Artificial intelligence (AI) based image generation using attribute-based slider control.

Abstract

This provides a method for generating AI-based images using attribute-based slider control. [Solution] A prompt indicating a description of the first image is received. The first image is generated based on the prompt by a text-image model. A set of attributes related to the description is determined based on a first language model. The set of attributes corresponds to the semantics related to the first image. A set of questions related to the prompt is generated based on a second language model. Slider boundary values ​​and initial slider values ​​are generated based on the set of questions and the first image. A set of sliders is generated based on the slider boundary values ​​and initial slider values. Each slider corresponds to an attribute. User input related to the set of sliders is received, and the second image is generated and rendered.

Description

Technical Field

[0001] The embodiments discussed in this disclosure relate to generative artificial intelligence (AI)-based image generation using attribute-based slider control.

Background Art

[0002] Generative artificial intelligence (AI) involves the use of various techniques to create new content such as images, music, or text that is not directly copied from existing data, but rather generated based on learned patterns and structures. One of the most prominent techniques in this field can be the use of adversarial generative networks (GANs). A GAN can include two neural networks, such as a generator and a discriminator. The two neural networks can operate cooperatively to create realistic images. One of the main challenges of GANs can be to identify specific directions in the latent space that correspond to meaningful edits within the generated images. This can make it difficult for a user to precisely adjust attributes such as age, gender, or hairstyle. When using attribute-based slider control, users may often experience inconsistent variations in the generated images. Just by slightly adjusting the slider, there can be unbalanced changes that compromise controllability and predictability in the image generation process. GANs can typically be trained on a specific dataset, which can limit the ability to generate diverse images. Due to these limitations, typical GANs may not be suitable for applications that require a diverse and general-purpose image generation function.

[0003] The subject matter claimed in this disclosure is not limited to embodiments that solve any disadvantages or that operate only in the environments as described above. Rather, this background art is provided only to illustrate an example of a technical field in which some of the embodiments described in this disclosure may be implemented.

Summary of the Invention

[0004] According to one aspect of the embodiment, the method may include a set of actions that include receiving a prompt indicating a description of a first image to be generated. The set of actions may further include generating the first image based on the prompt by a text-image model. The set of actions may further include determining a set of attributes (e.g., a person's age, a person's hair color, etc.) associated with the description of the first image (e.g., a young girl with curly hair) based on a first language model. The set of attributes may correspond to semantics associated with the first image. The set of actions may further include generating a set of questions associated with the prompt based on a second language model. The set of actions may include generating slider boundary values ​​and initial slider values ​​based on the set of questions associated with the prompt and the first image to generate a set of sliders associated with the set of attributes. The set of sliders may be generated based on the slider boundary values ​​and initial slider values. Each slider in the set of sliders may be associated with a corresponding attribute in the set of attributes. A set of actions may include receiving user input associated with a set of sliders, and generating a second image based on the user input associated with the set of sliders and the first image. Finally, the second image may be rendered on a display device.

[0005] The objectives and advantages of the embodiments will be realized and achieved, at least, by the elements, features, and combinations specifically indicated in the claims.

[0006] The above summary and the detailed description below are provided as examples and are explanatory; they do not limit the disclosures claimed.

[0007] Exemplary embodiments will be described and explained with further identification and detail using the attached drawings. [Brief explanation of the drawing]

[0008] [Figure 1] This diagram illustrates an exemplary environment for generative artificial intelligence (AI)-based image generation using attribute-based slider control. [Figure 2]This is a block diagram representing an exemplary electronic device for generative AI-based image generation using attribute-based slider control, as shown in Figure 1. [Figure 3] This diagram illustrates an exemplary execution pipeline for generative AI-based image generation using attribute-based slider control. [Figure 4] This diagram illustrates an exemplary electronic user interface (UI) showing a set of sliders for exploring the attribute space of a prompt. [Figure 5] This diagram illustrates an exemplary execution pipeline for determining slider boundary values ​​and initial slider values. [Figure 6A] This diagram collectively represents the scenarios of generated images based on variations in slider values. [Figure 6B] This diagram collectively represents the scenarios of generated images based on variations in slider values. [Figure 7] This diagram shows a flowchart illustrating a method for determining slider boundary values ​​based on Visual Question Response (VQA) scores. [Figure 8] This diagram shows a flowchart illustrating the method for determining the normalized value of the slider corresponding to the third image. [Figure 9] This diagram shows a flowchart of a generative AI-based image generation method using attribute-based slider control. [Modes for carrying out the invention]

[0009] All figures are in accordance with at least one embodiment described herein.

[0010] Some embodiments described in this disclosure may relate to methods and systems for generative artificial intelligence (AI)-based image generation using attribute-based slider control. In this disclosure, prompts may be received. Prompts may provide a description of a first image to be generated. Prompts may include, but are not limited to, text prompts, visual prompts, etc. A first image may be generated based on a prompt given by a user. A first image may be generated based on a text-image model. A set of attributes related to the description of the first image may be determined based on a first language model. The set of attributes may correspond to semantics related to the first image. The set of attributes may include, but are not limited to, a person's age, a person's hair color, a person's facial expression, a person's body type, a person's height, a person's face orientation, a person's gender, etc. A set of questions related to the prompt may be generated based on a second language model. Language models may include, but are not limited to, generative pre-trained transformers (GPT), bidirectional encoder representations from transformers (BERT), and text-to-text transfer transformers (T5). Slider boundary values ​​and initial slider values ​​may be generated based on a set of prompt-related questions and a first image. A set of sliders associated with a set of attributes may be generated based on slider boundary values ​​and initial slider values, with each slider in the set associated with a corresponding attribute in the set of attributes. User input associated with the set of sliders is received, and a second image may be generated based on the user input associated with the set of sliders and the first image. The second image may be rendered.

[0011] The field of generative AI-based image generation and parametric image editing can be improved by configuring an electronic device to generate an image (e.g., a second image) based on user input. The electronic device may receive a prompt indicating a description of the first image to be generated. The electronic device may generate the first image based on the prompt. The electronic device may determine a set of attributes related to the description of the first image. The set of attributes corresponds to the semantics associated with the first image. The electronic device may generate a set of questions related to the prompt. Furthermore, the electronic device may generate slider boundary values ​​and initial slider values ​​based on the set of questions related to the prompt and the first image. The electronic device may generate a set of sliders related to the set of attributes based on the slider boundary values ​​and initial slider values. Each slider in the set of sliders is associated with a corresponding attribute in the set of attributes. User input may be received to generate the second image. The user input is related to the set of sliders and the first image, and the second image may be rendered.

[0012] Generally, generative adversarial networks (GANs) may not be able to identify editing instructions. When a user attempts to modify a specific feature in a generated image, there may be no clear guidance on what to change, making it difficult to achieve the desired modification. This problem can be exacerbated by inconsistent variations in image generation when adjusting sliders. Users may experience that even slight adjustments can lead to unpredictable and non-uniform changes in the output. Therefore, adjustments may require a lot of trial and error and can be time-consuming. While GANs can create images that are very similar to the training data, generating entirely new and diverse images beyond the training set remains a complex task. This limitation may reduce the versatility of GANs in applications that require a wide range of image outputs. Users may often face difficulties when the initial alignment between the image and the presented prompt is unclear.

[0013] The disclosed approach may offer several advantages: 1. Flexible Thresholds - You can define the slider boundaries with flexible thresholds that cover various versions of an image or all meaningful images. 2. Consistent and predictable slider variation - Consistent slider variation allows for precise adjustments, ensures reliable data entry, and improves the overall effectiveness of slider variation. 3. Concept Slider - The concept slider allows for precise control of individual attributes with minimal interference. 4. Initial alignment of images and prompts - Concept sliders offer a more sophisticated approach, allowing for precise adjustment of individual attributes without affecting other attributes. 5. Use of Diffusion Models - By using diffusion models, it becomes possible to generate images with higher fidelity than GANs and to efficiently determine the direction of editing.

[0014] Traditional prompt-based image generation methods can generate images based on received prompts. This makes it difficult to determine how to adjust model parameters to make specific changes in the generated images. Adjusting control parameters (e.g., sliders) can lead to unpredictable and inconsistent changes in the generated images. This inconsistency can make it difficult for GAN models to precisely fine-tune or control their output. Furthermore, GANs may not be able to generate a wide variety of images outside of their training datasets. Also, current methods do not allow for systematic and controlled manipulation of specific attributes within generated images using clear and understandable parameters. This makes it difficult to explore and tune individual attributes in a precise and controlled manner.

[0015] The present disclosure can address these issues by providing generative AI-based image generation using attribute-based slider control. This approach can improve slider-driven image generation by enabling more efficient, consistent, and predictable slider variations.

[0016] Embodiments of the present disclosure are described with reference to the accompanying drawings.

[0017] FIG. 1 is a diagram illustrating an exemplary environment related to generative artificial intelligence (AI)-based image generation using attribute-based slider control in at least one embodiment described in the present disclosure. Referring to FIG. 1, environment 100 is shown. Environment 100 may include an electronic device 102, a text-image model 104, a communication network 106, a server 108, a database 110, and a display device 114. The electronic device 102, the server 108, and the database 110 can communicate with each other via the communication network 106. FIG. 1 further shows a prompt 102A, a generated image 112, a first image 116, and a second image 118.

[0018] The electronic device 102 may include suitable logic, circuitry, interfaces, and / or code configured to receive a prompt 102A indicating a description of the first image 116 and generate the first image 116 based on the prompt 102A using the text-image model 104. The prompt 102A may include a description of the image to be generated. For example, exemplary prompts 102A may include, but are not limited to, "portrait of a young woman," "a fantasy character with silver long hair standing in a magical forest with glowing eyes," "a child playing in the park with a bright smile," etc.

