Support Server for Improving Computational Efficiency of Generative AI Models through Prompt Similarity Matching and Method thereof

Abstract

The present invention relates to a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, wherein the server searches for and presents candidate images through similarity matching between an input prompt and a generation condition prompt of a stored base image, transmits the selected image and the prompt to a generative AI model, and receives a final image generated from the generative AI model and provides it to an external terminal. Here, the support server for improving the computational efficiency of a generative AI model through prompt similarity matching comprises: a mapping storage module that mutually maps and stores a base image and a generation condition prompt of the base image; a prompt module that receives a prompt from an external terminal; a candidate group presentation module that calculates the similarity between the prompt and the generation condition prompt and presents a base image with a similarity greater than or equal to a reference similarity as a candidate image to the external terminal; and a generation control module that receives a selected image chosen from the external terminal among the candidate images, transmits the selected image and the prompt to at least one generative AI model to request the generation of a final image reflecting the prompt, and then receives the generated final image and transmits it to the external terminal. A method for improving the computational efficiency of a generative AI model through prompt similarity matching comprises: (a) a step of mutually mapping and storing a base image and a generation condition prompt of the base image; (b) a step of receiving a prompt from an external terminal connected via a communication network; (c) a step of calculating the similarity between the prompt and the generation condition prompt and presenting a base image with a similarity greater than or equal to a preset reference similarity as a candidate image to the external terminal; (d) a step of receiving a selected image chosen by the external terminal among the candidate images; and (e) a step of transmitting the selected image and the prompt to at least one generative AI model to request the generation of a final image reflecting the prompt, and then receiving the generated final image and transmitting it to the external terminal.

Description

Technology Field

[0001] The present invention relates to a technology for searching for a pre-stored image generated by a generative AI model and presenting it to an external terminal. More specifically, the present invention provides a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, which searches for and presents candidate images by matching similarity between an input prompt and a generation condition prompt of a pre-stored base image, transmits the selected image and prompt to the generative AI model, and then receives the final image generated by the generative AI model and provides it to an external terminal. Background Technology

[0002] With the recent advancement of generative AI technologies based on diffusion models, services that generate high-quality images based on user text input—that is, prompts—are being commercialized. A typical Text-to-Image generation method features a structure where, upon user inputting a prompt, an AI model generates an image by starting from a state of random noise and removing the noise through an iterative process.

[0003] However, these existing technologies modify the prompt multiple times and regenerate the image to obtain the desired result. This process of repeatedly modifying the prompt and regenerating the image causes a problem of unnecessarily consuming GPU computational resources. Furthermore, generating a high-resolution image for each request increases the time required to obtain a satisfactory image for the user. Additionally, due to the probabilistic nature of the generative models in existing technologies, they generate different results each time even when the same or similar prompts are entered.

[0004] Accordingly, existing generative models have a problem in that it is difficult to reliably reproduce or control the desired results when users want to perform fine modifications to the generated output, such as inpainting or editing, or maintain a consistent style.

[0005] Furthermore, even if a user attempts to reproduce an image created in the past after a certain period of time, if the creation conditions at that time are not reproduced identically, there is a problem in that it is difficult to obtain the desired image again.

[0006] Therefore, there is a need for a device and method that can consistently and stably provide results that meet the user's intent, while reducing unnecessary computational resource consumption and latency caused by the iterative generation process starting from random noise without a selected image. Prior art literature

[0007] Republic of Korea Published Patent No. 10-2025-0179328 'AI-generated image evaluation device and method' (Publication date: Dec. 30, 2025) The problem to be solved

[0008] The objective of the present invention, devised to solve the aforementioned problems, is to provide a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, which searches for and presents candidate images through similarity matching between an input prompt and a generation condition prompt of a stored base image, transmits the selected image and prompt to a generative AI model to receive the generated final image, and provides it to an external terminal.

[0009] Furthermore, the objective of the present invention is to provide a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can improve the computational efficiency of a generative AI model by enabling the generative AI model to avoid a zero-base image generation process through prompt similarity matching and by reducing the consumption of computational resources and image generation delay time caused by repetitive prompt modification and image regeneration.

[0010] In addition, the objective of the present invention is to provide a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can reduce computational resource usage and infrastructure construction costs by deriving and presenting candidate images through low-cost similarity search without high-cost image generation operations in the step of presenting candidate images.

[0011] In addition, the objective of the present invention is to provide a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can reduce user drop-off caused by long waiting times by rapidly presenting candidate images to shorten the waiting time occurring during the user's image acquisition process.

[0012] In addition, the objective of the present invention is to provide a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can generate a final image that clearly reflects the user's intention by combining a prompt with a selected image chosen by the user and controlling a deformation strength parameter to adjust the degree of feature preservation of the selected image.

[0013] Furthermore, the objective of the present invention is to provide a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, wherein the final image, the corresponding prompt, and generation parameters are stored in a mapping storage module to continuously expand the searchable images and images that can be presented as candidate groups as operation time elapses, and as a result, candidate images similar to the prompt can be reused and reproduced for the same or similar prompts.

