Visual Asset Development Using Generative Adversarial Networks
A GAN trained on 3D digital representations addresses the high cost and uniformity issues in video game asset creation by generating diverse and visually consistent content, reducing production costs and enabling high-fidelity game development.
Patent Information
- Application Number
- JP2022574632
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2020-06-04
- Publication Date
- 2024-12-04
- Estimated Expiration
- 2040-06-04
AI Technical Summary
The high cost and resource-intensive process of creating visually consistent visual assets for video games, particularly in massively multiplayer online games, is exacerbated by the need for frequent content updates and the limitations of existing procedural content generation methods, which often result in uniform and repetitive outputs.
A computer-implemented method using a generative adversarial network (GAN) trained on 3D digital representations of visual assets, capturing images from various viewpoints and lighting conditions, to generate diverse and visually consistent content by iteratively refining a generator and discriminator based on loss functions.
The method enables efficient and diverse generation of visually consistent visual assets, reducing production costs and enabling small studios to create high-fidelity game content, while maintaining the desired visual identity of the game.
Smart Images

Figure 0007594611000001 
Figure 0007594611000002 
Figure 0007594611000003
Abstract
Description
Summary of the Invention [Problem to be solved by the invention]
[0001] background A significant portion of the budget and resources allocated to the production of video games is consumed by the process of creating the visual assets of the video game. For example, massively multiplayer online games contain thousands of player avatars and non-player characters (NPCs), which are typically created using three-dimensional (3D) templates that are manually customized during the development of the game to create individualized characters. In another example, the environment or context of a scene in a video game often contains a large number of virtual objects, such as trees, rocks, clouds, etc. These virtual objects are manually customized to avoid excessive repetition or homogeneity, as may occur if a forest contains hundreds of identical trees or repeating patterns of groups of trees. Procedural content generation has been used to generate characters and objects, but the content generation process is difficult to control and often produces output that is visually uniform, homogenous, or repetitive. The high cost of generating visual assets for video games drives up video game budgets and increases risk aversion on the part of video game creators. Additionally, the cost of content generation creates a significant barrier to entry for smaller studios (and correspondingly smaller budgets) looking to enter the market for high fidelity game design. Additionally, video game players, especially online players, have come to expect frequent content updates, which further exacerbates the problems associated with the high production costs of video assets. [Means for solving the problem]
[0002] overview The proposed solution particularly relates to a computer-implemented method comprising the steps of: capturing a first image of a three-dimensional (3D) digital representation of a visual asset; generating second images representing variations of the visual asset using a generator in a Generative Adversarial Network (GAN) and attempting to distinguish between the first and second images in a classifier in the GAN; updating at least one of a first model in the classifier and a second model in the generator based on whether the classifier successfully distinguishes between the first and second images; and generating a third image using the generator based on the updated second model. The first model is used by the generator as a basis for generating the second image, whereas the second model is used by the classifier as a basis for evaluating the generated second image. The variation of the first image generated by the generator may in particular relate to a variation of at least one image parameter of the first image, such as a variation of at least one or all pixel or texel values of the first image. Thus, the variation by the generator may relate for example to a variation of at least one of color, brightness, texture, granularity, or a combination thereof.
[0003] Machine learning has been used to generate images, for example using neural networks trained on image databases. One approach to image generation used in this context uses a machine learning architecture known as a Generative Adversarial Network (GAN), which uses a pair of interacting Convolutional Neural Networks (CNNs) to learn how to create different types of images. The first CNN (the generator) creates new images that correspond to images in the training dataset, and the second CNN (the discriminator) tries to distinguish between the generated images and "real" images from the training dataset. In some cases, the generator generates images based on hints and / or random noise that guide the image generation process, in which case the GAN is referred to as a Conditional GAN (CGAN). In general, a "hint" in this context may be a parameter that includes, for example, image content characterization in a computer-readable format. Examples of hints include labels associated with the image, shape information such as the outline of an animal or object, etc. The generator and discriminator then compete based on the images generated by the generator. The generator "wins" if the classifier classifies a generated image as a real image (or vice versa), and the classifier "wins" if it correctly classifies a generated image and a real image. The generator and the classifier may update their respective models based on a loss function that encodes victory or defeat as a "distance" from the correct model. The generator and the classifier continue to improve their respective models based on the results generated by other CNNs.