[0019] The electronic device 102 may determine a set of attributes related to the description of the first image 116 based on a first language model. The first language model may be a Large Language Model (LLM). The set of attributes may correspond to the semantics related to the first image 116. The electronic device 102 may generate a set of questions related to the prompt 102A based on a second language model. The second language model may also be a Large Language Model (LLM). Further, the electronic device 102 may generate slider boundary values and initial slider values based on the set of questions related to the prompt and the first image 116. A set of sliders may be generated in association with the set of attributes based on the slider boundary values and the initial slider values. Each slider in the set of sliders may be associated with a corresponding attribute in the set of attributes. Also, the electronic device 102 may receive user input related to the set of sliders and generate a second image 118 based on the user input related to the set of sliders and the first image 116. Finally, the second image 118 may be rendered on the display device 114. Based on the set of questions related to the prompt and the first image 116, slider boundary values and initial slider values can be generated. A set of sliders can be generated in association with the set of attributes based on the slider boundary values and the initial slider values. Each slider in the set of sliders can be associated with a corresponding attribute in the set of attributes. Also, the electronic device 102 can receive user input related to the set of sliders and generate a second image 118 based on the user input related to the set of sliders and the first image 116. Finally, the second image 118 can be rendered on the display device 114.

[0020] In an embodiment, the set of questions may include a first set of questions corresponding to questions independent of the slider and a second set of questions corresponding to questions dependent on the slider. The generation of the slider boundary values may be based on the first set of questions, and the generation of the initial slider values may be based on the second set of questions.

[0021] In an embodiment, the electronic device 102 may include the generation of slider boundary values corresponding to a first Visual Question Answering (VQA) model and the generation of initial slider values corresponding to a second VQA model. In some embodiments, the first VQA model may be the same as the second VQA model. In some other embodiments, the first VQA model may be different from the second VQA model.

[0022] In this embodiment, the electronic device 102 may generate a set of third images using a Low Rank Adaptation (LoRA) model. Based on the set of third images and the set of questions, the electronic device 102 may determine a VQA score using a first VQA model. The electronic device 102 may compare the VQA score to a first predetermined value and update the slider boundary value to the next value if the VQA score is less than the first predetermined value. The electronic device 102 may determine an upper limit for the slider boundary value if the VQA score is greater than the first predetermined value. The VQA score may also be compared to a second predetermined value, and the slider boundary value is updated to the previous value if the VQA score is less than the second predetermined value. Based on the VQA score being greater than the second predetermined value, a lower limit (or initial slider value) for the slider boundary value may be determined.

[0023] In one embodiment, the electronic device 102 can determine a Learned Perceptual Image Patch Similarity (LPIPS) score for each third image in a set of third images. Each third image is obtained by varying a slider value. Based on the estimation function of the LPIPS curve, the electronic device 102 can determine a mapping corresponding to a linear function associated with the LPIPS curve. Based on the mapping, a normalized value can be determined for the slider value corresponding to each third image.

[0024] The text-image model 104 may be a generative AI model capable of generating images based on natural language descriptions of images. The text-image model 104 may be trained on a set of text embeddings and an image dataset. The generative AI model may include a discriminator model that can be trained using a set of text embeddings and an image dataset. Training may be performed so that the discriminator model can classify whether the output generated by the generator model is associated with a real image (from the image dataset) or a fake image. The generator model may be trained to generate output images for text embeddings, and the discriminator model may not be able to reliably predict whether the generated output image is a real image from the image dataset or a fake image. Therefore, based on training, the generative AI model may be configured to generate images that may not be able to distinguish whether they are real or fake images. Examples of generative AI models include, but are not limited to, generative adversarial network (GAN) models, variational autoencoder (VAE) models, autoregressive models, transformer-based models, generative pre-trained transformers (GPT) models, or large-scale language models (LLMs).

[0025] The text-image model 104 may be applied to an received prompt 102A that provides a description of the first image 116. The description of the first image 116 may be processed to encode the description into a numerical format. The encoded description may then be mapped to a latent space. The latent space may be a high-dimensional space in which different features of the text are represented. Information from the latent space may be used to generate an image that matches the description hit in the text. Information from the latent space may be generated using a GAN model. The GAN model may be trained to produce an image that matches the description hit in the text. The display device 114 may be controlled to display the second image 118.

[0026] Server 108 may include logic, interfaces, and / or code that can be configured to store prompt 102A, information about the slider set, text-image model 104, and / or generated image 112 in database 110. Server 108 may be configured to read data (e.g., prompt 102A, information about the slider set, text-image model 104, and / or generated image 112) from database 110 and send the read data to electronic device 102.

[0027] Server 108 may be implemented as a cloud server and may perform operations through web applications, cloud applications, hypertext transfer protocol (HTTP) requests, repository operations, file transfers, etc. Other embodiments of Server 108 include, but are not limited to, a database server, file server, web server, media server, application server, mainframe server, cloud computing server, and / or any device with graph processing capabilities (e.g., a device with a set of graphics processor units (GPUs)).

[0028] In at least one embodiment, the server 108 may be implemented as a plurality of distributed cloud-based resources using some techniques well known to those skilled in the art. In certain embodiments, the functionality of the server 108 may be incorporated in whole or at least in part into the electronic device 102 without departing from the scope of this disclosure. In embodiments, the server 108 may be configured to train a text-image model 104, and the electronic device 102 may be configured to perform inference on a downstream prediction task (e.g., a task of producing a generated image 112 from a prompt 102A) based on the trained text-image model 104.

[0029] The database 110 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to store the generated images 112. The database 110 may further store the text-image models 104. The database 110 can be retrieved from data in relational or non-relational databases, or from a set of comma-separated value (CSV) files in conventional storage or big data storage. The database 110 may be stored or cached on a device such as a server 108 or an electronic device 102. The device storing the database 110 may be configured to receive queries for the generated images 112 or the text-image models 104. In response, the device storing the database 110 may be configured to read the generated images 112 or the text-image models 104 and send them to the electronic device 102.

[0030] Depending on the embodiment, the database 110 may be hosted on multiple servers stored in the same or different locations. The operation of the database 110 may be performed using hardware including a processor, a microprocessor (for example, one for performing or controlling one or more operations), a field-programmable gate array (FPGA), or an application-specific integrated circuit. In some other cases, the database 110 may be implemented using software.

[0031] Those skilled in the art will understand that the scope of this disclosure does not have to be limited to the implementation of database 110 and server 108 (or electronic device 102) as two separate entities. In certain embodiments, the functionality of database 110 may be incorporated into server 108 (or electronic device 102) in whole or at least in part without departing from the scope of this disclosure.

[0032] The communication network 106 may include various communication media through which the electronic device 102 can communicate with the server 108. Examples of the communication network 106 include, but are not limited to, the Internet, cloud networks, Wireless Fidelity (Wi-Fi) networks, personal area networks (PANs), local area networks (LANs), cellular networks (e.g., Long-Term Evolution (or 4G) cellular networks or 5G cellular networks), satellite networks (e.g., networks of low-Earth orbit satellites), and / or metropolitan area networks (MANs). Various devices in environment 100 may connect to the communication network 106 using various wired and wireless communication protocols, including TCP / IP, UDP, HTTP, FTP, ZigBee, EDGE, IEEE 802.11, Li-Fi, IEEE 802.16, multi-hop communication, wireless access points (APs), device-to-device communication, cellular communication protocols, and Bluetooth®.

[0033] The display device 114 may include logic, circuitry, and interfaces configured to display generated images (e.g., a first image 116, a second image 118). The first image 116 may be generated based on a received prompt 102A. The second image 118 may be generated based on user input associated with a set of sliders 214B. The display device 114 may be a touchscreen that allows a user to provide user input through the display device 114. The touchscreen may be at least one of a resistive touchscreen, a capacitive touchscreen, or a thermal touchscreen. The display device 114 may be implemented by several known technologies, including but not limited to liquid crystal displays (LCDs), light-emitting diode (LEDs), plasma displays, or organic LED (OLED) display technologies, or other display devices. Depending on the embodiment, the display device 114 may refer to a display screen for a head-mounted device (HMD), smart glasses device, see-through display, projection-based display, electrochromic display, or transmissive display.

[0034] During operation, the electronic device 102 may receive a prompt 102A indicating a description of the first image 116 to be generated. The prompt 102A may include, for example, a text prompt, a visual prompt, or an audio prompt. The description of the first image 116 may include, for example, "an image of a young woman riding a horse," or "an image of a man playing a game." The electronic device 102 may produce an output that includes a set of images or a set of videos merged as an image or video. The first image 116 may be generated by the text-image model 104 based on the prompt 102A. Receiving prompts is further illustrated, for example, in Figure 3.

[0035] In some embodiments, the electronic device 102 may determine a set of attributes associated with the description of the first image 116. The set of attributes may include, but is not limited to, “the person’s age,” “the person’s hair color,” “the person’s facial expression,” “the person’s body type,” “the person’s height,” “the person’s facial orientation,” or “the person’s gender.” The set of attributes may correspond to semantics associated with the first image 116. For example, prompt 102A may be “a realistic image of a young, muscular, well-built man with curly hair.” Prompt 102A may be processed by a first language model to generate a set of attributes. A set of attributes associated with the above-described exemplary prompt may be, for example, gender (i.e., male), age (i.e., young), hairstyle (i.e., curly hair), build (i.e., muscular), etc. Details regarding the determination of a set of attributes associated with the description of the first image 116 are given, for example, in Figure 3.

[0036] The electronic device 102 may be configured to generate a set of questions related to prompt 102A based on a second language model. Details regarding the generation of the set of questions related to prompt are given, for example, in Figure 4. The electronic device 102 may generate slider boundary values ​​and initial slider values ​​based on the set of questions related to prompt 102A and the first image 116. Details regarding the generation of slider boundary values ​​and initial slider values ​​are given, for example, in Figure 3.