[0014] The problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0015] A support server for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention for achieving the above-mentioned problem comprises: a mapping storage module that stores a base image and a generation condition prompt of the base image by mutually mapping them; a prompt module that receives a first prompt used to calculate similarity with the generation condition prompt from an external terminal, and receives a second prompt used to edit the selected image after receiving a selected image chosen from the external terminal among candidate images; a candidate group presentation module that calculates the similarity between the first prompt and the generation condition prompt and presents a base image with a similarity greater than or equal to a reference similarity to the external terminal as a candidate image; and a generation control module that receives the selected image chosen from the external terminal among the candidate images, transmits the selected image and the second prompt to at least one generative AI model to request the generation of a final image in which the second prompt is reflected in the selected image, and then receives the generated final image and transmits it to the external terminal.

[0016] It may further include a storage update module that collects creation history information including a final image and a prompt used to generate the final image, structures the collected creation history information into a creation history record, or updates the creation history record by mapping additional creation history information to the structured creation history record and stores it in a mapping storage module.

[0017] The candidate presentation module may include an embedding vector generation unit that generates a prompt embedding vector for a prompt, generates a generation condition prompt embedding vector for a generation condition prompt stored in a mapping storage module, or loads a previously stored generation condition prompt embedding vector, and a candidate presentation unit that presents a base image mapped with one or more generation condition prompts having a similarity score equal to or greater than a reference similarity as a candidate group.

[0018] The prompt module can receive from an external terminal a prompt including a first prompt used to calculate similarity with a generation condition prompt and a second prompt used to edit a selected image, and an additional reference image used for editing.

[0019] The generation control module can set or change a deformation strength parameter that determines the reflection level of the second prompt for the selected image.

[0020] The generation control module can lower the preservation strength of the selected image and increase the allowable degree of deformation according to the second prompt as the deformation strength parameter increases, and increase the preservation strength of the selected image and decrease the allowable degree of deformation according to the second prompt as the deformation strength parameter decreases, thereby enabling the final image to be generated in the corresponding generative AI model.

[0021] The generation control module compares the determined deformation strength parameter with a preset editing strength threshold value, and if the deformation strength parameter is greater than or equal to the preset editing strength threshold value, sets the preservation weight of the selected image to be smaller than the second prompt reflection weight so that a final image with the second prompt reflected in the selected image is generated, and if the deformation strength parameter is less than the preset editing strength threshold value, sets the preservation weight of the selected image to be larger than the second prompt reflection weight so that a final image in which the selected image is retained is generated.

[0022] The generation control module outputs a control element corresponding to the strain strength parameter to an external terminal, and can set or change the value of the strain strength parameter according to a control signal for the control element input from the external terminal.

[0023] The mapping storage module can structure the generation condition prompts used for generating multiple base images and the generation parameters used for generating multiple base images into a generation history record, map them to the base images, and store them.

[0024] The prompt module provides multiple selection elements to an external terminal, each having a different prompt set therein, and when a selection element is selected from the external terminal, it can receive the prompt pre-set for the selected selection element as a prompt.

[0025] A selection element can be mapped to at least one of a predefined phrase, tag, or parameter. Additionally, the prompt module can generate a prompt by combining selection signals for multiple selection elements and determine the placement order or application weight of each prompt component within the prompt based on the selection order or the number of selections for the multiple selection elements.

[0026] A method for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention for achieving another problem to be solved, wherein a support server improving the computational efficiency of a generative AI model through prompt similarity matching improves the computational efficiency of a generative AI model, comprising: (a) a step of mutually mapping and storing a base image and a generation condition prompt of the base image; (b) a step of receiving a first prompt used for calculating similarity with the generation condition prompt from an external terminal connected via a communication network; (c) a step of calculating similarity between the first prompt and the generation condition prompt, and presenting the base image, in which the similarity is greater than or equal to a preset reference similarity, as a candidate image to the external terminal; (d) a step of receiving a selected image selected from the candidate images by the external terminal; and (e) a step of receiving a second prompt used for editing the selected image from the external terminal. and (f) a step of transmitting the selected image and the second prompt to at least one generative AI model to request the generation of a final image in which the second prompt is reflected in the selected image, and then receiving the generated final image and transmitting it to the external terminal;

[0027] delete

[0028] delete

[0029] delete Effects of the invention

[0030] The present invention provides a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, which searches for and presents candidate images by matching similarity between an input prompt and a generation condition prompt of a stored base image, transmits the selected image and prompt to a generative AI model, and then receives the final image generated by the generative AI model and provides it to an external terminal.

[0031] In addition, the present invention provides a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can improve the computational efficiency of a generative AI model by enabling the generative AI model to avoid a zero-base image generation process through prompt similarity matching and by reducing the consumption of computational resources and image generation delay time caused by repetitive prompt modification and image regeneration.

[0032] In addition, the present invention provides a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can reduce computational resource usage and infrastructure construction costs by deriving and presenting candidate images through low-cost similarity search without high-cost image generation operations in the step of presenting candidate images.

[0033] In addition, the present invention provides a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can reduce user drop-off caused by long waiting times by rapidly presenting candidate images to shorten the waiting time that occurs during the user's image acquisition process.

[0034] In addition, the present invention provides a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, which can generate a final image that clearly reflects the user's intention by combining a prompt with a selected image chosen by the user and controlling a deformation strength parameter to adjust the degree of feature preservation of the selected image.