[0004] The generator in a trained GAN generates images that attempt to mimic the characteristics of people, animals, or objects in a training dataset. As described above, the generator in a trained GAN can generate images based on hints. For example, in response to receiving a hint that includes the label "bear," the trained GAN attempts to generate images that resemble bears. However, the images generated by the trained GAN are determined (at least in part) by the characteristics of the training dataset, which may fail to reflect the desired characteristics of the generated images. For example, video game designers often create visual identities for their games using fantasy or science fiction styles characterized by striking perspectives, image compositions, and lighting effects. In contrast, traditional image databases include real-world photos of a variety of different people, animals, or objects taken in a variety of environments under different lighting conditions. Furthermore, datasets of photographed faces are often pre-processed to include a limited number of perspectives that are rotated to ensure that the faces are not tilted or retouched by applying Gaussian blur to the background. Thus, GANs trained on traditional image databases are unable to generate images that maintain the visual identity created by the game designer. For example, images that mimic people, animals, or objects in real-world photographs would disrupt the visual consistency of scenes generated in a fantasy or science fiction style. Furthermore, large repositories of illustrations that could be used to train GANs are subject to issues of ownership, style conflicts, or simply lack the diversity necessary to build robust machine learning models.
[0005] Thus, the proposed solution provides a hybrid procedural pipeline for generating diverse and visually consistent content by training a generator and a classifier of a conditional generative adversarial network (CGAN) using images captured from a three-dimensional (3D) digital representation of a visual asset. The 3D digital representation includes a model of the 3D structure of the visual asset, and possibly includes textures applied to the surface of the model. For example, a 3D digital representation of a bear can be represented by a set of triangles, other polygons or patches, collectively referred to as primitives, and textures applied to the primitives to incorporate visual details having a higher resolution than that of the primitives, such as fur, teeth, claws and eyes. Training images ("first images") are captured using a virtual camera that captures images from different perspectives and possibly under different lighting conditions. Capturing training images of the 3D digital representation of the visual asset can result in an improved training dataset, resulting in diverse and visually consistent content composed of different second images that can be used separately or in combination in the 3D representation of the modified visual asset in the video game. Capturing training images ("first images") by the virtual camera may include capturing a set of training images associated with various viewpoints or lighting conditions of the 3D representation of the visual asset. The number of training images in the training set or the viewpoints or lighting conditions are predetermined by a user or an image capture algorithm. For example, at least one of the number of training images in the training set, the viewpoints and the lighting conditions may be preset or may be contingent on the visual asset for which the training images are captured. This includes, for example, that the capture of the training images may be performed automatically after loading the visual asset into an image capture system and / or after initiating an image capture process implementing the virtual camera.
[0006] The image capture system may also apply labels to the captured images, including labels indicating the type of object (e.g., bear), camera position, camera pose, lighting conditions, texture, color, etc. In some embodiments, the images are segmented into various portions of the visual asset, such as the animal's head, ears, neck, legs, and arms. The segmented portions of the image may be labeled to indicate various parts of the visual asset. The labeled images may be stored in a training database.
[0007] By training the GAN, the generator and the discriminator learn the distribution of parameters that represent the images in the training database generated from the 3D digital representation. That is, the GAN is trained using the images in the training database. First, the discriminator is trained to identify "real" images of the 3D digital representation based on the images in the training database. The generator then starts to generate (second) images in response to hints, such as, for example, a label or a digital representation of the outline of the visual asset. The generator and the discriminator can then iteratively and simultaneously update their corresponding models based on, for example, a loss function that indicates how well the generator generates images representing the visual asset (e.g., how well it "fools" the discriminator) and how well the discriminator distinguishes between the generated images and real images from the training database. The generator models the distribution of parameters in the training images, and the discriminator models the distribution of parameters inferred by the generator. Thus, the first model of the generator can include the distribution of parameters in the first image, and the second model of the discriminator includes the distribution of parameters inferred by the generator.
[0008] In some embodiments, the loss function includes a perceptual loss function that uses another neural network to extract features from the images and encodes the difference between the two images as a distance between the extracted features. In some embodiments, the loss function may receive a classification decision from the classifier. The loss function may also receive information indicating the identity (or at least the real or fake status) of the second image provided to the classifier. The loss function may then generate a classification error based on the received information. The classification error represents how well the generator and the classifier achieve their respective goals.
[0009] Once trained, the GAN is used to generate images representing visual assets based on the distribution of parameters inferred by the generator. In some embodiments, these images are generated in response to hints. For example, a trained GAN may generate an image of a bear in response to receiving a hint that includes the label "bear" or a representation of the outline of a bear. In some embodiments, these images are generated based on a composite of segmented portions of the visual assets. For example, a chimera may be generated by combining segments of images (indicated by their respective labels) representing different creatures, such as the head, torso, legs and tail of a dinosaur and the wings of a bat.
[0010] In some embodiments, at least one third image may be generated in the generator in the GAN to represent a variation of the visual asset based on the first model. And, generating the at least one third image may include, for example, generating the at least one third image based on at least one of a label associated with the visual asset or a digital representation of an outline of a portion of the visual asset. Alternatively or additionally, generating the at least one third image may include generating the at least one third image by combining at least one segment of the visual asset with at least one segment of another visual asset.
[0011] The proposed solution further relates to a system comprising a memory configured to store a first image captured from a three-dimensional (3D) digital representation of a visual asset, and at least one processor configured to realize a generative adversarial network (GAN) comprising a generator and a classifier, the generator configured to generate a second image representing a variation of the visual asset, for example at the same time as the classifier attempts to distinguish between the first image and the second image, and the at least one processor configured to update at least one of a first model in the classifier and a second model in the generator based on whether the classifier successfully distinguishes between the first image and the second image.
