Information processing system, control method for information processing system, and program

Abstract

This technology provides a way for users to easily select a pre-trained model that produces the desired results. [Solution] The information processing system includes a storage means that stores evaluation information by associating a trained model with information indicating the degree to which the generation results generated by the trained model change relative to the original data; a request acquisition means that acquires request information representing a request regarding the degree of change; and a selection means that selects at least one trained model from among a plurality of trained models held in the storage means based on the request information and the evaluation information.

Description

Technical Field

[0001] The present invention relates to an information processing system, a control method for an information processing system, and a program.

Background Art

[0002] In recent years, with the spread of pre-trained models trained for data generation, called generative AI, an environment has been gradually established in which various types of data (text, images, videos, audio, 3D models, etc.) can be easily generated in large quantities. For example, it includes the generation of news content in news programs and AI news anchors in the mass media industry, and the generation of AI talents and AI actors in the entertainment industry.

[0003] On the other hand, the data generated by the inference of generative AI may sometimes contain information different from facts. As the accuracy of generative AI improves, it becomes very difficult for those other than the data creator to distinguish between factual information and non-factual information contained in the generated data. Problems such as data creators intentionally mixing information different from facts into the generated data for the purpose of information manipulation or fraud are also increasing with the spread of generative AI.

[0004] Patent Document 1 discloses a technique for selecting a pre-trained model that satisfies performance and computational resources from a plurality of pre-trained models. Specifically, the performance of each of the plurality of pre-trained models is calculated using test data with correct information attached, and after screening the pre-trained models based on the performance, an optimal pre-trained model is selected based on the resource information of the user-side device. As a result, it becomes possible to select a pre-trained model that also takes into account the execution environment in which it will be executed from among a plurality of pre-trained models.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

[0006] However, while the technology described in Patent Document 1 allows for the selection of a pre-trained model considering performance and computational resources, it is unclear to what extent the generated results of the selected pre-trained model are based on factual information. Therefore, there is a challenge in that it is difficult for users to select a pre-trained model that will produce the generated results they desire.

[0007] This invention has been made in view of the above-mentioned problems, and aims to provide a technology for easily selecting a trained model that can produce the desired generation results for the user. [Means for solving the problem]

[0008] The information processing system according to the present invention, which achieves the above objective, A storage means that stores evaluation information by associating a trained model with information indicating the degree to which the generated results produced by the trained model change relative to the original data, A request acquisition means for acquiring request information representing the request regarding the degree of change, A selection means that selects at least one trained model from among a plurality of trained models held in the holding means based on the request information and the evaluation information, It is characterized by being equipped with [the following features]. [Effects of the Invention]

[0009] According to the present invention, users can easily select a pre-trained model that produces the desired generation results. [Brief explanation of the drawing]

[0010] [Figure 1] An explanatory diagram illustrating the scenario in which the information processing device according to the first and third embodiments selects a trained model. [Figure 2]A diagram showing an example of a plurality of learned models recorded in the holding unit of the information processing apparatus according to the first and third embodiments. [Figure 3] A flowchart showing the flow of processing related to the authenticity evaluation information recorded in the recording unit according to an embodiment. [Figure 4] An explanatory diagram regarding the calculation of authenticity according to an embodiment. [Figure 5] An explanatory diagram regarding a learned model that performs Text-to-Image. [Figure 6] A block diagram showing the hardware configuration of the information processing system according to an embodiment. [Figure 7] A block diagram showing the module configuration of the information processing system according to the first embodiment. [Figure 8] A flowchart showing the flow of processing executed by the information processing apparatus according to the first and second embodiments. [Figure 9] A diagram showing an example of the authenticity evaluation information acquired by the information processing apparatus according to an embodiment. [Figure 10] A flowchart showing the flow of the authenticity calculation process according to the first embodiment. [Figure 11] An explanatory diagram of the authenticity calculation process according to the first embodiment. [Figure 12] An explanatory diagram of the usage form of the information processing apparatus according to the second embodiment. [Figure 13] A block diagram showing the module configuration of the information processing system according to the second embodiment. [Figure 14] A diagram showing an example of the authenticity evaluation information acquired by the information processing apparatus according to the second embodiment. [Figure 15] A block diagram showing the module configuration of the information processing system according to the third embodiment. [Figure 16] A flowchart showing the flow of processing executed by the information processing apparatus according to the third embodiment. [Figure 17] A diagram showing an example of the display screen of the display UI system according to an embodiment.

Embodiments for Carrying Out the Invention

[0011] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the invention according to the claims. Although a plurality of features are described in the embodiments, not all of these plurality of features are essential for the invention, and the plurality of features may be arbitrarily combined. Further, in the accompanying drawings, the same or similar configurations are denoted by the same reference numerals, and redundant descriptions are omitted.

[0012] (First Embodiment) First, an overview of the environment in which the information processing apparatus according to the first embodiment is used will be described. Here, the "degree representing how much the generation result by the generative AI is based on facts" or the "degree of change in the generation result by the generative AI with respect to the original data (information to be changed)" is defined as the "truth degree". Alternatively, information indicating the degree to which the generation result by the generative AI has not changed with respect to the original data may be used. Alternatively, an index representing the degree of addition to the original data, the amount of modification of the original data, the degree of leaving the original data, the degree of deformation of the original data, etc. may be used.

[0013] In the present embodiment, a situation where a learned model is selected based on the truth degree requested by a user who generates an image will be described as an example.