[0035] In addition, the present invention provides a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching, wherein the final image, the corresponding prompt, and generation parameters are stored in a mapping storage module to continuously expand the searchable images and images that can be presented as candidate groups as operation time elapses, and as a result, candidate images similar to the prompt can be reused and reproduced for the same or similar prompts. Brief explanation of the drawing

[0036] FIG. 1 is a block diagram of a support server that improves the computational efficiency of a generative AI model through prompt similarity matching according to an embodiment of the present invention. FIG. 2 is a diagram showing the relationship between a support server and an external terminal that improves the computational efficiency of a generative AI model through prompt similarity matching according to an embodiment of the present invention. FIG. 3 is a flowchart of a method for improving the computational efficiency of a generative AI model through prompt similarity matching according to an embodiment of the present invention. Figure 4 is a diagram showing the mapping storage module of Figure 1. Figure 5 is a diagram showing the state in which a prompt is input to the prompt module of Figure 1. FIG. 6 is a diagram showing the state in which one of the candidate images presented by the candidate group presentation module of FIG. 1 is selected from an external terminal, and the prompt module receives a prompt from the external terminal. Figure 7 is a diagram showing the state in which the generation control module of Figure 1 receives a final image from a generative AI model and transmits it to an external terminal. Figure 8 is a diagram showing the state in which the storage update module of Figure 1 stores the final image in the mapping storage module. FIGS. 9 to 12 are drawings showing the selection elements provided by the prompt module of FIG. 1. Specific details for implementing the invention

[0037] The advantages and features derived by the present invention, and the technical configurations for achieving them, become clear through the embodiments described below in combination with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. The embodiments of this specification are provided to fully convey the technical concept of the invention to those skilled in the art, and the actual scope of protection of the present invention is defined solely by the description in the claims.

[0038] Therefore, specific embodiments or drawings described in this specification cannot be used as grounds to limit the interpretation of the invention.

[0039] Additionally, as used herein, the terms “part” or “module” refer to software or hardware components, and “part” or “module” perform the roles defined for each. However, “part” or “module” is not limited to software or hardware. “Part” or “module” may be configured to reside in an addressable storage medium or may be configured to run one or more processors. For example, “part” or “module” may include components such as object-oriented software components, class components, and task components, and at least one of processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, or variables. Such “part” or “module” may be combined with a smaller number of components and other “parts” or “modules” provided therein, or further separated into additional components and “parts” or “modules.” In addition, singular expressions in this specification include plural expressions unless the context clearly indicates that they are singular, and plural expressions also include singular expressions unless the context clearly indicates that they are plural.

[0040] Hereinafter, with reference to FIGS. 1 to 12, a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching according to an embodiment of the present invention will be described in detail.

[0041] However, for the sake of brevity and clarity in the description of the present invention, the overall configuration and operation of a support server that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention will first be described in general terms with reference to FIG. 1. Subsequently, with reference to FIG. 2 and FIG. 3, a support server that improves the computational efficiency of a generative AI model through prompt similarity matching and a method thereof according to an embodiment of the present invention will be described.

[0042] A support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to an embodiment of the present invention searches for and presents candidate images through similarity matching between an input prompt and a generation condition prompt of a stored basic image, transmits the selected image and the prompt to a generative AI model, receives the final image generated by the generative AI model, and provides it to an external terminal.

[0043] The support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention calculates the similarity between an input prompt and a generation condition prompt of a stored base image through prompt similarity matching, derives a base image in which the similarity is greater than or equal to a reference similarity, and provides the base image as initial input data for initiating computation of the generative AI model. At this time, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention avoids the zero-base inverse diffusion process in which the generative AI model reconstructs all pixels starting from random noise, and performs incremental generation that reflects deformation elements according to the input prompt by utilizing latent spatial information corresponding to the base image as an initial value.

[0044] The image generation method of the support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention reduces the number of inference steps required to reach the final image, alleviates the occupation of GPU computational resources due to repetitive prompt modifications, and reduces the latency required for generating high-resolution images. Furthermore, by reducing the number of GPUs and the scale of server expansion required to achieve the same throughput, the infrastructure construction costs for providing generative AI services are reduced.

[0045] Furthermore, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention can precisely control the degree of preservation of visual features such as composition, style, color tone, and object placement of a selected image according to a deformation strength parameter, thereby enabling the generation of an image in which the user's editing intention is precisely reflected in the generative AI model. In addition, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention may generate a video with an additional prompt as an object, i.e., the main character, after setting the selected image as a background image. At this time, the present invention enables fine editing of the video as well as the image.

[0046] A support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention provides a candidate image selected based on an input prompt to an external terminal (B) and receives a selected image among the candidate images from the external terminal (B). Subsequently, among the input prompts, it identifies and extracts the remaining elements excluding the elements corresponding to the search conditions of the candidate image as editing attributes, and transmits the editing attributes along with the selected image to a generative AI model (G) to request the generation of a final image in which the editing attributes are incrementally reflected in the selected image. Alternatively, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention may receive a first prompt used for selecting candidate images and a second prompt that is incrementally reflected in the selected image from the external terminal (B), provide the candidate image selected based on the first prompt to the external terminal (B), and then request the generative AI model (G) to generate a final image in which the second prompt is incrementally reflected in the selected image.