[0012] The proposed system may be particularly configured to implement embodiments of the proposed method.
[0013] The present disclosure may be better understood, and its numerous features and advantages made apparent to those skilled in the art by referencing the accompanying drawings, in which: The use of the same reference symbols in different drawings indicates similar or identical elements. [Brief description of the drawings]
[0014] [Figure 1] FIG. 1 is a block diagram of a video game processing system implementing a hybrid procedural machine language (ML) pipeline for technology development according to some embodiments. [Diagram 2] FIG. 1 is a block diagram of a cloud-based system implementing a hybrid procedural ML pipeline for technology development according to some embodiments. [Diagram 3] FIG. 1 is a block diagram of an image capture system for capturing images of a digital representation of a visual asset according to some embodiments. [Figure 4]1 is a block diagram of an image of a visual asset and labeled data representing the visual asset according to some embodiments. [Diagram 5] FIG. 1 is a block diagram of a generative adversarial network (GAN) trained to generate images that are variations of a visual asset, according to some embodiments. [Figure 6] FIG. 1 is a flow diagram of a method for training a GAN to generate image variations of a visual asset, according to some embodiments. [Figure 7] FIG. 1 illustrates the evolution of a ground truth distribution of parameters characterizing images of a visual asset and the distribution of corresponding parameters generated by a generator in a GAN, according to some embodiments. [Figure 8] FIG. 1 is a block diagram of a portion of a GAN trained to generate images that are variations of a visual asset, according to some embodiments. [Figure 9] FIG. 2 is a flow diagram of a method for generating variations of an image of a visual asset according to some embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0015] Detailed Description 1 is a block diagram of a video game processing system 100 implementing a hybrid procedural machine (ML) pipeline for technology development, according to some embodiments. The processing system 100 includes or has access to a system memory 105 or other storage elements implemented using a non-transitory computer-readable medium, such as Dynamic Random-Access Memory (DRAM). However, some embodiments of the memory 105 are implemented using other types of memory, including static RAM (SRAM), non-volatile RAM, and the like. The processing system 100 also includes a bus 110 for supporting communication between entities implemented in the processing system 100, such as the memory 105. Some embodiments of the processing system 100 include other buses, bridges, switches, routers, and the like, which are not shown in FIG. 1 for clarity.
[0016] The processing system 100 includes a central processing unit (CPU) 115. Some embodiments of the CPU 115 include multiple processing elements (not shown in FIG. 1 for clarity) that execute instructions simultaneously or in parallel. These processing elements may be referred to as processor cores, compute units, or using other terminology. The CPU 115 is connected to a bus 110, through which the CPU 115 communicates with a memory 105. The CPU 115 executes instructions, such as program code 120, stored in the memory 105, and the CPU 115 stores information, such as results of executed instructions, in the memory 105. The CPU 115 may also initiate graphics processing by issuing draw calls.
[0017] The input / output (I / O) engine 125 processes input or output operations associated with the display 130, which displays images or videos on the screen 135. In the illustrated embodiment, the I / O engine 125 is connected to the game controller 140, which provides control signals to the I / O engine 125 in response to a user pressing one or more buttons on the game controller 140 or interacting with the game controller 140 in other ways, such as using movements detected by an accelerometer. The I / O engine 125 also provides signals to the game controller 140 for activating responses in the game controller 140, such as vibrations, turning on lights, etc. In the illustrated embodiment, the I / O engine 125 reads information stored in an external storage element 145, which is implemented using a non-transitory computer-readable medium, such as a compact disk (CD), a digital video disk (DVD), etc. The I / O engine 125 also writes information, such as results of processing by the CPU 115, to the external storage element 145. Some embodiments of I / O engine 125 are coupled to other elements of processing system 100, such as a keyboard, a mouse, a printer, an external disk, etc. I / O engine 125 is coupled to bus 110 such that I / O engine 125 communicates with memory 105, CPU 115, or other entities connected to bus 110.
[0018] The processing system 100 includes a graphics processing unit (GPU) 150 that renders and displays images on the screen 135 of the display 130, for example by controlling the pixels that make up the screen 135. For example, the GPU 150 renders objects to generate pixel values that are provided to the display 130, which uses these pixel values to display images representing the rendered objects. The GPU 150 includes one or more processing elements, such as an array of compute units 155, that execute instructions simultaneously or in parallel. Some embodiments of the GPU 150 are used for general-purpose computing. In the embodiment shown, the GPU 150 communicates with the memory 105 (and other entities connected to the bus 110) via the bus 110. However, some embodiments of the GPU 150 communicate with the memory 105 via a direct connection or via other buses, bridges, switches, routers, etc. The GPU 150 executes instructions stored in the memory 105, and the GPU 150 stores information in the memory 105, such as results of executed instructions. For example, memory 105 stores instructions representing program code 160 to be executed by GPU 150 .