[0037] The electronic device 102 can generate a set of sliders associated with a set of attributes based on slider boundary values ​​and initial slider values. Each slider in the set of sliders can be associated with a corresponding attribute in the set of attributes. The initial slider values ​​can be supplied based on a set of questions and the first image 116, instead of the user selecting a value for alignment with prompt 102A. Details regarding the generation of the slider set are given, for example, in Figure 3.

[0038] The electronic device 102 may be configured to receive user input associated with a set of sliders. The electronic device 102 may generate a second image 118 based on the user input associated with the set of sliders and the first image 116. The electronic device 102 may then render the generated second image 118 to the display device 114. Details regarding the receiving of user input, generation of the second image, and rendering of the second image are given, for example, in Figure 3.

[0039] Figure 2 is a block diagram representing the exemplary electronic device of Figure 1 for generative AI-based image generation using attribute-based slider control, according to at least one embodiment described herein. Figure 2 is described in relation to the elements from Figure 1. Referring to Figure 2, a block diagram 200 of the electronic device 102 is shown. The electronic device 102 may include a processor 202, memory 204, I / O device 206, network interface 208, first language model 210, second language model 212, and text-image model 104. The I / O device 206 may include a display device 114. The memory 204 may include the generated images 112 (e.g., first image 116 and second image 118), a set of attributes 214A, and a set of sliders 214B.

[0040] The processor 202 may include appropriate logic, circuitry, and / or interfaces that can be configured to execute program instructions related to various operations performed by the electronic device 102. These operations may include, but are not limited to, receiving prompts, generating a first image, determining attributes, determining questions, generating slider boundary values ​​and initial slider values, generating sliders, receiving user input, generating a second image, and rendering control. The processor 202 may include any appropriate specialized or general-purpose computer, computing entity, or processing device, including various computer hardware or software modules, and may be configured to execute instructions stored in any appropriate computer-readable storage medium. For example, the processor 202 may include a microprocessor, a μ-controller, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or any other digital or analog circuitry configured to interpret and / or execute program instructions and / or process data.

[0041] Although represented as a single processor in Figure 2, processor 202 may include any number of processors configured to individually or collectively perform or direct any number of operations of electronic device 102, as described in this disclosure. Furthermore, one or more processors may reside in one or more different electronic devices 102, such as different servers.

[0042] In some embodiments, the processor 202 may be configured to interpret and / or execute program instructions stored in memory 204, and / or process data stored in memory. In some embodiments, the processor 202 may fetch program instructions from memory 204 and load them into memory 204. After the program instructions are loaded into memory 204, the processor 202 may execute them. Some examples of the processor 202 may be a graphical processing unit (GPU), a central processing unit (CPU), a reduced instruction set computer (RISC) processor, an application-specific integrated circuit (ASIC) processor, a composite instruction set computer (CISC) processor, a coprocessor, and / or a combination thereof.

[0043] Memory 204 may include appropriate logic, circuitry, and / or interfaces that can be configured to store program instructions executable by the processor 202. In certain embodiments, memory 204 may be configured to store information such as, but not limited to, a generated image 112, a set of attributes 214A, and a set of sliders 214B. Memory 204 may further store a text-image model 104, a first language model 210, and a second language model 212.

[0044] Memory 204 may include computer-readable storage media that carry or store computer-executable instructions or data structures. Such computer-readable storage media may include any available media accessible by a general-purpose or specialized computer, such as a processor 202. Such computer-readable storage media may include, but are not limited to, tangible or non-temporary computer-readable storage media, including, CPU caches, hard disk drives (HDDs), solid-state drives (SSDs), random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disk read-only memory (CD-ROM), secure digital (SD) cards or other optical disk storage, magnetic disk storage or other magnetic storage devices, or flash memory devices (e.g., solid-state memory devices). Computer-readable storage may also include any other storage media accessible by a general-purpose or specialized computer that can be used to carry or store specific program code in the form of computer-executable instructions or data structures. Combinations of the above may also be included within the scope of computer-readable storage media. Computer executable instructions may include, for example, instructions or data configured to cause the processor 202 to perform a specific operation or group of operations related to the electronic device 102.

[0045] The first language model 210 and the second language model 212 may be large-scale language models (LLMs). An LLM may be an advanced AI system trained on large amounts of text data so that it can perform a wide range of natural language processing tasks, such as translation, summarization, and text generation. An LLM may use, for example, a transformer architecture, which enables it to process and generate text efficiently. During training, an LLM may learn statistical relationships between words and phrases by analyzing language datasets. Through this training, an LLM can learn how to determine the context, syntax, and meaning associated with any natural language text so that it can generate consistent and contextually relevant responses. Large-scale language models may include, but are not limited to, generative pre-trained transformers (GPT), bidirectional encoded representations with transformers (BERT), and text-to-text translators (T5).

[0046] The I / O device 206 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to receive user input. The I / O device 206 may be further configured to supply outputs in response to user input. The I / O device 206 may include various input and output devices that can be configured to communicate with the processor 202 and other components such as the network interface 208. For example, the input may include the first image 116 and the output may include the second image 118. Examples of input devices include, but are not limited to, a touchscreen, keyboard, mouse, joystick, and / or microphone. An example of an output device is, but is not limited to, the display device 114. The I / O device 206 may be configured within or outside the electronic device 102.

[0047] The network interface 208 may communicate wirelessly with networks such as the Internet, intranets, and / or cellular telephone networks, wireless local area networks (LANs), and / or metropolitan area networks (MANs). Wireless communication may use any of the following: Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (W-CDMA), Long-Term Evolution (LTE), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Bluetooth, Wireless Fidelity (Wi-Fi) (e.g., IEEE 802.11a, IEEE 802.11b, IEEE 802.11g, and / or IEEE 802.11n), Voice over Internet Protocol (VoIP), Light Fidelity (Li-Fi), or Wi-MAX.

[0048] Without departing from the scope of this disclosure, modifications, additions, or deletions may be made to the exemplary electronic device 102. For example, in some embodiments, the exemplary electronic device 102 may include any number of other components that are not explicitly illustrated or described for the sake of brevity.

[0049] Figure 3 is a diagram representing an exemplary execution pipeline for generative AI-based image generation using attribute-based slider control, according to at least one embodiment described in this disclosure. Figure 3 may be described in relation to elements from Figures 1 and 2. Referring to Figure 3, an exemplary execution pipeline 300 is shown. The exemplary execution pipeline 300 may include a series of operations that can be performed by the processor 202 of the electronic device 102 in Figure 1 for generative AI-based image generation using attribute-based slider control.

[0050] The execution pipeline 300 may include operations such as receiving a prompt 302, generating a first image 304, generating a set of attributes 306, generating a set of questions 312, generating slider-independent questions 314A, generating question-dependent questions 314B, aligning images and text 316A and 316B, determining slider boundary values ​​318B, determining initial slider values ​​318B, generating a set of sliders 310 (e.g., slider 1, slider 2, ... and slider N), and merging sliders 310A. The execution pipeline 300 may also include a set of attributes 308 (e.g., attribute 1 308A, attribute 2 308B, ... and attribute N 308N), a generated image 112, and an edited image or second image 320.

[0051] At 302, the operation of receiving a prompt may be performed. Processor 202 may be configured to receive a prompt 102A indicating a description of the first image 116 to be generated (e.g., generated image 112). Example prompts 102A may include, but are not limited to, "portrait of a young woman," "fantasy character with long silver hair standing in a magical forest with glowing eyes," or "child playing in a park with a bright smile." Prompt 102A may include a natural language description that may be input for generating the first image 116. Prompt 102A may be received as user input using text, gesture, haptic, or audio input devices associated with electronic device 102.

[0052] At 304, the operation of generating the first image may be performed. The processor 202 may be configured to generate the first image 116 (i.e., the generated image 112) based on the prompt 102A using the text-image model 104. The text-image model 104 may be a machine learning model that can be configured to generate images based on natural language descriptions. For example, if the received prompt 102A is "cat wearing a hat", the text-image model 104 may generate an image representing a cat wearing a hat.

[0053] In 306, the operation of generating a set of attributes may be performed. The processor 202 may be configured to determine a set of attributes 214A related to the description of the first image 116 based on a first language model (e.g., LLM). The set of attributes 214A may include age, hair color, facial expression, hand gestures, eye movements, etc. The set of attributes 214A may be adjusted based on user requests. Adjustments may include, for example, adding new attributes, updating attributes, deleting attributes, etc. The set of attributes 214A may be generated based on a first language model 210. The first language model 210 may be an LLM. For example, an LLM may analyze an input prompt 102A to understand the various attributes described. The set of attributes may include object, color, size, location, and other descriptive details. An LLM may extract a set of attributes 214A from a prompt 102A such as "a red apple on a wooden table". The set of attributes 214A may be "red", "apple", and "wooden table". The attribute set 214A can be mapped to visual elements, each accompanied by a visual representation of the attribute. For example, "red" may be mapped to a specific shade of red, and "apple" may be mapped to the shape and texture of an apple. In one example, an attribute set 308 (e.g., attribute 1 308A, attribute 2 308B, ..., and attribute N 308N) may be generated, as shown in Figure 3.

