, comprising the following steps:
[0038] S100, preprocessing the face image, using a face detection model to determine the face position information, cropping to obtain a separate face, and extracting a face feature vector using a multi-layer convolutional neural network;
[0039] S200, text preprocessing, using encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture the semantic and contextual information of the text, and extract text feature vectors;
[0040] S300, noise injection preprocessing, using a noise injection method to inject random noise into the reference image, obtain a diffused noise image as the starting point of the reverse diffusion process, and extract a noise feature vector;
[0041] S400, preprocessing the reference image, using a multi-layer convolutional neural network to extract useful features in the image, converting it into a latent vector space representation, and extracting a reference image feature vector;
[0042] S500, multimodal feature information fusion, receiving and guiding the face feature vector, the text feature vector, the noise feature vector and the reference image feature vector to fuse with each other, to obtain a multimodal information feature vector;
[0043] S600, acquiring a denoised image feature vector, receiving a multimodal information feature vector, and obtaining a denoised image feature vector through an inverse diffusion process;
[0044] S700, Image Restoration, uses a multi-layer convolutional neural network to convert denoised image feature vectors into true images at the pixel level;
[0045] S800, image post-processing, performing super-resolution operation on the real image to improve the resolution and clarity to obtain a designated portrait image.
[0046] Optionally, in the method for generating a specified portrait image in the above embodiment, step S100 includes:
[0047] S110, face image detection, using a face detection model to determine face position information;
[0048] S120, face cropping, removing other interference information, and cropping the face to obtain a single face;
[0049] S130, extracting facial feature vectors, using a multi-layer convolutional neural network to extract facial feature vectors.
[0050] Furthermore, in the method for generating a specified portrait image in the above embodiment, the face detection model is based on a convolutional neural network.
[0051] Furthermore, in the method for generating a specified portrait image in the above embodiment, the encoding technology uses a CLIP (Contrastive Language-Image Pre-training) model to perform unsupervised learning on a large-scale text corpus to obtain rich language representation capabilities.
[0052] Furthermore, in the method for generating a specified portrait image in the above embodiment, the CLIP model uses a multi-layer Transformer encoder to encode the embedding vector to generate a text feature vector of a fixed length. The encoding layer captures the relationship between words and context dependencies in the sentence through a self-attention mechanism.
[0053] Optionally, in the method for generating a designated portrait image in any of the above embodiments, the noise injection method in step S300 includes using standard Gaussian noise and injecting it into the initial image by generating a random vector with Gaussian distribution.
[0054] Optionally, in the method for generating a specified portrait image in any of the above embodiments, step S500 includes:
[0055] S510, generating a noise image, using the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, and obtains a noise image that integrates the text feature vector and the face feature vector;
[0056] S520, generating a denoised image feature vector, gradually denoising the noisy image that is a fusion of the text feature vector and the face feature vector, to generate a denoised image feature vector.
[0057] Furthermore, in the method for generating a specified portrait image in the above embodiment, step S510 includes:
[0058] S511, downsampling, fusing text feature vector and noise image to guide the generation process;
[0059] S512, up-sampling, fusing the face feature vector and the noise image to guide the generation process, and obtaining a noise image that fuses the text feature vector and the face feature vector.
[0060] The present invention uses a face detection model, a coding technology, a noise injection method, and a multi-layer convolutional neural network to process the face image in sequence, and then performs multi-modal feature information fusion, denoises through an inverse diffusion process, performs image restoration and super-resolution operations, obtains a specified portrait picture, accurately obtains information provided by the face image, avoids a large difference in image elements between the generated image and the provided image and insufficient details of the generated image, reduces noise, and improves image resolution.
[0061] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 is a structural diagram illustrating a system for generating a designated portrait picture according to an exemplary embodiment;
[0063] Figure 2 is a flowchart illustrating a method of generating a designated portrait picture according to an exemplary embodiment. DETAILED DESCRIPTION
[0064] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.
[0065] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is schematically exaggerated in some places in the drawings.
[0066] The inventor has designed a system for generating a specified portrait image, such as Figure 1 As shown, including:
[0067] The information preprocessing module preprocesses the face image and text, extracts the face feature vector, text feature vector, noise feature vector and reference image feature vector, and sends them to the multimodal information fusion module; the information preprocessing module includes:
[0068] The face detection and extraction unit uses a face detection model based on a convolutional neural network to determine the face position information, crop the face to obtain a single face, and use a multi-layer convolutional neural network to extract the face feature vector. The face feature vector contains rich face semantic information, including facial details, gender, hairstyle, and retains the identity information of the target person;
[0069] The text encoder unit uses encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture the semantics and contextual information of the text, and extract text feature vectors. The encoding technology uses the CLIP (Contrastive Language-Image Pre-training) model to perform unsupervised learning on a large-scale text corpus to obtain rich language representation capabilities. The CLIP model uses a multi-layer Transformer encoder to encode the embedded vector and generate a fixed-length text feature vector. The encoding layer uses a self-attention mechanism to capture the relationship between words in a sentence and contextual dependencies.