[0047] Alternatively, a support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention may request generation by transmitting a selected image, a second prompt, and a reference image to a generative AI model (G) so that a final image is generated in which the second prompt and the reference image are incrementally reflected in the selected image. For example, when a first prompt, “a photo of a woman holding a handbag,” is input, the support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention may present a photo of a woman holding a handbag corresponding to the first prompt as several candidate images. Then, when a selected image selected from an external terminal and a second prompt, “Generate an additional reference image with a different handbag drawn on it, and a photo in which the handbag in the selected image is replaced with the handbag drawn in the additional reference image,” are input, the support server may request the generative AI model to generate a final image in which the content of the second prompt is incrementally reflected in the selected image. In addition, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching of the present invention may identify the generative AI model (G) that generated the base image based on identification information, i.e., metadata, of the base image collected from the generative AI model (G), and then transmit a selected image and a prompt to the identified generative AI model so that a final image is generated in which an incremental editing prompt is reflected in the selected image. Here, the identification information is information including at least one of the server ID, API endpoint, model ID, and model version of the generative AI model.

[0048] The support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention can quickly provide a final image to an external terminal without its own GPU computation load by calling and routing a pre-established generative AI model without performing zero-base computation starting from random noise.

[0049] The support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention includes a mapping storage module (110), a prompt module (120), a candidate group presentation module (130), a generation control module (140), and a storage update module (150), as shown in FIG. 1.

[0050] To explain in more detail the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching according to the present invention, as shown in FIG. 2, the present invention searches for and presents candidate images from a mapping storage module (110) based on vector similarity for a prompt input from an external terminal (B) based on a basic image collected from a plurality of generative AI models (G). Subsequently, when a selected image is determined from the external terminal among the candidate images, the prompt input from the external terminal (B), an additional reference image, and deformation strength parameters are transmitted to the generative AI model (G) to request the generative AI model (G) to generate a final image in which the prompt is reflected in the selected image, and the generated final image is received and provided to the external terminal (B). In this way, the support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching of the present invention can be connected to communicate with an external terminal (B) and a generative AI model (G) via a communication network (A). Here, the generative AI model (G) may include an artificial intelligence image generation model trained to perform image generation of various styles. Additionally, in this specification, the generative AI model (G) may be used as a concept encompassing the image generation model and a server or service where the image generation model is executed, and depending on the implementation, it may be provided by an external server or executed within the support server (100).

[0051] In addition, the method for improving the computational efficiency of a generative AI model through prompt similarity matching, which is conducted by a support server that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention, as illustrated in FIG. 3, includes a series of steps: (a) a step (S110) of mapping and storing a generation condition prompt to a base image; (b) a step (S120) of receiving a prompt from an external terminal (B); (c) a step (S130) of presenting a candidate image; (d) a step (S140) of receiving a selected image chosen by the external terminal among the candidate images; and (e) a step (S150) of receiving a final image generated by the generative AI model and presenting it to the external terminal. Additionally, it may include a step (f) (S160) of storing the final image in a mapping storage module. At this time, if the prompt is divided into a first prompt and a second prompt, step (b) (S120) may be a step of receiving the first prompt used to calculate similarity with the generation condition prompt, and step (c) (S130) may be a step of generating a prompt embedding vector for the first prompt, generating a generation condition prompt embedding vector for the generation condition prompt stored in the mapping storage module (110), or loading a previously stored generation condition prompt embedding vector, calculating the similarity between the first prompt and the generation condition prompt, and presenting a base image mapped to one or more generation condition prompts whose similarity score is greater than or equal to the reference similarity as a candidate group. And step (d) (S140) may be a step of receiving a second prompt used to edit the selected image and an additional reference image used for editing from an external terminal after receiving the first prompt. And (e) step (S150) may be a step of requesting a generative AI model to generate a final image for a selected image and a second prompt, and after receiving the final image generated from the generative AI model, presenting it to an external terminal.

[0052] Hereinafter, with reference to FIGS. 4 to 8, a mapping storage module (110), a prompt module (120), a candidate group presentation module (130), a generation control module (140), and a storage update module (150) included in a support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention, and steps that are the subject of each module will be described.

[0053] When the mapping storage module (110) stores the base image (C), it may store the base image (C) by mapping a source identification tag to the base image (C) to distinguish the source of the base image (C). Here, the source identification tag may be set based on at least one of the collection path of the base image, file header / EXIF, model identification information of the generation history record, or user input. Alternatively, if the base image (C) is a captured image, it may be set as a captured image, or if it is an image generated by a generative AI, it may be set as an AI image.