[0054] In step 312, the operation of generating a set of questions may be performed. The processor 202 may be configured to generate a set of questions related to prompt 102A based on a second language model (e.g., LLM). The set of attributes 214A may include age, hair color, facial expression, hand gestures, eye movements, etc. The set of attributes 214A may be adjusted based on user requests. Adjustments may include, for example, adding new attributes, updating attributes, deleting attributes, etc. The set of attributes 214A may be generated based on a first language model 210. The first language model 210 may be LLM. For example, LLM may analyze input prompt 102A to determine various attributes described. The set of attributes may include object, color, size, location, and other descriptive details. LLM may extract a set of attributes 214A from prompt 102A such as "a red apple on a wooden table". The set of attributes 214A may be "red", "apple", "wooden table". The attribute set 214A can be mapped to visual elements, each accompanied by a visual representation of the attribute. For example, "red" may be mapped to a specific shade of red, and "apple" may be mapped to the shape and texture of an apple. In one example, an attribute set 308 (e.g., attribute 1 308A, attribute 2 308B, ..., and attribute N 308N) may be generated, as shown in Figure 3. Questions may include, for example, "curly hair," "young girl," etc.

[0055] In 314A, the operation of generating slider-independent questions may be performed. Processor 202 may be configured to generate slider-independent questions. The set of questions may include a first set of questions corresponding to the slider-independent questions. Generation of slider boundary values ​​may be based on the first set of questions. The slider boundary values ​​may correspond to a first VQA model. For example, the first VQA model may be a computer vision model that can be configured to determine the context associated with an image and answer context-related text questions.

[0056] In 314B, the operation of generating slider-dependent questions may be performed. Processor 202 may be configured to generate slider-dependent questions. The set of questions may include a second set of questions corresponding to the slider-dependent questions. The generation of initial slider values ​​may be based on the second set of questions. The generation of initial slider values ​​may correspond to a second VQA model. In some embodiments, the first VQA model may be the same as the second VQA model. In some further embodiments, the first VQA model may be different from the second VQA model. For example, the second VQA model may be a computer vision model that can determine the context associated with an image and answer context-related text questions.

[0057] In 316A and 316B, image and text alignment operations may be performed. Processor 202 may be configured to align the image with the text from prompt 102A. A VQA model may be used to align the generated image with prompt 102A received as input. The VQA model may be configured to answer questions about the generated image 112 in both visual content and text questions. The VQA model may generate sets of questions. Sets of questions may include a first set of questions and a second set of questions. The first set of questions may correspond to slider-independent questions, and the second set of questions may correspond to slider-dependent questions. The generated image and the text in prompt 102A may be combined using a fusion model. The fusion model may determine the relationship between the visual content of the generated image and the sets of questions. In some embodiments, text prompts may be used to generate a composite image to help answer questions. This may involve translating the text prompt into a visual representation using a vision-language model. Next, the visual representation can be analyzed to generate an answer. Slider boundary values ​​can be generated based on a first set of questions. Initial slider values ​​can be generated based on a second set of questions. Sets of questions influenced by the slider set 214B (e.g., slider-dependent questions) can be sent to the VQA model along with the generated image 112 to determine the exact slider values ​​(e.g., initial slider values). For example, for the age slider, the relevant question might be "Is this person young?". The VQA model then considers the image and selects a value that maximizes the probability of a "yes" response to the question "Is this person young?". Questions not influenced by the slider set 214B (e.g., slider-independent questions) can be sent to the VQA model along with the generated image to determine boundary slider values. For example, for the age slider, an uninfluenced question might be "Does this person have curly hair?". Next, the VQA model selects boundary values ​​such that the probability of answering "yes" to the question "Does this person have curly hair?", taking the image into account, falls below a certain threshold.

[0058] In 318A and 318B, operations may be performed to determine the slider boundary value and the initial slider value, respectively. Processor 202 may be configured to generate the slider boundary value based on a first set of questions corresponding to slider-independent questions (e.g., questions determined in 314A). Processor 202 may be configured to generate the initial slider value based on a second set of questions corresponding to slider-dependent questions (e.g., questions determined in 314B). The slider set 214B may be generated based on the generated slider boundary value and initial slider value. The generation of the slider boundary value may be based on a first VQA model. The generation of the initial slider value may be based on a second VQA model. Further details regarding the generation of the slider boundary value and initial slider value are described, for example, in Figure 5.

[0059] At 310, the operation of generating the slider set 214B may be performed. The processor 202 may be configured to generate slider boundary values ​​and initial slider values ​​(as described, for example, in 318A and 318B). The processor 202 may generate the slider set 214B based on the generated slider boundary values ​​and initial slider values. The slider set 214B, along with the initial slider values ​​and slider boundary values, may be mapped to the second image 118.

[0060] At 310A, the slider merging operation may be performed. The processor 202 may be configured to merge the set of sliders 214B using various techniques such as diffusion-based generative AI models, GANs, autoencoders, image blending techniques, feature extraction and manipulation. The user may supply input related to the variation of slider values ​​in order to obtain a desired image from the original generated image 112. Based on the slider variation related to the user input, a second image (or edited image) 320 may be generated. Thus, the slider values ​​may be merged to create a final image or edited image (e.g., second image 118 or edited image 320).

[0061] Typically, generative adversarial networks (GANs) may not be able to identify editing instructions. When a user attempts to modify a specific feature in a generated image, there may be no clear guidance on what to change, making it difficult to achieve the desired modification. This problem can be exacerbated by inconsistent variations in image generation when adjusting sliders. Users may experience that even slight adjustments can lead to unpredictable and uneven changes in the output. Therefore, adjustments may require a lot of trial and error and can be time-consuming. While GANs can create images that are very similar to the training data, generating entirely new and diverse images beyond the training set remains a complex task. This limitation can reduce the versatility of GANs in applications that require a wide range of image outputs. Users may often face difficulties when the initial alignment between the image and the presented prompt is unclear.

[0062] The disclosed approach may offer several advantages: 1. Flexible Thresholds - You can define the slider boundaries with flexible thresholds that cover various versions of an image or all meaningful images. 2. Consistent and predictable slider variation - Consistent slider variation allows for precise adjustments, ensures reliable data entry, and improves the overall effectiveness of slider variation. 3. Concept Slider - The concept slider allows for precise control of individual attributes with minimal interference. 4. Initial alignment of images and prompts - Concept sliders offer a more sophisticated approach, allowing for precise adjustment of individual attributes without affecting other attributes. 5. Use of Diffusion Models - By using diffusion models, it becomes possible to generate images with higher fidelity than GANs and to efficiently determine the direction of editing.

[0063] Traditional prompt-based image generation methods can generate images based on received prompts. This makes it difficult to determine how to adjust model parameters to make specific changes in the generated images. Adjusting control parameters (e.g., sliders) can lead to unpredictable and inconsistent changes in the generated images. This inconsistency can make it difficult for GAN models to precisely fine-tune or control their output. Furthermore, GANs may not be able to generate a wide variety of images outside of their training datasets. Also, current methods do not allow for systematic and controlled manipulation of specific attributes within generated images using clear and understandable parameters. This makes it difficult to explore and tune individual attributes in a precise and controlled manner.

[0064] This disclosure may address these challenges by providing generative AI-based image generation using attribute-based slider control. This approach can improve slider-driven image generation by enabling more efficient, consistent, and predictable slider variations.

[0065] In one example, selecting which slider to choose from a database of hundreds of sliders can be difficult for a user. This disclosure enables the retrieval of relevant sliders based on a received prompt 102A. Furthermore, the electronic device 102 can also identify whether a given attribute associated with prompt 102A can be modified by multiple sliders. Thus, the electronic device 102 can enable a user to create custom sliders by recommending attributes that the user can instruct the creation of.

[0066] Figure 4 is a diagram representing an exemplary electronic user interface (UI) showing a set of sliders for exploring the attribute space of a prompt, according to embodiments of the present disclosure. Figure 4 is described in relation to elements from Figures 1, 2, and 3. An exemplary electronic UI 400 is shown with reference to Figure 4.

[0067] The electronic UI 400 may include various UI elements such as a text prompt UI element 402, an attribute selection UI element 410, a first generated image 408A, a second generated image 408B, a third generated image 408C, an attribute insertion UI element 412, a slider set 414, a generated image 112, an image editing option 416, and an image selection option from the gallery 422. The electronic UI 400 may further include a UI element 418 for displaying the generated image, an image upload UI element 418A, a save button 420, and a UI element 422 for selecting an image from the gallery.

[0068] The user may input prompt 102A via a text prompt UI element 402 (e.g., a text box). The entered prompt 102A may contain text 404 such as, for example, "young girl with curly hair". The user may submit prompt 102A and click or press a button such as "Generate" 406 to generate an image. When the user inputs prompt 102A, a first image 116 may be generated and the generated image 112 may be displayed. The electronic UI 400 may display multiple images based on user input. For example, the electronic UI 400 may render a first generated image 408A, a second generated image 408B, and a third generated image 408C. The generated image 112 may be saved and reused for further editing.

[0069] In this embodiment, the attribute set 214A may be generated based on a prompt 102A. The prompt 102A may include user input such as, for example, "young girl with curly hair". The user may select the attribute set 214A using the attribute selection UI element 410. The attribute set 214A may be, for example, "age", as shown in Figure 4. A dropdown function may be provided for selecting multiple attributes based on user requests. The electronic UI 400 may insert new attributes based on user input obtained using the attribute insertion UI element 412. One or more new attributes (e.g., "hair type") may be added by inserting attributes along with existing attributes (e.g., age). The attribute insertion UI element 412 can function as an attribute filter and is not limited to inserting attributes. In some scenarios, attributes may be removed from the attribute set 214A by the attribute insertion UI element 412.

[0070] In one embodiment, a set of sliders 214B may be generated based on a set of attributes 214A. In some aspects, each attribute in the set of attributes 214A may include a slider to edit the image based on the slider's variation. In one example, the slider "age" 414 is shown in Figure 4. Each slider in the set of sliders 214B may include a predefined range of values. The user can adjust the set of sliders 214B to obtain the desired image. The predefined range of values ​​may vary, for example, between (-4, 4). The electronic UI 400 may include UI elements such as buttons to accept user commands to edit the image after it has been generated. For example, an image editing option 416 may be used by the user to further edit the image. Finally, the image may be displayed in part of the electronic UI 400. The electronic UI 400 may include UI elements for selecting an image from a gallery and a save button 420 for saving the generated image 112. Furthermore, the electronic UI 400 may have an image upload UI element 418A for uploading an image to be edited from the gallery.