[0019] In the illustrated embodiment, CPU 115 and GPU 150 execute corresponding program code 120, 160 to implement a video game application. For example, user input received via game controller 140 is processed by CPU 115 to change the state of the video game application. CPU 115 then sends draw calls to instruct GPU 150 to render and display images representing the state of the video game application on screen 135 of display 130. As described herein, GPU 150 may also perform general-purpose computing related to video games, such as running a physics engine or machine learning algorithms.
[0020] The CPU 115 or GPU 150 also executes program code 165 for implementing a hybrid procedural machine language (ML) pipeline for technology development. The hybrid procedural ML pipeline includes a first portion that captures images 170 of a three-dimensional (3D) digital representation of a visual asset from various viewpoints and possibly under various lighting conditions. In some embodiments, a virtual camera captures the first or training images of the 3D digital representation of the visual asset from various viewpoints and / or under various lighting conditions. The images 170 may be captured by the virtual camera automatically, i.e., based on an image capture algorithm included in the program code 165. The images 170 captured by the first portion of the hybrid procedural ML pipeline, e.g., the portion including the model and the virtual camera, are stored in the memory 105. The visual asset into which the images 170 are captured may be generated by a user (e.g., using a computer-aided design tool) and stored in the memory 105.
[0021] A second part of the hybrid procedural ML pipeline includes a generative adversarial network (GAN) represented by program code and associated data (e.g., model parameters), shown by box 175. The GAN 175 includes a generator and a discriminator implemented as different neural networks. The generator generates a second image representing a variation of the visual asset at the same time that the discriminator attempts to distinguish between the first and second images. Parameters defining the ML model in the discriminator or generator are updated based on whether the discriminator successfully distinguishes between the first and second images. Parameters defining the model implemented in the generator determine the distribution of parameters in the training images 170. Parameters defining the model implemented in the discriminator determine the distribution of parameters inferred by the generator, for example, based on the model of the generator.
[0022] The GAN 175 is trained to generate different versions of the visual asset based on hints or random noise provided to the trained GAN 175, in which case the trained GAN 175 may be referred to as a conditional GAN. For example, if the GAN 175 is trained based on a set of images 170 of a digital representation of a red dragon, the generator in the GAN 175 generates images representing variations of the red dragon (e.g., a blue dragon, a green dragon, a bigger dragon, a smaller dragon, etc.). The images or training images 170 generated by the generator are selectively provided to a discriminator (e.g., by randomly selecting between the training images 170 and the generated images), and the discriminator attempts to distinguish between the "real" training images 170 and the "fake" images generated by the generator. Parameters of the models implemented in the generator and discriminator are then updated based on a loss function whose value is determined based on whether the discriminator successfully distinguished between the real and fake images. In some embodiments, the loss function also includes a perceptual loss function that uses another neural network to extract features from the real and fake images and encodes the differences between the two images as distances between the extracted features.
[0023] Once trained, the generator in the GAN 175 generates variations of training images, which are used to generate images or videos for a video game. The processing system 100 shown in FIG. 1 performs image capture, GAN model training, and subsequent image generation using the trained model, although these operations are performed using other processing systems in some embodiments. For example, a first processing system (configured in a manner similar to the processing system 100 shown in FIG. 1) can perform image capture and store images of visual assets in a memory accessible to a second processing system, or transmit these images to a second processing system. The second processing system can perform model training of the GAN 175 and store parameters defining the trained model in a memory accessible to a third processing system, or transmit these parameters to a third processing system. The third processing system can then be used to generate images or videos for a video game using the trained model.
[0024] FIG. 2 is a block diagram of a cloud-based system 200 implementing a hybrid procedural ML pipeline for technology development, according to some embodiments. The cloud-based system 200 includes a server 205 interconnected with a network 210. Although one server 205 is shown in FIG. 2, some embodiments of the cloud-based system 200 include two or more servers connected to the network 210. In the embodiment shown, the server 205 includes a transceiver 215 that transmits signals toward and receives signals from the network 210. The transceiver 215 can be implemented using one or more separate transmitters and receivers. The server 205 also includes one or more processors 220 and one or more memories 225. The processor 220 executes instructions, such as program code, stored in the memory 225, and the processor 220 stores information, such as results of executed instructions, in the memory 225.
[0025] The cloud-based system 200 includes one or more processing devices 230, such as computers, set-top boxes, gaming consoles, etc., connected to a server 205 via a network 210. In the embodiment shown, the processing device 230 includes a transceiver 235 that transmits signals toward and receives signals from the network 210. The transceiver 235 may be implemented using one or more separate transmitters and receivers. The processing device 230 also includes one or more processors 240 and one or more memories 245. The processor 240 executes instructions, such as program code, stored in the memory 245, and the processor 240 stores information, such as results of executed instructions, in the memory 245. The transceiver 235 is connected to a display 250 that displays images or video on a screen 255, a game controller 260, and other text or voice input devices. Thus, some embodiments of the cloud-based system 200 are used by a cloud-based game streaming application.