[0054] Additionally, the mapping storage module (110) can store a generated image (basic image or final image) from a stored basic image or at least one pre-configured generative AI model (G) and the generated condition prompt of the image by mutual mapping. At this time, the mapping storage module (110) can store the generated history information, which includes the prompt and generation parameters used to generate the image, by structuring it into metadata in the form of a generation history record and then mapping it to the image. For example, the mapping storage module (110) stores the basic image (C) and the generated condition prompt of the basic image (C) by mutual mapping. Furthermore, identification information regarding the generative AI model (G) that generated the basic image (C) can be included in the generated history information and stored. Additionally, when a base image (C) is selected and generated as a final image (F), the mapping storage module (110) may generate a generation history record containing the prompt and deformation strength parameters used until the base image (C) is generated as a final image (F), and then map and store it on the base image (C). For example, the mapping storage module (110) may map and store a first generation history record for the first base image (C1), which includes at least one of the following: image identification information of Image ID=IMG_001, a generation condition prompt of 'A+B+C' as shown in FIG. 4, or a keyword tag constituting the generation condition prompt, a seed value of S=12345, generative AI model identification information such as Model ID=M1, sampling-related parameters such as the number of steps and guidance scale, and post-processing parameters such as resolution and upscaling coefficient.Additionally, the mapping storage module (110) can map and store a second generation history record to the second base image (C2), which includes at least one of the following: image identification information of Image ID=IMG_002, a generation condition prompt of 'A+B+D' as shown in FIG. 4, or a keyword tag constituting the generation condition prompt, a seed value such as S=67890, generative AI model identification information such as Model ID=M2, sampling-related parameters such as the number of steps and guidance scale, and post-processing parameters such as resolution and upscaling coefficient.

[0055] By structuring and storing the creation history records in this manner, the mapping storage module (110) can quickly load the image creation context when searching for prompt similarity in the future, and can reuse the loaded creation context as basic data for incremental reflection requests. Here, the creation context may include a creation condition prompt, a seed, model identification information, sampling-related parameters, and post-processing parameters.

[0056] The prompt module (120) can receive a prompt from an external terminal (B) (S120). At this time, the prompt may include a first prompt used to calculate similarity with a generation condition prompt and a second prompt used to edit a selected image. Such first and second prompts may be transmitted together as a single prompt, or they may be transmitted separately at different times. That is, the second prompt may be transmitted after the first prompt has been transmitted. In addition, when the second prompt is transmitted, an additional reference image used for editing the second prompt may be transmitted together to the prompt module (120).

[0057] The candidate group presentation module (130) calculates the similarity between the prompt and the generation condition prompt as shown in FIG. 6, and presents a base image with a calculated similarity greater than or equal to the reference similarity as a candidate image (D) to an external terminal (B) (S130). In this case, the candidate group presentation module (130) can calculate the similarity between the first prompt and the generation condition prompt when the prompt is the first prompt.

[0058] The candidate group presentation module (130) derives the top N candidate images (D1~D5) whose similarity is greater than or equal to the reference similarity, and presents the derived candidate images in the form of thumbnails to an external terminal (B), thereby allowing the user to select the image closest to their intention. For example, the candidate group presentation module (130) can derive the top 5 candidate images (D1~D5) whose similarity is greater than or equal to the reference similarity by using a Sentence-BERT or CLIP model to calculate the vector similarity between the prompt and the generation condition prompt stored in the mapping storage module (110).

[0059] Such a candidate group presentation module (130) may include an embedding vector generation unit (1301) and a candidate group presentation unit (1302). Here, the embedding vector generation unit (1301) generates an embedding vector for a prompt and generates a generation condition prompt embedding vector for each generation condition prompt stored in the mapping storage module (110), or loads a generation condition prompt embedding vector previously stored in the mapping storage module (110). If necessary, the embedding vector generation unit (1301) may generate embedding vectors for a first prompt and a second prompt, respectively, to update storage or enhance search, and link them to a generation history record. At this time, the embedding vector generation unit (1301) may calculate a vector similarity score between the first prompt embedding vector and the generation condition prompt embedding vector. For example, the embedding vector generation unit (1301) can calculate a vector similarity score using at least one of cosine similarity, inner product-based similarity, distance-based similarity, or a variation thereof, and can determine that the higher the calculated score, the higher the semantic similarity between the first prompt and the generation condition prompt. Here, the embedding vector is a value that quantifies the sentence meaning of the prompt and expresses it as coordinates in a multidimensional vector space, and is used as input data to calculate the semantic similarity between prompts. For example, the embedding vector generation unit (1301) can generate a first prompt embedding vector by inputting the first prompt, “a photo of three toiletries displayed on a handkerchief,” into the embedding model. Additionally, the embedding vector generation unit (1301) can generate a generation condition prompt embedding vector by inputting a generation condition prompt, for example, “A+B+C” and “A+B+D” stored in the mapping storage module (110), into the embedding model, or load a generation condition prompt embedding vector previously stored in the mapping storage module (110).

[0060] The candidate group presentation unit (1302) can set a base image mapped to one or more generation condition prompts, for which the similarity score calculated by the embedding vector generation unit (1301) is greater than or equal to the reference similarity, as a candidate group and present the candidate group as a candidate image to an external terminal (B). Additionally, when a reference image is input instead of or together with a text prompt, the candidate group presentation unit (1302) can calculate the visual similarity between the reference image and the base image within the mapping storage module (110) using a CLIP (Contrastive Language-Image Pre-training) model and derive an optimal candidate image based on the visual similarity. That is, the candidate group presentation module (130) can derive and generate an optimal candidate image through a search that performs at least one of a text-based semantic similarity search and an image-based visual similarity search.