[0071] It should be noted that the electronic UI400 in Figure 4 is provided for illustrative purposes only and should not be interpreted as limiting the scope of this disclosure.

[0072] Figure 5 is a diagram representing an exemplary execution pipeline for determining slider boundary values ​​and initial slider values ​​according to embodiments of the present disclosure. Figure 5 is described in relation to elements from Figures 1, 2, 3, and 4. Referring to Figure 5, an exemplary execution pipeline 500 is shown. The execution pipeline 500 may include an operation 502 for inputting prompts, an operation 504 for applying a text image model, an operation 504A for selecting slider values, and an operation 512 for determining slider boundaries and slider alignment. The operations of the execution pipeline 500 may be performed by an electronic device 102 or any computing device such as a processor 202.

[0073] At 502, a prompt input operation may be performed. Processor 202 may be configured to receive prompt input, for example, prompt 102A. Prompt 102A may be, for example, "young girl with curly hair". The user may input prompt 102A to generate an image.

[0074] In step 504, the operation of applying a text-image model may be performed. Processor 202 may be configured to apply the text-image model 104 to the received prompt 102A. The text-image model 104 may be a low-rank adaptive (LoRA) model. LoRA models can be a technique used in machine learning to reduce the number of model parameters to efficiently fine-tune the model by utilizing low-rank decomposition. This involves decomposing the weight matrix of the text-image model 104 into a lower-dimensional matrix, which can make training easier. The text-image model 104 may generate a first image based on the prompt 102A.

[0075] Therefore, the text-image model 104 may be a machine learning model that can be configured to generate an image (e.g., a first image 116) based on a natural language description. A prompt 102A may be supplied to the text-image model 104 to obtain a generated image 112 that matches the description in prompt 102A. For example, if the user enters the prompt "cat wearing a hat", the model may generate an image representing a cat wearing a hat. The text-image model 104 may include, for example, a diffusion-based generative AI model, a GAN, an autoencoder, an image blending technique, and a model for feature extraction and manipulation.

[0076] In 506, the input prompt 102A may be received in a Davidson Scene Graph. The DSG may be an automated graph-based framework for question generation and answering (QG / A). It can enhance the reliability of fine-grained evaluation of text-image generation models. The DSG can generate unique, indivisible questions organized in a dependency graph. By organizing questions in a dependency graph, the DSG can ensure comprehensive semantic coverage. This helps to accurately evaluate the alignment between the generated image 112 and the input prompt 102A. The DSG can avoid inconsistent answers by structuring questions in a way that avoids contradictions. The input prompt 102A may be analyzed to generate a set of contextually relevant questions. These questions may be designed to examine various aspects of the image matched with the text. For example, given the input text "a red car parked under a tree," a DSG might generate questions such as "Is there a car in the image?", "What color is the car?", "Is the car parked under something?", and "What is the car parked under?". These questions can be answered by a VQA model, and the answers are compared to expected responses to assess the accuracy of the image. DSGs can support large-scale language models (LLMs), including but not limited to the Generative Pre-trained Transducers (GPT) series, Bidirectional Encoded Representations with Transducers (BERT), and Text-to-Text Transducers (T5).

[0077] At 504A, the operation of selecting a slider value may be performed. The processor 202 may be configured to select a slider value based on the application of the text-image model 104 to prompt 102A. For example, the slider value may initially be selected as (-8,-2)U(2,8). However, the selected slider value may not be limited to (-8,-2)U(2,8) and may be any range of natural numbers without departing from the scope of this disclosure.

[0078] At 508, the operation of generating a set of images (e.g., a set of third images) may be performed. Processor 202 may generate a set of images based on various slider values ​​associated with a predefined range. For example, a set of images may include image 1 corresponding to a slider value of "2", image 2 corresponding to a slider value of "3", ... and image N corresponding to a slider value of "-2". The set of images and the set of questions may be used to determine a visual question answer (VQA).

[0079] At 510, the VQA score determination operation may be performed. Processor 202 may determine the VQA score based on the VQA model, using the set of third images (determined at 508) and the set of questions (determined at 506). The determination of the VQA score may be represented by the following pseudocode: [Table 1]

[0080] The processor 202 may compare the VQA score to a first predetermined value. The first predetermined value may be, for example, (0,4). Based on the fact that the VQA score is less than the first predetermined value, the processor 202 may update the slider value (e.g., slider boundary value) to the next value. Furthermore, based on the fact that the VQA score is greater than the first predetermined value, the processor 202 may determine an upper limit for the slider boundary value.

[0081] For example, to determine the upper bound, referencing pseudocode for how the VQA score is determined, "U_bound" (upper bound variable) and "L_bound" (lower bound variable) may be initialized to the value "0". These variables may be used to store the upper and lower bounds, respectively. The outer loop may iterate through half of the sets of third images, for example, len(images) / 2 times. "I" may represent a loop variable that can represent the current index in the list of sets of third images. For each image, the inner loop iterates through the list of questions. "VQA_score" may be calculated using the function P(Yes / question,image[i+4]). This function may correspond to a VQA model that considers the question and the image (e.g., one of the third images in the set of third images) to predict the probability that the answer is "yes". Image[i+4] means that for each iteration of "I", the function may evaluate the image at index "i+4". If "VQA_score" is less than "0.6", it may indicate that the VQA model has low confidence that the answer is "yes". If the condition is met, "U_bound" can be set to the value of "i+4", and the inner loop is broken. This means that the upper bound can be determined based on the index where the VQA score first falls below "0.6".

[0082] The processor 202 may compare the VQA score to a second predetermined value. Furthermore, the processor 202 may update the slider boundary value to the previous value based on the fact that the VQA score is less than the second predetermined value. The processor 202 may determine a lower limit for the slider boundary value based on the fact that the VQA score is greater than the second predetermined value.

[0083] For example, to determine the lower bound, referencing pseudocode, "U_bound" and "L_bound" may be initialized to "0". These variables can be used to store the upper and lower bounds, respectively. The outer loop may iterate through half of the sets of third images, for example, len(images) / 2 times. This means that if there are 10 images, the loop will run 5 times. "I" may represent a loop variable that can represent the current index in the list of sets of third images. For each image, the inner loop iterates through the list of questions. "VQA_score" may be calculated using the function P(Yes / question,image[-i-4]). This function can represent a VQA model that predicts the probability that the answer is "yes" considering the question and image, where image[-i-4] represents each iteration of "I", and the function may evaluate the image at index "-i-4". Negative index "-i" starts counting from the end of the list. If "VQA_score" is less than "0.6", it may indicate that the VQA model has low confidence that the answer is "yes". If the condition is met, "L_bound" can be set to the value "-i-4", and the inner loop is broken. This means that the lower bound can be determined based on the index where the VQA score first falls below "0.6".

[0084] At 512, the operation of determining the slider boundary and slider alignment may be performed. Processor 202 may determine the slider boundary and perform slider alignment. The slider boundary may limit the range in which the slider can operate. The slider boundary may typically be defined by upper and lower limits that can constrain the range of values ​​the slider can take. Slider alignment can be said to place the slider within its defined boundary. Slider alignment may ensure that the slider moves smoothly and precisely within its range.

[0085] Figures 6A and 6B are diagrams that collectively represent scenarios of generated images based on variations in slider values, according to embodiments of the present disclosure. Figures 6A and 6B are described in relation to elements from Figures 1, 2, 3, 4, and 5.

[0086] Referring to Figure 6A, an exemplary first scenario 600A is shown that can represent an image of person 602A. An initial slider value may be given to allow the user to select the best match with prompt 102A. For example, for prompt 102A, a first image 116 may be generated. The image of person 602A may be a photograph of a person with a smile (represented by 604A) that can correspond to features such as a blurred street background, a realistic effect, and a toothless face of person 602A with "8K" resolution. The user may be provided with a smile slider to adjust the range of person 602A's smile based on the user's request. For example, the generated image of person 602A may correspond to a slider value of "-2". The slider value can vary, for example, between (-4 and 4), and the user may select a slider value based on the desired output.

[0087] Referring to Figure 6B, an exemplary second scenario 600B is shown that can represent an image of person 602B. The second scenario 600B may represent slider changes made to the user based on user requests and prompts 102A. For example, the image of person 602B may be a photograph of person 602B smiling with their teeth showing, with a blurred street background, realistic effects, and "8K" resolution. Image 600B may be obtained by changing the slider value to, for example, "0".

[0088] It should be noted that Scenario 1 600A in Figure 6A and Scenario 2 600B in Figure 6B are for illustrative purposes only and should not be interpreted as limiting the scope of this disclosure.

[0089] Figure 7 is a diagram illustrating a flowchart of a method for determining slider boundary values ​​based on visual question answer (VQA) scores according to embodiments of the present disclosure. Figure 7 is described in relation to elements from Figures 1, 2, 3, 4, 5, 6A, and 6B. Referring to Figure 7, an exemplary flowchart 700 of a method for determining slider boundary values ​​based on VQA scores is shown. Flowchart 700 may include operations 702-718 that can be performed by the processor 202 (in Figure 2) of the electronic device 102 (in Figure 1).