[0026] The processor 220, the processor 240, or a combination thereof executes program code for performing image capture, GAN model training, and subsequent image generation using the trained model. The division of labor between the processor 220 in the server 205 and the processor 240 in the processing device 230 varies in different embodiments. For example, the server 205 may train a GAN using images captured by a remote video capture processing system and provide parameters that define a model in the trained GAN to the processor 220 via the transceivers 215, 235. The processor 220 may then use the trained GAN to generate images or videos that are variations of the visual assets used to capture the training images.
[0027] 3 is a block diagram of an image capture system 300 for capturing images of a digital representation of a visual asset, according to some embodiments. The image capture system 300 is implemented using some embodiments of the processing system 100 shown in FIG. 1 and the processing system 200 shown in FIG.
[0028] The image capture system 300 includes a controller 305 implemented using one or more processors, memories or other circuits. The controller 305 is connected to a virtual camera 310 and a virtual light source 315, although not all connections are shown in FIG. 3 for clarity. The image capture system 300 is used to capture images of a visual asset 320 that is displayed as a digital 3D model. In some embodiments, the 3D digital representation of the visual asset 320 (in this example, a dragon) is represented by a set of triangles, other polygons or patches collectively referred to as primitives, and textures that are applied to the primitives to incorporate visual details that have a higher resolution than that of the primitives, such as the texture and color of the dragon's head, claws, wings, teeth, eyes and tail. The controller 305 selects a position, orientation or pose of the virtual camera 310, such as the three positions of the virtual camera 310 shown in FIG. 3. The controller 305 also selects the luminosity, direction, color and other characteristics of the light generated by the virtual light source 315 to illuminate the visual asset 320. Different light features or characteristics are used at different exposures of the virtual camera 310 to generate different images of the visual asset 320. Selection of the position, orientation or pose of the virtual camera 310 and / or selection of the luminosity, direction, color and other characteristics of the light generated by the virtual light source 315 may be based on user selection or may be determined automatically by image capture algorithms executed by the image capture system 300.
[0029] The controller 305 labels the images (e.g., by generating metadata associated with the images) and stores them as labeled images 325. In some embodiments, these images are labeled with metadata indicating the type of visual asset 320 (e.g., dragon), the position of the virtual camera 310 when the image was captured, the pose of the virtual camera 310 when the image was captured, the lighting conditions produced by the light source 315, the texture applied to the visual asset 320, the color of the visual asset 320, etc. In some embodiments, these images are segmented into different portions of the visual asset 320 that indicate different parts of the visual asset 320 that may be modified in the proposed technology development process, such as the head, claws, wings, teeth, eyes, and tail of the visual asset 320. The segmented portions of the images are labeled to indicate different parts of the visual asset 320.
[0030] 4 is a block diagram of an image 400 of a visual asset and labeled data 405 representing the visual asset, according to some embodiments. The image 400 and labeled data 405 are generated by some embodiments of the image capture system 300 shown in FIG. 3. In the embodiment shown, the image 400 is an image of a visual asset including a bird in flight. The image 400 is segmented into various parts including a head 410, a beak 415, wings 420, 421, a body 425, and a tail 430. The labeled data 405 represents the image 400 and the associated label "bird." The labeled data 405 includes the image 400For example, labeled data 405 includes image portion 410 and an associated label of "head," image portion 415 and an associated label of "beak," image portion 420 and an associated label of "wing," image portion 421 and an associated label of "wing," image portion 425 and an associated label of "body," and image portion 430 and an associated label of "tail."
[0031] In some embodiments, image portions 410, 415, 420, 421, 425, and 430 are used to train a GAN to create corresponding portions of other visual assets. For example, image portion 410 is used to train a generator of a GAN to create a "head" of another visual asset. Training of the GAN with image portion 410 is performed in conjunction with training of the GAN with other image portions that correspond to the "head" of one or more other visual assets.
[0032] 5 is a block diagram of a GAN 500 trained to generate images that are variations of a visual asset, according to some embodiments. The GAN 500 is implemented in some embodiments of the processing system 100 shown in FIG. 1 and the cloud-based system 200 shown in FIG.
[0033] The GAN 500 includes a generator 505, implemented using a neural network 510, that generates images based on a model distribution of parameters. Some embodiments of the generator 505 generate images based on input information such as random noise 515, hints in the form of labels or outlines of visual assets 520, etc. The GAN 500 also includes a discriminator 525, implemented using a neural network 530, that attempts to distinguish between images generated by the generator 505 and labeled images 535 of the visual assets that represent ground truth images. Thus, the discriminator 525 receives either an image generated by the generator 505 or one of the labeled images 535, and outputs a classification decision 540, which indicates whether the discriminator 525 believes the received image to be a (fake) image generated by the generator 505 or a (real) image from the set of labeled images 535.