[0061] Even if such a candidate group presentation module (130) distinguishes between a captured image and an AI image by a source distinction tag, the candidate group search and presentation operation based on the similarity between the first prompt and the generation condition prompt can be performed in the same way.

[0062] The generation control module (140) receives a selected image (E) selected from an external terminal (B) among candidate images and transmits the selected image (E) and a prompt to at least one generative AI model (G). At this time, the generation control module (140) may transmit the selected image (E) and a second prompt to a generative AI model (G) corresponding to the identification information of the generative AI model (G) mapped to the selected image (E). For example, the generation control module (140) may determine the generative AI model corresponding to the identification information as the generative AI model (G) that processes the selected image and the prompt. For example, when the generation control module (140) determines that the first generative AI model (G1) is the server to process the selected image (E) and the prompt, it requests the first generative AI model (G1) to generate a final image by reflecting the prompt in the selected image (E). Then, the final image (F) is received from the first generative AI model (G1) and transmitted to an external terminal (B) (S150). Furthermore, the generation control module (140) includes an editing intensity control unit (1401) that determines the level of reflection of the prompt for the selected image (E), and can set or change the deformation intensity parameter of the selected image (E). More specifically, the editing intensity control unit (1401) compares the set deformation intensity parameter with a preset editing intensity threshold value (e.g., 0.6) and, based on the comparison result, controls so that the degree of preservation of the selected image (E) and the degree of reflection of the prompt do not conflict. For example, when the deformation strength parameter is set to be greater than or equal to the editing strength threshold, the editing strength control unit (1401) sets the preservation weight of the selected image to be smaller than the second prompt reflection weight among the prompts, thereby lowering the preservation strength for the composition, background, color tone, etc. of the selected image (E) and increasing the allowable degree of deformation according to the second prompt, so that a final image (F) in which the second prompt is relatively significantly reflected in the selected image (E) is generated in the generative AI model.

[0063] On the other hand, when the deformation strength parameter is set to be less than the editing strength threshold, the editing strength control unit (1401) sets the preservation weight of the selected image to be greater than the second prompt reflection weight to increase the preservation strength regarding the composition, background, color tone, etc. of the selected image (E), and limits the allowable degree of deformation according to the second prompt, thereby ensuring that the final image (F) in which the visual characteristics of the selected image (E) are preferentially preserved is generated in the generative AI model. Here, preservation strength refers to a control strength that induces the preservation of the visual characteristics of the selected image (E), and can be implemented as at least one of the selected image preservation weight, the second prompt reflection weight, the amount of noise injected from the selected image, the step interval controlled to preserve the selected image characteristics among the number of inverse diffusion steps, or a combination thereof.

[0064] Additionally, the editing strength control unit (1401) can output a control element corresponding to the deformation strength parameter to the interface screen of the external terminal (B) and set or update the value of the deformation strength parameter based on the operation input for the control element. For example, at least one of a slider, a dial, or an increase / decrease button capable of continuously adjusting the value of the deformation strength parameter may be output as a control element on the interface screen of the external terminal (B). The external terminal (B) can transmit a control signal including a target value of the deformation strength parameter based on the operation input for the control element to a support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching, and the editing strength control unit (1401) can receive the target value included in the control signal and set or update the deformation strength parameter. In other words, the editing strength control unit (1401) can set or change the deformation strength parameter that determines the level of reflection of the second prompt on the derived selection image (E).

[0065] In this way, the generation control module (140) can control the generation of the final image (F) by lowering the degree of preservation of the selected image (E) and increasing the degree of deformation allowance according to the second prompt as the deformation strength parameter increases, and by controlling the generation of the final image (F) by increasing the degree of preservation of the selected image (E) and limiting the degree of deformation allowance according to the second prompt as the deformation strength parameter decreases. Here, the degree of preservation of the selected image (E) and the degree of reflection of the second prompt can be implemented as a preservation weight and a prompt reflection weight, respectively. For example, when the deformation strength parameter is greater than or equal to the editing strength threshold, the generation control module (140) can set the preservation weight to be smaller than the prompt reflection weight so that the degree of reflection of the second prompt becomes relatively larger, thereby ensuring that the overall characteristics of the selected image (E) are relatively less preserved. On the other hand, when the deformation strength parameter is less than the editing strength threshold, the preservation weight can be set to be larger than the prompt reflection weight so that the overall characteristics of the selected image (E) are relatively more preserved.

[0066] The support server (100) for improving the computational efficiency of a generative AI model through prompt similarity matching of the present invention can precisely control the deformation strength parameter through such a generation control module (140) to adjust the degree of feature preservation of the selected image (E), thereby enabling the generation of a final image (F) in which the change elements intended by the user are preferentially reflected.

[0067] Furthermore, the support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention can reduce user dropout due to waiting time by deriving and presenting candidate images through low-cost similarity search without generating images through high-cost GPU computation during the step of presenting candidate images, thereby presenting candidate images quickly without self-generated GPU computation. For example, the support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching according to the present invention can reduce server infrastructure costs compared to existing methods and can reduce user dropout occurring during the generation waiting time through immediate visual feedback.