[0090] In 702, a set of third images may be generated by a low-rank adaptive (LoRA) model. Processor 202 may be configured to generate a set of third images based on the LoRA model. Prompt 102A may be supplied as input to the LoRA model. The LoRA model may be a technique used in machine learning to reduce the number of model parameters to efficiently fine-tune the model by utilizing low-rank decomposition. This involves decomposing the weight matrix of the text-image model 104 into a lower-dimensional matrix, which can make training easier. The text-image model 104 may generate a set of third images based on prompt 102A and the LoRA model. The LoRA model may generate a set of third images with different slider values ​​for determining the VQA score. The VQA score may be determined based on the first VQA model using the set of third images and the set of questions. The first VQA model may be the same as the second VQA model. In some embodiments, the first VQA model may be different from the second VQA model. Further details regarding image generation are described, for example, in Figures 3 and 5.

[0091] In 704, a first VQA score may be determined by a first VQA model based on a set of third images and a set of questions, and a second VQA score may be determined by a second VQA model. The processor 202 may be configured by the first VQA model to determine a first VQA score based on a set of third images and a set of questions (e.g., a first set of questions corresponding to slider-independent questions). The processor 202 may be configured by the second VQA model to determine a second VQA score based on a set of third images and a set of questions (e.g., a second set of questions corresponding to slider-dependent questions). Based on the first VQA score, slider boundary values ​​may be determined, and based on the second VQA score, initial slider values ​​may be determined. The determination of the first and second VQA scores is further described, for example, in Figures 3 and 5.

[0092] In 706A, the first VQA score can be compared to a first predetermined value. The processor 202 may be configured to compare the first VQA score of the generated set of third images to the first predetermined value. The first predetermined value may correspond to the mean of the VQA scores. The first VQA score may be determined based on a first VQA model that can be expressed using a function P(Yes / question,image[i+4]). The function may correspond to a first VQA model that takes a question (e.g., a set of first questions) and a set of third images to predict the probability that the answer is "yes".

[0093] At 708, it may be determined whether the first VQA score is less than a first predetermined value. Processor 202 may be configured to compare the first VQA score with the first predetermined value. If the first VQA score is less than the first predetermined value, control may be passed to 710. Otherwise, control may be passed to 712.

[0094] At 710, the slider boundary value may be updated to the following value. The processor 202 may be configured to update the slider boundary value to the following value based on the fact that the first VQA score is less than a first predetermined value. For example, the first VQA score may be less than a first predetermined value. In such a case, if the slider boundary value was "2", the slider boundary value may be moved to the following value, i.e., "3".

[0095] In step 712, the upper limit of the slider boundary value can be determined. The processor 202 may be configured to determine the upper limit of the slider boundary value based on whether the first VQA score is greater than a first predetermined value. For example, the first VQA score may be greater than a first predetermined value. In such a case, if the first VQA score was "5", then the upper limit of the slider boundary value would be "5".

[0096] In 706B, the second VQA score can be compared to a second predetermined value. Processor 202 may be configured to compare the second VQA score of the generated set of third images to the second predetermined value. The second predetermined value may correspond to the mean of the VQA scores. The second VQA score may be determined based on a second VQA model that can be expressed using the function P(Yes / question,image[-i-4]). The function may correspond to a second VQA model that considers a question (e.g., a set of second questions) and a set of third images to predict the probability that the answer is "yes".

[0097] At 714, it may be determined whether the second VQA score is less than a second predetermined value. Processor 202 may be configured to compare the second VQA score with the second predetermined value. If the second VQA score is less than the second predetermined value, control may be passed to 716. Otherwise, control may be passed to 718.

[0098] In 716, the slider boundary value can be updated to the previous value. The processor 202 may be configured to update the slider boundary value to the previous value based on the fact that the second VQA score is less than the second predetermined value. For example, the second VQA score may be less than the second predetermined value. In such a case, if the slider boundary value was "-2", the slider boundary value may be moved to the previous value, i.e., "-3".

[0099] At 718, the lower limit of the slider boundary value can be determined. The processor 202 may be configured to determine the lower limit of the slider boundary value based on the fact that the second VQA score is greater than a second predetermined value. For example, the second VQA score may be greater than a second predetermined value. In such a case, if the first VQA score was "5", then the lower limit of the slider boundary value will be "5". Control can then proceed to termination.

[0100] While flowchart 700 is represented as individual operations such as 702, 704, 706A, 706B, 708, 710, 712, 714, 716, and 718, the disclosure is not limited in this way. However, in certain embodiments, such individual operations may be further divided into additional operations, combined into fewer operations, or deleted, depending on the particular implementation, without deviating from the essence of the disclosed embodiments.

[0101] Figure 8 is a diagram showing a flowchart of a method for determining a normalized value of a slider value corresponding to a third image, according to embodiments of the present disclosure. Figure 8 is described in relation to elements from Figures 1, 2, 3, 4, 5, 6A, 6B, and 7. Referring to Figure 8, an exemplary flowchart 800 of a method for determining a normalized slider value corresponding to an image is shown. Flowchart 800 may include operations 802-808 that can be performed by the processor 202 (in Figure 2) of the electronic device 102 (in Figure 1).

[0102] In 802, a trained Perceptual Image Patch Similarity (LPIPS) score may be determined for a set of third images, where each third image may correspond to a slider value. Processor 202 may be configured to determine the LPIPS score for a set of third images, where each third image in the set may correspond to a slider value. The LPIPS score may be used to measure the perceptual similarity between two images. LPIPS corresponds to a deep learning-based metric that compares two images by passing them through a pre-trained neural network and calculating the distance between their feature representations. The LPIPS score may be determined by loading each third image in the set of third images and a reference image. A pre-trained LPIPS model may be used to obtain the similarity score.

[0103] In 804, the LPIPS curve may be estimated based on the LPIPS score for each third image in a set of third images. Processor 202 may be configured to estimate the LPIPS curve based on the LPIPS score for each third image in a set of third images. Estimating the LPIPS curve may involve collecting the LPIPS scores, fitting a function to the LPIPS scores, and then using the fitted function to determine or predict the trend of perceptual similarity. The LPIPS scores may be organized into a structured format, and a mathematical model may be selected to represent the relationship between the LPIPS scores and the corresponding variables. A statistical or machine learning (ML) model may be selected to fit the selected model to the data containing the collected LPIPS scores. This involves finding the parameters of the statistical or ML model that best describe the data. The performance of the fitted model may be evaluated using appropriate metrics (e.g., R-squared error and mean squared error) to ensure that the LPIPS curve accurately represents the relationship. Once it is evaluated that the LPIPS curve accurately represents the relationship, the LPIPS curve may be used to predict the LPIPS score for new data points.

[0104] Processor 806 may determine a mapping corresponding to a linear function associated with the LPIPS curve. Processor 202 may be configured to determine a mapping corresponding to a linear function associated with the LPIPS curve. A linear model can be fitted to the LPIPS score using linear regression. The performance of the linear model can be evaluated using appropriate metrics. The fitted linear model can be used to predict the LPIPS score for new data points or to understand the underlying trends.

[0105] In 808, the normalized values ​​of the slider values ​​corresponding to each third image in the set of third images may be determined based on the mapping. The processor 202 may be configured to determine the normalized values ​​for the normalized slider values ​​corresponding to each third image in the set of third images based on the mapping. The slider values ​​may be normalized based on a fitted linear model so as to ensure that each slider value falls within a specific range, for example, between 0 and 1. Control can then be moved to termination.

[0106] While flowchart 800 is represented as individual operations such as 802, 804, 806, and 808, the disclosure is not limited in this way. However, in certain embodiments, such individual operations may be further divided into additional operations, combined into fewer operations, or deleted, depending on the particular implementation, without deviating from the essence of the disclosed embodiments.

[0107] Figure 9 is a diagram representing a flowchart of a generative AI-based image generation method using attribute-based slider control according to embodiments of the present disclosure. Figure 9 is described in relation to elements from Figures 1, 2, 3, 4, 5, 6A, 6B, 7, and 8. Referring to Figure 9, an exemplary flowchart 900 of a generative AI-based image generation method using attribute-based slider control is shown. Flowchart 900 may include operations 902-920 that can be performed by the processor 202 (in Figure 2) of the electronic device 102 (in Figure 1). Flowchart 900 may begin at 902 and continue at 904.

[0108] At 904, a prompt indicating a description of the first image to be generated may be received. The processor 202 may be configured to receive a prompt 102A indicating a description of the first image 116 to be generated. Prompt 102A may be received from a user or any electronic device. Prompt 102A may be a text prompt or an image obtained from a user gallery or saved in any database. The image may be, for example, a real image, a cartoon image, a painting, etc. The receipt of prompts is described further, for example, in Figure 3.

[0109] In 906, the first image may be generated by a text-image model based on a prompt. Processor 202 may be configured to generate the first image 116 based on prompt 102A by text-image model 104. The first image 116 may be, for example, a real image, a cartoon image, a painting, etc., generated based on prompt 102A. Text-image model 104 may use machine learning techniques such as a GAN model or a diffusion model to create an image that matches a given text (e.g., prompt 102A). Processor 202 may convert the input text into a numerical representation (or encoded text) using techniques such as embedding or transformer. This can capture the semantic meaning of the text. The encoded text may be fed into a generative model, which produces an image that matches the text description. This process may include multiple layers of a neural network that progressively refine the image. Text-image model 104 may generate an image based on prompt 102A that provides a description. The generation of the first image is further described, for example, in Figure 3.

[0110] In 908, a set of attributes related to the description of the first image may be determined based on a first language model, where the set of attributes may correspond to semantics related to the first image. The processor 202 may be configured to determine a set of attributes 214A related to the description of the first image 116 based on a first language model 210. The set of attributes 214A may correspond to semantics related to the first image 116. In embodiments, the set of attributes 214A may include, but are not limited to, a person's age, a person's hair color, a person's facial expression, a person's body type, a person's height, a person's facial orientation, or a person's gender. Consider the example prompt “young girl with curly hair.” The set of attributes 214A may be, for example, “young,” “hair,” or “girl.” The determination of the set of attributes is further described, for example, in Figure 3.