[0034] The loss function 545 receives the classification decision 540 from the discriminator 525. The loss function 545 also receives information indicative of the identity (or at least the real or fake status) of the corresponding image provided to the discriminator 525. The loss function 545 then generates a classification error based on the received information. This classification error represents how well the generator 505 and the discriminator 525 achieve their respective goals. In the illustrated embodiment, the loss function 545 also includes a perceptual loss function 550 that extracts features from the real and fake images and encodes the difference between the real and fake images as a distance between the extracted features. The perceptual loss function 550 is realized using a neural network 555 that is trained based on the labeled images 535 and the images generated by the generator 505. Thus, the perceptual loss function 550 contributes to the overall loss function 545.
[0035] The goal of the generator 505 is to fool the discriminator 525, i.e. to have the discriminator 525 identify a (fake) generated image as a (real) image extracted from the labeled image 535, and to have the discriminator 525 identify a real image as a fake image. The model parameters of the neural network 510 are therefore trained to maximize the classification error (between real and fake images), represented by the loss function 545. The goal of the discriminator 525 is to correctly distinguish between real and fake images. The model parameters of the neural network 530 are therefore trained to minimize the classification error, represented by the loss function 545. The training of the generator 505 and the discriminator 525 is performed iteratively, and the parameters defining their corresponding models are updated during each iteration. In some embodiments, a gradient ascent method is used to update the parameters defining the model implemented in the generator 505 to increase the classification error. Gradient descent is used to update the parameters that define the model implemented in the classifier 525 so as to reduce the classification error.
[0036] 6 is a flow diagram of a method 600 for training a GAN to generate image variations of a visual asset according to some embodiments. The method 600 is implemented in some embodiments of the processing system 100 shown in FIG. 1, the cloud-based system 200 shown in FIG. 2, and the GAN 500 shown in FIG.
[0037] In block 605, a first neural network implemented in a GAN classifier is first trained to identify images of visual assets using a set of labeled images captured from those visual assets. Some embodiments of the labeled images are captured by the image capture system 300 shown in FIG.
[0038] At block 610, a second neural network implemented in a generator of the GAN generates images representing variations of the visual asset. In some embodiments, the images are generated based on input random noise, hints, or other information. At block 615, either the generated images or images selected from the set of labeled images are provided to a classifier. In some embodiments, the GAN randomly selects between the (fake) generated images and the (real) labeled images provided to the classifier.
[0039] At decision block 620, the classifier attempts to distinguish between real and fake images received from the generator. The classifier makes a classification decision indicating whether the classifier identifies the image as real or fake and provides the classification decision to a loss function, which determines whether the classifier correctly identified the image as real or fake. If the classification decision from the classifier is correct, method 600 flows to block 625. If the classification decision from the classifier is incorrect, method 600 flows to block 630.
[0040] At block 625, model parameters defining the model distribution used by the first neural network in the generator are updated to reflect the fact that the image generated by the generator did not successfully fool the discriminator. At block 630, model parameters defining the model distribution used by the second neural network in the discriminator are updated to reflect the fact that the discriminator did not correctly identify whether the received image was real or fake. Although the method 600 shown in FIG. 6 illustrates the model parameters in the generator and discriminator being updated independently, some embodiments of GANs update the model parameters of the generator and discriminator simultaneously based on a loss function determined in response to the discriminator providing a classification decision.
[0041] At decision block 635, the GAN determines whether the training of the generator and the discriminator has converged. Convergence is evaluated based on the magnitude of change in the parameters of the models implemented in the first and second neural networks, the fractional change in the parameters, the rate of change of the parameters, a combination thereof, or other criteria. If the GAN determines that the training has converged, the method 600 flows to block 640, where the method 600 ends. If the GAN determines that the training has not converged, the method 600 flows to block 610, where another iteration is performed. Although each iteration of the method 600 is performed on one (real or fake) image, some embodiments of the method 600 provide multiple real and fake images to the discriminator in each iteration, and then update the loss function and model parameters based on the classification decisions returned by the discriminator for these multiple images.
[0042] 7 illustrates the evolution of a ground truth distribution of parameters characterizing images of a visual asset and a corresponding distribution of parameters generated by a generator in a GAN, according to some embodiments. These distributions are shown for three successive time intervals 701, 702, 703, which correspond to successive iterations of training a GAN, for example according to method 600 shown in FIG. 6. The values of parameters corresponding to labeled images (real images) captured from the visual asset are indicated by open circles 705, although for clarity only one is indicated by a reference number in each of time intervals 701-703.
[0043] In the first time interval 701, the values of the parameters corresponding to the images (fake images) generated by the generator in the GAN are indicated by black circles 710, but only one is indicated by a reference number for clarity. The distribution of the parameters 710 of the fake images is significantly different from the distribution of the parameters 705 of the real images. Therefore, the chances that the classifier in the GAN will successfully distinguish between real and fake images are large during the first time interval 710. Therefore, the neural network realized in the generator is updated to improve its ability to generate fake images that fool the classifier.