[0070] The noise injection unit injects random noise into the reference image using a noise injection method.
[0071] The diffused noise image is obtained as the starting point of the reverse diffusion process, and the noise feature vector is extracted; the random noise adopts standard Gaussian noise, which is injected into the initial image by generating a random vector with Gaussian distribution; the noise injection method is to directly add the noise vector to the image pixel value;
[0072] The image encoder unit uses a multi-layer convolutional neural network to extract useful features from the image, converts them into a latent vector space representation, and extracts the reference image feature vector. The latent vector space representation will be used to guide the process of generating new images.
[0073] The multimodal information fusion module receives and guides the face feature vector, text feature vector, noise feature vector and reference image feature vector to fuse with each other, obtains the multimodal information feature vector, and sends it to the image generation module. The multimodal information fusion module includes:
[0074] The cross attention unit receives and guides the fusion of face feature vector, text feature vector, noise feature vector and reference image feature vector.
[0075] The image generation module receives the multimodal information feature vector, obtains the denoised image feature vector through the inverse diffusion process, and sends it to the post-processing module; the image generation module includes:
[0076] The conditional generation network unit uses the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, thereby obtaining a noise image that fuses the text feature vector and the face feature vector; the conditional generation network unit includes a downsampling part and an upsampling part, wherein the downsampling part fuses the text feature vector and the noise image to guide the generation process, and the upsampling part fuses the face feature vector and the noise image processed by the downsampling part to guide the generation process, thereby obtaining a noise image that fuses the text feature vector and the face feature vector;
[0077] The inverse diffusion process unit gradually denoises the noisy image by fusing the text feature vector and the face feature vector to generate a denoised image feature vector.
[0078] The post-processing module receives the denoised image feature vector, performs image restoration and super-resolution operations, and obtains a specified portrait image; the post-processing module includes:
[0079] The image decoder unit uses a multi-layer convolutional neural network to convert the denoised image feature vector into a real image at the pixel level;
[0080] The super-resolution unit performs super-resolution operations on real images at the pixel level to improve the resolution and clarity, generate high-resolution images from low-resolution images, and obtain specified portrait images.
[0081] The information preprocessing module, the multimodal information fusion module, the image generation module and the post-processing module are sequentially communicatively connected;
[0082] The information preprocessing module preprocesses the face image and text, extracts the face feature vector, text feature vector, noise feature vector and reference image feature vector, and sends them to the multimodal information fusion module for fusion to obtain the multimodal information feature vector, which is sent to the image generation module to obtain the denoised image feature vector through the inverse diffusion process, and is sent to the post-processing module for super-resolution operation to obtain the specified portrait image.
[0083] Based on the above embodiments, the inventor provides a method for generating a specified portrait image, such as Figure 2 As shown, the following steps are included:
[0084] S100, preprocessing of face images, using a face detection model based on a convolutional neural network to determine face position information, cropping to obtain a single face, and extracting face feature vectors using a multi-layer convolutional neural network; specifically including:
[0085] S110, face image detection, using a face detection model based on a convolutional neural network to determine face position information;
[0087] S120, face cropping, removing other interference information, and cropping the face to obtain a single face;
[0089] S130, extracting facial feature vectors, using a multi-layer convolutional neural network to extract facial feature vectors.
[0090] S200, text preprocessing, uses encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture the semantic and contextual information of the text, and extract text feature vectors; the encoding technology uses the CLIP (Contrastive Language-Image Pre-training, multimodal pre-training) model to perform unsupervised learning on a large-scale text corpus to obtain rich language representation capabilities. The CLIP model uses a multi-layer Transformer encoder to encode the embedded vector and generate a fixed-length text feature vector. The encoding layer uses a self-attention mechanism to capture the relationship between words in a sentence and contextual dependencies.
[0091] S300, noise injection preprocessing, using a noise injection method to inject random noise into the reference image, obtain a diffused noise image, use it as the starting point of the inverse diffusion process, and extract the noise feature vector. The noise injection method includes using standard Gaussian noise and injecting it into the initial image by generating a random vector with Gaussian distribution.
[0092] S400, reference image preprocessing, using a multi-layer convolutional neural network to extract useful features in the image, converting it into a latent vector space representation, and extracting a reference image feature vector.
[0093] S500, multimodal feature information fusion, receiving and guiding the face feature vector, text feature vector, noise feature vector and reference image feature vector to fuse with each other, to obtain a multimodal information feature vector; specifically including:
[0094] S510, generating a noise image, using the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, and obtains a noise image that integrates the text feature vector and the face feature vector; specifically including:
[0095] S511, downsampling, fusing text feature vector and noise image to guide the generation process;
[0096] S512, up-sampling, fusing the face feature vector and the noise image to guide the generation process, and obtaining a noise image that fuses the text feature vector and the face feature vector.
[0097] S520, generating a denoised image feature vector, gradually denoising the noisy image that is a fusion of the text feature vector and the face feature vector, and generating a denoised image feature vector.
[0098] S600, acquiring a denoised image feature vector, receiving a multimodal information feature vector, and obtaining a denoised image feature vector through an inverse diffusion process.