[0111] In 910, a set of questions related to the prompt may be generated based on a second language model. The processor 202 may be configured to generate a set of questions related to the prompt 102A based on a second language model 212. The first language model 210 and the second language model 212 may correspond to an LLM. In one embodiment, the first language model 210 may be the same as the second language model 212. In another embodiment, the first language model 210 may be different from the second language model 212. The LLM may be an advanced AI system trained on large amounts of text data so that it can perform a wide range of natural language processing tasks such as translation, summarization, and text generation. The LLM may use, for example, a transformer architecture, which allows for efficient processing and generation of text. During training, the LLM may learn statistical relationships between words and phrases by analyzing language datasets. Through this training, the LLM learns how to determine the context, syntax, and meaning of any natural language text so that it can generate consistent and contextually relevant responses. Large-scale language models may include, but are not limited to, generative pre-trained transformers (GPT), bidirectional encoded representations with transformers (BERT), and text-to-text translation transformers (T5). The generation of question sets is further described, for example, in Figure 3.

[0112] In 912, the slider boundary value and initial slider value may be generated based on the set of questions associated with prompt 102A and the first image 116. The processor 202 may be configured to generate the slider boundary value and initial slider value based on the set of questions associated with prompt 102A and the first image 116. The generation of the slider boundary value and initial slider value is further described, for example, in Figure 3.

[0113] In 914, a set of sliders associated with a set of attributes can be formed based on slider boundary values ​​and initial slider values. Here, each slider in the set of sliders can be associated with a corresponding attribute in the set of attributes. The processor 202 may be configured to generate a set of sliders 214B associated with a set of attributes 214A based on the slider boundary values ​​and initial slider values. Each slider in the set of sliders 214B can be associated with a corresponding attribute in the set of attributes 214A. The slider boundary values ​​may indicate a range of slider values ​​from which the user can generate various images by changing the slider values. The initial slider values ​​may indicate the initial value of the slider from which the slider values ​​can start, and the user can generate various images by changing the slider values. The generation of slider boundary values ​​and initial slider values ​​may be based on a set of questions and a first image 116. The set of questions may be generated based on a first VQA model and a second VQA model. The first VQA model may be the same as the second VQA model. In some embodiments, the first VQA model may differ from the second VQA model. The first and second VQA scores may be compared to predetermined values ​​(e.g., a first predetermined value and a second predetermined value, respectively). The slider boundary value may be updated to the next value based on the first VQA score being less than the first predetermined value. The upper limit of the slider boundary value may be determined based on the first VQA score being greater than the first predetermined value.

[0114] In another embodiment, the second VQA score may even be compared to a second predetermined value. The initial slider value may be updated based on the second VQA score being less than the second predetermined value. The lower limit may be determined relative to the initial slider value based on the second VQA score being greater than the second predetermined value. The first and second predetermined values ​​may be the same. In some embodiments, the first and second predetermined values ​​may be different.

[0115] At 916, user input related to a set of sliders may be received. Processor 202 may be configured to receive user input related to a set of sliders 214B. The user may provide input by changing a set of slider values. The user may change one or more sliders in the set of sliders 214B. The reception of user input is further described, for example, in Figure 3.

[0116] In 918, a second image may be generated based on user input associated with a set of sliders and the first image. Processor 202 may be configured to generate a second image 118 based on user input associated with a set of sliders 214B and the first image 116. The second image 118 may be generated by changing the slider values. The user may change the slider values ​​based on requirements. In the example of "young girl with curly hair," text-image model 104 may generate an image based on the prompt "young girl with curly hair." The user may provide user input associated with a set of sliders 214B. The set of sliders 214B may, for example, be "age." The user may change the slider value for age based on requirements. For example, if the slider value was set to "4," the user may select a slider value of "2" to indicate that the user wants the girl in the image to be younger. The generation of the second image is further described, for example, in Figure 3.

[0117] At 920, the second image may be rendered. The processor 202 may be configured to render the second image 118. The display device 114 may display the first image 116 and the second image 118. The display device 114 may display real-time changes in slider values ​​based on user input. The display device 114 may display more than one image on the display simultaneously. The rendering of the second image is further described, for example, in Figure 4. Control can then be terminated.

[0118] While flowchart 900 is represented as individual operations such as 902, 904, 906, 908, 910, 912, 914, 916, 918, and 920, the disclosure is not limited in this way. However, in certain embodiments, such individual operations may be further divided into additional operations, combined into fewer operations, or deleted, depending on the particular implementation, without deviating from the essence of the disclosed embodiments.

[0119] Various embodiments of the present disclosure may provide one or more non-temporary computer-readable storage media configured to store instructions causing an electronic device (e.g., electronic device 102) to perform an action in response to being executed. The action may include receiving a prompt (e.g., prompt 102A) indicating a description of a first image to be generated (e.g., first image 116). The action may further include generating the first image 116 based on prompt 102A using a text-image model (e.g., text-image model 104). The action may include determining a set of attributes (e.g., attribute set 214A) that may be related to the description of the first image 116, based on a first language model (e.g., first language model 210). Furthermore, attribute set 214A may correspond to semantics related to the first image 116. The action may further include generating a set of questions related to prompt 102A using a second language model (e.g., second language model 212). The operation may include generating slider boundary values ​​and initial slider values ​​based on a set of questions associated with prompt 102A and the first image 116. The operation may further include generating a set of sliders (e.g., slider set 214B) associated with a set of attributes 214A based on the slider boundary values ​​and initial slider values. Each slider in slider set 214B may be associated with a corresponding attribute in attribute set 214A. The operation may include receiving user input associated with slider set 214B. The operation may further include generating a second image (e.g., second image 118) based on the user input associated with slider set 214B and the first image 116, and rendering the second image 118.

[0120] As used in this disclosure, the terms “module” or “component” may refer to a specific hardware implementation configured to perform the operation of a module or component, and / or a software object or software routine that can be executed and / or stored in general-purpose hardware of a computing system (e.g., computer-readable media, processing devices, etc.). In some embodiments, the various components, modules, engines, and services described herein may be implemented as objects or processes executed in a computing system (e.g., as separate threads). While some of the systems and methods described herein are generally described as being implemented in software (stored and / or executed by general-purpose hardware), specific hardware implementations or combinations of software and specific hardware implementations are also possible and intended. herein, “computation entity” may be any computing system as previously defined herein, or any module or combination of modules executed in a computing system.

[0121] In this disclosure, the terms used in particular in the attached claims (e.g., the text of the attached claims) are generally intended to be “open” terms (for example, the word “including” should be interpreted as “including but not limited to,” the word “having” should be interpreted as “at least having,” and the word “includes” should be interpreted as “including but not limited to,” etc.).

[0122] Furthermore, if a specific number is intended in an introduced claim recitation, that intention must be clearly stated in the claim; if there is no such statement, then no such intention exists. For example, to facilitate understanding, subsequent appended claims may use introductory phrases such as "at least one" and "one or more" to introduce a claim. However, the use of such phrases should not be interpreted as suggesting that a particular claim containing the introduced claim recitation is limited to cases that include only one instance of that item, even if the claim includes both an introductory phrase such as "one or more" or "at least one" and an indefinite article such as "a" or "an" (for example, "a" and / or "an" should be interpreted as meaning "at least one" or "one or more"). The same applies when introducing a claim recitation using a definite article.

[0123] Furthermore, even if a specific number is explicitly stated in the introduced claim description, it will be understood by those skilled in the art that such description should generally be interpreted to mean at least the number stated (for example, if there is a description of only "two descriptions" without any other modifiers, this description means at least two descriptions, or two or more descriptions). Furthermore, when a notation similar to "at least one of A, B, and C, etc." or "one or more of A, B, and C, etc." is used, such a structure is generally intended to include A only, B only, C only, both A and B, both A and C, both B and C, and / or all of A, B, and C, etc.

[0124] Furthermore, any disjunctions and / or disjunctions representing two or more selectable terms, whether in the specification, claims, or drawings, should be understood as intended to include the possibility of including one of those terms, either of those terms, or both of those terms. For example, the phrase "A or B" should be understood to include the possibility of "A or B" or "A and B".

[0125] All examples and conditional language cited herein are intended for educational purposes to help readers understand the concepts and inventions to which the inventors have contributed to the advancement of the art, and should be construed as not being limited to such specifically cited examples and conditions. While embodiments of this disclosure have been described in detail, various modifications, substitutions, and alternatives may be made without departing from the spirit and scope of this disclosure.