[0044] In the second time interval 702, the values of the parameters corresponding to the images generated by the generator are indicated by black circles 715, but only one is indicated by a reference number for clarity. The distribution of the parameters 715 representing the fake images is more similar to the distribution of the parameters 705 representing the real images, which means that the neural network in the generator is successfully trained. However, the distribution of the parameters 715 of the fake images is still significantly different (albeit not much) from the distribution of the parameters 705 of the real images. Therefore, the probability that the classifier in the GAN will successfully distinguish between real and fake images is large during the second time interval 702. Again, the neural network realized in the generator is updated to improve its ability to generate fake images that fool the classifier.
[0045] In the third time interval 703, the values of the parameters corresponding to the images generated by the generator are indicated by black circles 720, but only one is indicated by a reference number for clarity. Here, the distribution of parameters 720 representing fake images is nearly indistinguishable from the distribution of parameters 705 representing real images, which means that the neural network in the generator is successfully trained. Thus, the chances of the discriminator in the GAN successfully discriminating between real and fake images are small during the third time interval 703. Thus, the neural network realized in the generator has converged to a model distribution for generating variations of visual assets.
[0046] FIG. 8 is a block diagram of a portion 800 of a GAN trained to generate images that are variations of a visual asset, according to some embodiments. The portion 800 of the GAN is implemented in some embodiments of the processing system 100 shown in FIG. 1 and the cloud-based system 200 shown in FIG. 2. The portion 800 of the GAN includes a generator 805 implemented using a neural network 810 that generates images based on a model distribution of parameters. As described herein, the model distribution of parameters has been trained based on a set of labeled images captured from the visual asset. The trained neural network 810 is used to generate images or videos 815 that represent variations of the visual asset, for use by, for example, a video game. Some embodiments of the generator 805 generate images based on input information such as random noise 820, hints 825 in the form of labels or outlines of the visual asset, etc.
[0047] 9 is a flow diagram of a method 900 for generating variations of an image of a visual asset according to some embodiments. The method 900 is implemented in some embodiments of the processing system 100 shown in FIG. 1, the cloud-based system 200 shown in FIG. 2, the GAN 500 shown in FIG. 5, and the portion 800 of the GAN shown in FIG. 8.
[0048] At block 905, a hint is provided to the generator. In some embodiments, the hint is a sketched digital representation of a portion of the visual asset (such as an outline). The hint may also include a label or metadata used in generating the image. For example, the label may indicate a type of visual asset, such as a "dragon" or a "tree". In another example, if the visual asset is segmented, the label may indicate one or more of the segments.
[0049] At block 910, random noise is provided to the generator. The random noise may be used to add a degree of randomness to the variation of images generated by the generator. In some embodiments, both the hints and the random noise are provided to the generator. However, in other embodiments, either the hints or the random noise or the other is provided to the generator.
[0050] In block 915, the generator generates images representing variations of the visual asset based on the hints, random noise, or a combination thereof. For example, if the label indicates a type of visual asset, the generator generates images of the variations of the visual asset using images with the corresponding labels. In another example, if the label indicates a segment of the visual asset, the generator generates images of the variations of the visual asset based on images of the segments with the corresponding labels. Thus, multiple variations of the visual asset can be created by combining different labeled images or segments. For example, a chimera can be created by combining the head of one animal with the body of another animal and the wings of a third animal.
[0051] In some embodiments, certain aspects of the above techniques may be realized by one or more processors of a processing system executing software. The software includes one or more sets of executable instructions stored or tangibly embodied in a non-transitory computer-readable storage medium. The software may include instructions and specific data that, when executed by one or more processors, operate the one or more processors to perform one or more aspects of the above techniques. The non-transitory computer-readable storage medium may include, for example, magnetic or optical disk storage devices, solid-state storage devices such as flash memory, caches, Random Access Memory (RAM) or other non-volatile memory devices, etc. The executable instructions stored in the non-transitory computer-readable storage medium may be source code, assembly language code, object code, or other instruction formats that are interpreted or executable by one or more processors.
[0052] A computer readable storage medium may include any storage medium or combination of storage media that is accessible by a computer system during use to provide instructions and / or data to the computer system. Such storage media may include, but are not limited to, optical media (e.g., compact discs (CDs), digital versatile discs (DVDs), Blu-ray discs), magnetic media (e.g., floppy disks, magnetic tapes or magnetic hard drives), volatile memory (e.g., random access memory (RAM) or cache), non-volatile memory (e.g., read-only memory (ROM) or flash memory), or Microelectromechanical System (MEMS) based storage media. The computer readable storage medium may be incorporated into the computing system (e.g., system RAM or ROM), permanently attached to the computing system (e.g., a magnetic hard drive), removably attached to the computing system (e.g., an optical disk or Universal Serial Bus (USB)-based flash memory), or coupled to the computer system via a wired or wireless network (e.g., Network Accessible Storage (NAS)).