[0068] The storage update module (150) collects generation history information including the final image (F) and the prompt and deformation strength parameters used to generate the final image. At this time, the storage update module (150) can structure the collected generation history information into a generation history record. Alternatively, it can update the existing generation history record by mapping additional collected generation history information to the structured generation history record and store it in the mapping storage module (110). This storage update module (150) can update and store the generation history record of the base image stored in the mapping storage module (110) into a generation history record based on the final image (S160). At the same time, it expands the candidate group of search target images stored in the mapping storage module (110). That is, the storage update module (150) establishes a resource recycling structure that automatically registers the generated final image to the mapping storage module, thereby expanding the candidate images of the mapping storage module. More specifically, the storage update module (150) receives the final image (F) generated by the generative AI model from the generation control module (140). Additionally, the storage update module (150) receives the prompt used to generate the final image (F) from the prompt module (120) and collects generation history information including deformation strength parameters from the generation control module (140). At this time, the mapping storage module (110) configures and stores the final image (F) and the generation history information as a single generation history record so that they are mutually linked, thereby expanding the search target image candidate group stored in the mapping storage module (110).

[0069] Additionally, the storage update module (150) generates or acquires an embedding vector to be linked to a creation history record corresponding to the final image (F) and provides it to the mapping storage module (110), and the mapping storage module (110) stores the embedding vector linked to the creation history record. Here, the embedding vector may be an embedding vector for a first prompt or a second prompt, and the linking of the embedding vector may be implemented by directly storing it in a vector field within the creation history record, by storing the embedding vector in a separate vector storage table and linking the Image ID as a common key, or by a combination thereof. For example, the storage update module (150) may assign image identification information, e.g., Image ID=IMG_2026_0101, to the final image (F) and provide a set of creation history information including the Image ID to the mapping storage module (110). At this time, the mapping storage module (110) can configure and store a creation history record R_2026_0101 using an Image ID as a key. Here, the creation history record R_2026_0101 may include at least one of image identification information, a final image data file path, a first prompt, a second prompt, a deformation strength parameter, and a creation parameter.

[0070] As the candidate image search target of the mapping storage module (110) is expanded by the storage update module (150) in this manner, the range in which the candidate presentation module (130) can search for candidate images is expanded. And, it becomes possible to present candidate images with improved suitability.

[0071] In addition, according to another embodiment of the present invention, a prompt module (120) of a support server that improves the computational efficiency of a generative AI model through prompt similarity matching of the present invention provides a screen as shown in FIG. 9 to an interface screen of an external terminal (B) and can receive a first prompt from the external terminal (B) of “summer sale 50% discount banner, bright and cool-feeling photo”.

[0072] since, The prompt module (120) can provide a plurality of selection elements (1201) as illustrated in FIG. 10 to an external terminal (B). At this time, the selection elements (1201) may be implemented in the form of buttons mapped to at least one of a predefined phrase, tag, or generation parameter, and may be divided into a first prompt selection element and a second prompt selection element. At this time, when the second prompt selection element is selected through the prompt module (120), the prompt mapped to the second prompt selection element may be incrementally reflected in the selection image (E).

[0073] The prompt module (120) can generate a second prompt by combining user selection signals for multiple selection elements (1201), but can perform intelligent combination logic rather than simple listing. Specifically, the prompt module (120) can determine the emphasis ranking within the prompt by tracking the selection sequence of multiple buttons and placing the prompt token corresponding to the first selected button at the beginning of the generated phrase or increasing its importance. Additionally, the prompt module (120) can determine the weight of the prompt by counting the number of times a user selects a specific button. For example, when the 'bright' button is clicked once, the corresponding token is reflected as a default value, and when it is clicked repeatedly two or more times, a weighted emphasis phrase is applied to the corresponding token, so that the generation control module (140) can express the corresponding feature relatively more strongly in the generative AI model.

[0074] Meanwhile, as illustrated in FIG. 11, the candidate group presentation module (130) provides the basic images (C) of the candidate group retrieved from the mapping storage module (110) to an external terminal (B) and allows the user to select at least one of the candidate images. At this time, the candidate group presentation module (130) presents the top 15 basic images with the highest similarity scores among the basic images that have a reference similarity level or higher as a candidate group, so that at least one selected image (E) among the basic images (C) of the candidate group can be determined according to the selection signal received from the external terminal (B).

[0075] Additionally, the editing strength control unit (1401) of the present invention can output a control element corresponding to a deformation strength parameter to the interface screen of an external terminal. A control signal generated as a user manipulates the control element is received by a support server (100) that improves the computational efficiency of a generative AI model through prompt similarity matching in real time, and accordingly, the value of the deformation strength parameter can be set or changed.

[0076] When the selected image is confirmed, the prompt module (120) provides an interface to an external terminal (B) that allows selecting the ratio of the image as shown in FIG. 12 and inputting additional text prompts such as brand name, product name, and main copy, and can receive a second prompt.

[0077] The generation control module (140) enables the generation of a final image in which the degree of feature preservation of the selected image and the degree of reflection of the second prompt are precisely reflected based on the prompt and deformation strength parameters intuitively received through the UI in this way, so that the final image is generated in the generative AI model.