[0126] In addition to the embodiments described above, the following additional information is disclosed. (Note 1) The way in which a processor executes You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, the slider boundary value and initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the previous second image and A method of having. (Note 2) The set of questions includes at least one of the following: a first set of questions corresponding to questions that do not depend on the slider, and a second set of questions corresponding to questions that do depend on the slider. The generation of the slider boundary value is based on the set of questions in the first question, The generation of the initial slider value is based on the second set of questions, The method described in Appendix 1. (Note 3) The aforementioned text-image model corresponds to a Generative Adversarial Network (GAN) model. The method described in Appendix 1. (Note 4) The first language model and the second language model each correspond to a large-scale language model (LLM). The method described in Appendix 1. (Note 5) The first language model is the same as the second language model. The method described in Appendix 1. (Note 6) The first language model differs from the second language model. The method described in Appendix 1. (Note 7) The aforementioned set of attributes is, A person's age, The aforementioned person's hair color, The facial expression of the aforementioned person, The aforementioned person's body type, The height of the aforementioned person, The direction of the person's face, or The gender of the aforementioned person Corresponding to at least one of the following, The method described in Appendix 1. (Note 8) The generation of the aforementioned slider boundary values ​​corresponds to the first visual question answering (VQA) model, The generation of the initial slider value corresponds to the second VQA model. The method described in Appendix 1. (Note 9) The first VQA model is the same as the second VQA model. The method described in Appendix 8. (Note 10) The first VQA model is different from the second VQA model. The method described in Appendix 8. (Note 11) The third image set is generated using a low-rank adaptive (LoRA) model, The first VQA model determines the first VQA score based on the set of third images and the set of questions, The first VQA score is compared with a first predetermined value, Based on the fact that the first VQA score is smaller than the first predetermined value, the slider boundary value is updated to the following value, Based on the fact that the first VQA score is greater than the first predetermined value, the upper limit of the slider boundary value is determined. The method described in Appendix 8, further comprising the above. (Note 12) The second VQA score is determined based on the set of third images and the set of questions using the second VQA model, The second VQA score is compared with a second predetermined value, If the slider boundary value is updated to the previous value based on the fact that the second VQA score is smaller than the second predetermined value, Based on the fact that the second VQA score is greater than the second predetermined value, the lower limit of the slider boundary value is determined. The method described in Appendix 11, further comprising the above. (Note 13) The process involves determining a trained Perceptual Image Patch Similarity (LPIPS) score for each third image in the aforementioned set of third images, where each third image corresponds to a slider value. To estimate the LPIPS curve for each third image in the set of the third images based on the LPIPS score, To determine the mapping corresponding to the linear function associated with the LPIPS curve, Based on the mapping, a normalized value is determined for the slider value corresponding to each third image in the set of third images. The method described in Appendix 11, further comprising the above. (Note 14) When executed, it is configured to store instructions that cause the system to perform actions. The aforementioned operation is, You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, the slider boundary value and initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the previous second image and Non-temporary computer-readable storage media, including [specific data / information]. (Note 15) The set of questions includes at least one of the following: a first set of questions corresponding to questions that do not depend on the slider, and a second set of questions corresponding to questions that do depend on the slider. The generation of the slider boundary value is based on the set of questions in the first question, The generation of the initial slider value is based on the second set of questions, A non-temporary computer-readable storage medium as described in Appendix 14. (Note 16) The generation of the aforementioned slider boundary values ​​corresponds to the first visual question answering (VQA) model, The generation of the initial slider value corresponds to the second VQA model. A non-temporary computer-readable storage medium as described in Appendix 14. (Note 17) The aforementioned operation is, The third image set is generated using a low-rank adaptive (LoRA) model, The VQA score is determined based on the set of the third images and the set of questions using the first VQA model, The VQA score is compared with a first predetermined value, Based on the fact that the VQA score is smaller than the first predetermined value, the slider boundary value is updated to the following value, Based on the fact that the VQA score is greater than the first predetermined value, the upper limit of the slider boundary value is determined. This also includes, A non-temporary computer-readable storage medium as described in Appendix 16. (Note 18) The aforementioned operation is, The VQA score is compared with a second predetermined value, If the slider boundary value is updated to the previous value based on the fact that the VQA score is smaller than the second predetermined value, Based on the fact that the VQA score is greater than the second predetermined value, the lower limit of the slider boundary value is determined. This also includes, A non-temporary computer-readable storage medium as described in Appendix 17. (Note 19) The aforementioned operation is, The process involves determining a trained Perceptual Image Patch Similarity (LPIPS) score for each third image in the aforementioned set of third images, where each third image corresponds to a slider value. To estimate the LPIPS curve for each third image in the set of the third images based on the LPIPS score, To determine the mapping corresponding to the linear function associated with the LPIPS curve, Based on the mapping, a normalized value is determined for the slider value corresponding to each third image in the set of third images. This also includes, A non-temporary computer-readable storage medium as described in Appendix 17. (Note 20) Memory configured to store instructions, A processor coupled to the aforementioned memory and configured to execute instructions and perform a process, It has, The aforementioned process, You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, the slider boundary value and initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the previous second image and Electronic devices, including those mentioned above. [Explanation of symbols]

[0127] 102 Electronic Devices 102A Prompt 104 Text-Image Model 106 Communication Networks 108 servers 110 Databases 112 Generated Images 114 Display Devices 116 Image 1 118,320 Image 2 202 processors 204 memory 206 I / O devices 208 Network Interfaces 210 First Language Model 212 Second Language Model 214A,308 Attribute Sets 214B Slider Set

Claims

1. The way in which a processor executes You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, a slider boundary value and an initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the aforementioned second image and A method of having.

2. The set of questions includes at least one of the following: a first set of questions corresponding to questions that do not depend on the slider, and a second set of questions corresponding to questions that do depend on the slider. The generation of the slider boundary value is based on the set of questions in the first question, The generation of the initial slider value is based on the second set of questions, The method according to claim 1.

3. The aforementioned text-image model corresponds to a Generative Adversarial Network (GAN) model. The method according to claim 1.

4. The first language model and the second language model each correspond to a large-scale language model (LLM). The method according to claim 1.

5. The first language model is the same as the second language model. The method according to claim 1.

6. The first language model is different from the second language model. The method according to claim 1.

7. The aforementioned set of attributes is, A person's age, The aforementioned person's hair color, The facial expression of the aforementioned person, The aforementioned person's body type, The height of the aforementioned person, The direction of the person's face, or The gender of the aforementioned person Corresponding to at least one of the following, The method according to claim 1.

8. The generation of the slider boundary values ​​corresponds to the first visual question answering (VQA) model, The generation of the initial slider value corresponds to the second VQA model. The method according to claim 1.

9. The first VQA model is the same as the second VQA model. The method according to claim 8.

10. The first VQA model is different from the second VQA model. The method according to claim 8.

11. The third image set is generated using a low-rank adaptive (LoRA) model, The first VQA model determines the first VQA score based on the set of third images and the set of questions, Comparing the first VQA score with a first predetermined value, Based on the fact that the first VQA score is smaller than the first predetermined value, the slider boundary value is updated to the following value, Based on the fact that the first VQA score is greater than the first predetermined value, the upper limit of the slider boundary value is determined. The method according to claim 8, further comprising:

12. The second VQA model determines the second VQA score based on the set of third images and the set of questions, The second VQA score is compared with a second predetermined value, If the slider boundary value is updated to the previous value based on the fact that the second VQA score is smaller than the second predetermined value, Based on the fact that the second VQA score is greater than the second predetermined value, the lower limit of the slider boundary value is determined. The method according to claim 11, further comprising:

13. The process involves determining a trained perceptual image patch similarity (LPIPS) score for each third image in the aforementioned set of third images, where each third image corresponds to a slider value. For each third image in the set of the third images, the LPIPS curve is estimated based on the LPIPS score, To determine the mapping corresponding to the linear function associated with the LPIPS curve, Based on the mapping, a normalized value is determined for the slider value corresponding to each third image in the set of third images. The method according to claim 11, further comprising:

14. When executed, it is configured to store instructions that cause the system to perform actions. The aforementioned operation is, You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, a slider boundary value and an initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the aforementioned second image and Non-temporary computer-readable storage media, including [specific data / information].

15. The set of questions includes at least one of the following: a first set of questions corresponding to questions that do not depend on the slider, and a second set of questions corresponding to questions that do depend on the slider. The generation of the slider boundary value is based on the set of questions in the first question, The generation of the initial slider value is based on the second set of questions, A non-temporary computer-readable storage medium according to claim 14.

16. The generation of the slider boundary values ​​corresponds to the first visual question answering (VQA) model, The generation of the initial slider value corresponds to the second VQA model. A non-temporary computer-readable storage medium according to claim 14.

17. The aforementioned operation is, The third image set is generated using a low-rank adaptive (LoRA) model, The VQA score is determined based on the set of the third images and the set of questions using the first VQA model, The VQA score is compared with a first predetermined value, Based on the fact that the VQA score is smaller than the first predetermined value, the slider boundary value is updated to the following value, Based on the fact that the VQA score is greater than the first predetermined value, the upper limit of the slider boundary value is determined. This also includes, A non-temporary computer-readable storage medium according to claim 16.

18. The aforementioned operation is, The VQA score is compared with a second predetermined value, If the slider boundary value is updated to the previous value based on the fact that the VQA score is smaller than the second predetermined value, Based on the fact that the VQA score is greater than the second predetermined value, the lower limit of the slider boundary value is determined. This also includes, A non-temporary computer-readable storage medium according to claim 17.

19. The aforementioned operation is, The process involves determining a trained perceptual image patch similarity (LPIPS) score for each third image in the aforementioned set of third images, where each third image corresponds to a slider value. For each third image in the set of the third images, the LPIPS curve is estimated based on the LPIPS score, To determine the mapping corresponding to the linear function associated with the LPIPS curve, Based on the mapping, a normalized value is determined for the slider value corresponding to each third image in the set of third images. This also includes, A non-temporary computer-readable storage medium according to claim 17.

20. Memory configured to store instructions, A processor coupled to the aforementioned memory and configured to execute instructions and perform a process, It has, The aforementioned process, You will receive a prompt indicating a description of the first image that will be generated, The text-image model generates the first image based on the prompt, Based on a first language model, a set of attributes related to the description of the first image is determined, and the set of attributes corresponds to the semantics associated with the first image. Based on a second language model, generate a set of questions related to the prompt, Based on the set of questions related to the prompt and the first image, a slider boundary value and an initial slider value are generated. Based on the slider boundary value and the initial slider value, a set of sliders associated with the set of attributes is generated, wherein each slider in the set of sliders is associated with the corresponding attribute in the set of attributes. Receiving user input related to the aforementioned set of sliders, A second image is generated based on the user input and the first image related to the set of sliders, Rendering the aforementioned second image and Electronic devices, including those mentioned above.

Images