[0099] S700, Image Restoration, uses a multi-layer convolutional neural network to convert the denoised image feature vector into a true image at the pixel level.
[0100] S800, image post-processing, performing super-resolution operation on the real image to improve the resolution and clarity, and obtain a designated portrait image.
[0101] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A system for generating a specified portrait image, characterized in that: include: The information preprocessing module preprocesses the face image and text, extracts the face feature vector, text feature vector, noise feature vector and reference image feature vector, and sends them to the multimodal information fusion module; A multimodal information fusion module receives and guides the face feature vector, the text feature vector, the noise feature vector and the reference image feature vector to fuse with each other, obtains a multimodal information feature vector, and sends it to an image generation module; An image generation module receives the multimodal information feature vector, obtains a denoised image feature vector through an inverse diffusion process, and sends the denoised image feature vector to a post-processing module; A post-processing module receives the denoised image feature vector, performs image restoration and super-resolution operations, and obtains a specified portrait image; The information preprocessing module, the multimodal information fusion module, the image generation module and the post-processing module are communicatively connected in sequence.
2. The system for generating a designated portrait image according to claim 1, characterized in that: The information preprocessing module comprises: A face detection and extraction unit, using a face detection model to determine face position information, cropping to obtain a single face, and using a multi-layer convolutional neural network to extract the face feature vector; A text encoder unit, using encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture semantic and contextual information of the text, and extract the text feature vectors; A noise injection unit, which uses a noise injection method to inject random noise into the reference image, obtains a diffused noise image as the starting point of the reverse diffusion process, and extracts the noise feature vector; The image encoder unit uses a multi-layer convolutional neural network to extract useful features in the image, converts it into a latent vector space representation, and extracts the reference image feature vector.
3. The system for generating a designated portrait image as claimed in claim 2, characterized in that: The face detection model is based on a convolutional neural network.
4. The system for generating a designated portrait image according to claim 1, wherein: The encoding technology uses the CLIP model to perform unsupervised learning on a large-scale text corpus to obtain rich language representation capabilities.
5. The system for generating a designated portrait image according to claim 1, characterized in that: The multimodal information fusion module includes: A cross attention unit receives and guides the face feature vector, the text feature vector, the noise feature vector and the reference image feature vector to be merged with each other.
6. The system for generating a designated portrait image according to claim 1, characterized in that: The image generation module comprises: A conditional generative network unit uses the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, thereby obtaining a noise image that fuses the text feature vector and the face feature vector; The inverse diffusion process unit gradually denoises the noise image of the fused text feature vector and the face feature vector to generate a denoised image feature vector.
7. The system for generating a designated portrait image according to claim 1, characterized in that: The post-processing module comprises: An image decoder unit, which converts the denoised image feature vector into a real image at the pixel level using a multi-layer convolutional neural network; The super-resolution unit performs a super-resolution operation on the real image at the pixel level to improve the resolution and clarity, and generates a high-resolution image from the low-resolution image to obtain a designated portrait image.
8. A method for generating a specified portrait image, using the system for generating a specified portrait image as claimed in any one of claims 1 to 6, characterized in that: The steps include: S100, preprocessing the face image, using a face detection model to determine the face position information, cropping to obtain a separate face, and extracting a face feature vector using a multi-layer convolutional neural network; S200, text preprocessing, using encoding technology to convert natural language descriptions into high-dimensional semantic feature vectors, capture the semantic and contextual information of the text, and extract text feature vectors; S300, noise injection preprocessing, using a noise injection method to inject random noise into the reference image, obtain a diffused noise image as the starting point of the reverse diffusion process, and extract a noise feature vector; S400, preprocessing the reference image, using a multi-layer convolutional neural network to extract useful features in the image, converting it into a latent vector space representation, and extracting a reference image feature vector; S500, multimodal feature information fusion, receiving and guiding the face feature vector, the text feature vector, the noise feature vector and the reference image feature vector to fuse with each other to obtain a multimodal information feature vector; S600, acquiring a denoised image feature vector, receiving the multimodal information feature vector, and obtaining a denoised image feature vector through an inverse diffusion process; S700, image restoration, using a multi-layer convolutional neural network to convert the denoised image feature vector into a real image at the pixel level; S800 , image post-processing, performing super-resolution operation on the real image to improve the resolution and clarity, and obtaining a designated portrait image.
9. The method for generating a designated portrait image according to claim 8, characterized in that: The step S100 includes: S110, face image detection, using the face detection model to determine the face position information; S120, face cropping, removing other interference information, and performing face cropping to obtain the single face; S130, extracting facial feature vectors, using the multi-layer convolutional neural network to extract the facial feature vectors.
10. The method for generating a designated portrait image according to claim 8, characterized in that: The step S500 includes. S510, generating a noise image, using the text feature vector and the face feature vector to guide the generation process, so that the generated image is consistent with the input text description and contains the specified face information, and obtains a noise image that fuses the text feature vector and the face feature vector; S520, generating a denoised image feature vector, denoising the noisy image by step by step, by merging the text feature vector and the face feature vector, to generate a denoised image feature vector.