[0053] It should be noted that not all of the operations or elements described above in the summary are required, some of the particular operations or devices may be unnecessary, one or more additional operations may be performed in addition to those described above, and one or more additional elements may be included. Furthermore, the order in which the operations are listed is not necessarily the order in which they are performed. Also, concepts have been described with reference to specific embodiments. However, those skilled in the art will appreciate that various modifications and changes can be made without departing from the scope of the present disclosure, as set forth in the following claims. Thus, the specification and drawings should be interpreted in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of the present disclosure.
[0054] Benefits, other advantages, and solutions to problems have been described above with respect to specific embodiments. However, these benefits, advantages, solutions to problems, and any features that may give rise to or make more prominent any benefit, advantage, or solution should not be construed as key, necessary, or essential features of any or all claims. Moreover, the disclosed subject matter may be modified and practiced in different but equivalent manners apparent to those skilled in the art having the benefit of the teachings herein. No limitations are intended to the details of construction or design shown herein, other than as set forth in the following claims. It is therefore apparent that the specific embodiments disclosed above may be altered or modified, and all such variations are considered to be within the scope of the disclosed subject matter. Accordingly, the protection sought herein is as set forth in the following claims.
Claims
1. 1. A computer-implemented method comprising: capturing a first image of a three-dimensional (3D) digital representation of a visual asset; generating second images representing variations of the visual asset using a generator in a generative adversarial network (GAN) and attempting to distinguish between the first and second images in a classifier in the GAN; updating at least one of a first model in the classifier and a second model in the generator based on whether the classifier successfully distinguishes between the first image and the second image; generating a third image using the generator based on the updated second model; The method, wherein the 3D digital representation includes primitives comprised of a set of polygons or patches, and textures that are applied to the primitives to incorporate visual detail having a higher resolution than the resolution of the primitives.
2. 2. The method of claim 1 , wherein capturing the first images from the 3D digital representation of the visual asset comprises capturing the first images using a virtual camera that captures the first images from different viewpoints and under different lighting conditions.
3. 1. A computer-implemented method comprising: capturing a first image of a three-dimensional (3D) digital representation of a visual asset; generating second images representing variations of the visual asset using a generator in a generative adversarial network (GAN) and attempting to distinguish between the first and second images in a classifier in the GAN; updating at least one of a first model in the classifier and a second model in the generator based on whether the classifier successfully distinguishes between the first image and the second image; generating a third image using the generator based on the updated second model; Capturing the first images from the 3D digital representation of the visual asset includes capturing the first images using a virtual camera that captures the first images from different viewpoints and under different lighting conditions; The method, wherein capturing the first image includes labeling the first image based on at least one of a type of the visual asset, a position of the virtual camera, a pose of the virtual camera, a texture applied to the visual asset, and a color of the visual asset.
4. 4. The method of claim 3, wherein capturing the first image includes segmenting the first image into portions associated with different portions of the visual asset and labeling the portions of the first image to indicate the different portions of the visual asset.
5. 5. The method of claim 4, wherein generating a third image using the generator based on the updated second model includes generating at least one third image in the generator in the GAN to represent a variation of the visual asset based on the updated second model by combining at least one labeled portion of the visual asset with at least one labeled portion of another visual asset.
6. 6. The method of claim 1, wherein updating at least one of the first model and the second model comprises applying a loss function indicative of at least one of a first likelihood that the second image is not distinguishable from the first image by the classifier and a second likelihood that the classifier successfully distinguishes between the first image and the second image.
7. The method of claim 6 , wherein the first model comprises a first distribution of a parameter in the first image and the second model comprises a second distribution of a parameter inferred by the generator.
8. 8. The method of claim 7, wherein applying the loss function comprises applying a perceptual loss function that extracts features from the first image and the second image and encodes differences between the first image and the second image as a distance between the extracted features.
9. 9. The method of claim 1, further comprising: generating, in the generator in the GAN, at least one third image to represent a variation of the visual asset based on the first model.
10. 10. The method of claim 9, wherein generating the at least one third image comprises generating the at least one third image based on at least one of a label associated with the visual asset or a digital representation of an outline of a portion of the visual asset.
11. 11. The method of claim 9 or 10, wherein the step of generating the at least one third image includes the step of generating the at least one third image by combining at least one portion of the visual asset with at least one portion of another visual asset.
12. A computer program embodying a set of executable instructions, said set of executable instructions operating at least one processor to perform a method according to any one of claims 1 to 11.
13. A memory configured to store the computer program of claim 12; and at least one processor configured to execute the computer program.
Citation Information
Patent Citations
Image processing device and program for generating two-dimensional image
JP2019185295A
Generating Synthetic Image Data
US20190057519A1