[0078] Through this, the support server (100) that improves the computational efficiency of the generative AI model through prompt similarity matching of the present invention can provide an environment in which even a user without professional knowledge of writing prompts can precisely and intuitively adjust the details of the generated result by only using a combination of buttons and slider operation.

[0079] As described above, specific embodiments of the invention regarding a support server and a method for improving the computational efficiency of a generative AI model through prompt similarity matching according to one embodiment of the present invention have been described in detail in this specification. However, a person skilled in the art who understands the spirit of the present invention can easily propose other inventions or other embodiments included within the scope of the spirit of the present invention by adding, changing, deleting, etc., other components within the scope of the same spirit. Explanation of the symbols

[0080] 100: Support server improving the computational efficiency of generative AI models through prompt similarity matching 110: Mapping Storage Module 120: Prompt Module 130: Candidate Presentation Module 1301: Embedding Vector Generation Unit 1302: Candidate Presentation Unit 140: Generation Control Module 1401: Editing Intensity Control Unit 150: Save / Update Module A: Communication network B: External terminal C: Base Image D: Candidate Image E: Selected Image F: Final Image G: Generative AI model

Claims

Claim 1 A support server for improving the computational efficiency of a generative AI model through prompt similarity matching, comprising: a mapping storage module (110) that stores a base image and a generation condition prompt of the base image by mutually mapping them; a prompt module (120) that receives a first prompt used to calculate similarity with the generation condition prompt from an external terminal (B), and receives a second prompt used to edit the selected image after receiving a selected image selected from the external terminal among the candidate images; a candidate group presentation module (130) that calculates the similarity between the first prompt and the generation condition prompt and presents a base image that is greater than or equal to a reference similarity to the external terminal as a candidate image; and a generation control module (140) that receives the selected image selected from the external terminal among the candidate images, transmits the selected image and the second prompt to at least one generative AI model to request the generation of a final image in which the second prompt is reflected in the selected image, and then receives the generated final image and transmits it to the external terminal. Claim 2 A support server for improving the computational efficiency of a generative AI model through prompt similarity matching, further comprising a storage update module (150) that collects generation history information including the final image and the prompt used to generate the final image, structures the collected generation history information into a generation history record, or updates the generation history record by mapping additional generation history information to the structured generation history record and stores it in a mapping storage module (110). Claim 3 In claim 1, the candidate group presentation module (130) comprises an embedding vector generation unit (1301) that generates a prompt embedding vector for the prompt and generates a generation condition prompt embedding vector for the generation condition prompt stored in the mapping storage module (110), or loads a previously stored generation condition prompt embedding vector, and a candidate group presentation unit (1302) that presents a base image mapped to one or more generation condition prompts having a similarity score greater than or equal to a reference similarity as a candidate group, thereby supporting a generative AI model that improves computational efficiency through prompt similarity matching. Claim 4 In paragraph 3, the prompt module (120) is a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, comprising a first prompt used to calculate similarity with the generation condition prompt and a second prompt used to edit the selected image, and receiving an additional reference image used for editing from the external terminal. Claim 5 In paragraph 4, the generation control module (140) is a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, which sets or changes a deformation strength parameter that determines the level of reflection of the second prompt for the selected image. Claim 6 In claim 5, the generation control module (140) is a support server that improves the computational efficiency of a generative AI model through prompt similarity matching, wherein as the deformation strength parameter increases, the preservation strength of the selected image is lowered and the allowable degree of deformation according to the second prompt is increased, and as the deformation strength parameter decreases, the preservation strength of the selected image is increased and the allowable degree of deformation according to the second prompt is lowered, and the final image is generated in the generative AI model. Claim 7 In claim 6, the generation control module (140) compares the determined deformation strength parameter with a preset editing strength threshold value, and if the deformation strength parameter is greater than or equal to the preset editing strength threshold value, sets the preservation weight of the selected image to be smaller than the second prompt reflection weight so that a final image with the second prompt reflected in the selected image is generated, and if the deformation strength parameter is less than the preset editing strength threshold value, sets the preservation weight of the selected image to be larger than the second prompt reflection weight so that a final image in which the selected image is maintained is generated, thereby improving the computational efficiency of a generative AI model through prompt similarity matching. Claim 8 A method for improving the computational efficiency of a generative AI model through prompt similarity matching, comprising: (a) a step of mutually mapping and storing a base image and a generation condition prompt of the base image; (b) a step of receiving a first prompt used to calculate similarity with the generation condition prompt from an external terminal (B) connected via a communication network; (c) a step of calculating similarity between the first prompt and the generation condition prompt, and presenting the base image, in which the similarity is greater than or equal to a preset reference similarity, as a candidate image to the external terminal (B); (d) a step of receiving a selected image selected from the candidate images by the external terminal (B); and (e) a step of receiving a second prompt used to edit the selected image from the external terminal. and (f) a step of transmitting the selected image and the second prompt to at least one generative AI model to request the generation of a final image in which the second prompt is reflected in the selected image, and then receiving the generated final image and transmitting it to the external terminal; a method for improving the computational efficiency of a generative AI model through prompt similarity matching.

Images