[0014] <Usage form> Figure 1 is an explanatory diagram illustrating how the information processing device 103 according to this embodiment selects a trained model recorded in the storage unit 104. User 102 inputs the desired truth value as truth value request information 105. The information processing device 103 obtains truth value evaluation information 109 from the storage unit 104 in order to select a trained model suitable for the input truth value request information 105. In this embodiment, the truth value evaluation information is a management table that records the truth value for each model in association with models A106, B107, and C108 recorded in the storage unit 104. Based on the acquired truth value evaluation information 109 and the input truth value request information 105, the information processing device 103 selects a selected model 110 from the storage unit 104. In the example in Figure 1, the information processing device 103 and the storage unit 104 are depicted as separate components, but the information processing device 103 may also be configured to include the storage unit 104.

[0015] <Pre-trained model> Figures 2(a) and 2(c) illustrate an example of the training of multiple different trained models recorded in the storage unit 104. Referring to Figure 2(a), model A106 will be described. This model A106 is a style transfer model, and specifically, it outputs a style transfer image 113 that resembles the person in the input image, which is the style transfer image 112, from the person in the input image, which is the reference image 111.

[0016] Refer to Figure 2(b) to explain Model B107. Model B107 is a model that removes local regions. Specifically, it outputs a region-removed image 122 based on the input image, which is the mask image 121, by removing local regions from the input image, which is the reference image 120. In the example in Figure 2(b), the region on the left has been removed.

[0017] Refer to Figure 2(c) to explain Model C108. Model C108 is a noise removal model that specifically detects and removes noise components contained in an input image, which is a noisy image 130, and outputs a noiseless image 131.

[0018] The multiple different pre-trained models in this embodiment may be CNN-based neural network models (hereinafter referred to as NN models) or GANs with Generator / Discriminator. CNN stands for Convolutional Neural Network, and GAN stands for Generative Adversarial Network.

[0019] <Recording and processing of truthfulness evaluation information> Figure 3 is a flowchart showing the procedure for recording truthfulness evaluation information 109 stored in the holding unit 104 according to this embodiment. The series of processes are performed by the information processing device 103.

[0020] In step S140, the information processing device 103 selects one trained model from among multiple trained models recorded in the storage unit 104. In step S141, the information processing device 103 prepares test data in advance and performs the generation process by inputting the prepared test data into the trained model selected in step S140. The test data may be data previously recorded in the storage unit 104, or data arbitrarily prepared by the user who has recorded the trained model in the storage unit 104. Correct answer information may also be attached to the test data. The generation process of the trained model in this embodiment is intended to generate an image. However, the target of generation is not limited to images, and may also be audio, text, 3D models, etc.

[0021] In step S142, the information processing device 103 calculates the truth score based on the results generated in step S141. The specific details of the method for calculating the truth score will be described later. In step S143, the information processing device 103 associates the trained model selected in step 140 with the truth score calculated in step S142 and records it as truth score evaluation information 109. The recording method in this embodiment is a table associating the trained model and the truth score. However, it is not limited to this, and it is sufficient to record the trained model and the truth score in a way that allows them to be mutually indexed. For example, it may be a list format in which the trained model and the truth score are managed with the same index.

[0022] In step S144, the information processing device 103 determines whether the truth value calculation has been completed for all trained models recorded in the storage unit 104. If it has been completed, the series of processes ends. If it has not been completed, the process returns to step S140.

[0023] <Calculation of Truthfulness> Next, with reference to Figures 4(a)-4(c), the truthfulness calculation process according to this embodiment will be explained. Figure 4(a) shows an example of performing the truthfulness calculation process in step S142 of the processing flow in Figure 3 using similarity. If model A106 is selected in step S140, test data 150 is input to model A106 and a style-transformed image 151 is generated. Based on the generated style-transformed image 151 and the test data 150, the image similarity is calculated. In this embodiment, SSIM (Structural similarity) is used for the similarity calculation process. The average pixel values ​​of the style-transformed image 151 and the test data 150 are μ x , μ y , variance σ x , σ y , covariance σ xy Let's assume that... Furthermore, using the constants C1 and C2, SSIM can be expressed by the following equation (1).

[0024]

number

[0025] SSIM values ​​closer to 0 indicate lower similarity, while values ​​closer to 1 indicate higher similarity. Therefore, the value calculated using equation (1) may be used as the truth value. Alternatively, the calculated similarity may be converted to a percentage and used as the truth value, or the calculated similarity may be divided into arbitrary intervals, such as 0 ≤ SSIM < 0.2 for truth value rank A, and 0.2 ≤ SSIM < 0.4 for truth value rank B, and a truth value may be assigned to each divided range. The method for calculating similarity is not limited to SSIM; it may also be calculated using cosine similarity or the difference in simple pixel values. Furthermore, in this embodiment, the truth value was determined by calculating the similarity between the input test data 150 and the style-converted image 151, but the truth value may also be determined by the similarity between the ground truth information attached to the test data and the style-converted image 151.

[0026] Figure 4(b) shows an example of calculating the truthfulness of an image using the area ratio of the modified region, as performed in step S142 of the processing flow in Figure 3. If model B107 is selected in step S140, test data 160 is input to model B107 to generate a region-deleted image 161. The generated region-deleted image 161 is compared with the test data 160, and the ratio of the area of ​​the deleted region in the region-deleted image 161 to the area of ​​the entire image is calculated. This calculated area ratio is then used as the truthfulness of the image. For example, if the area of ​​the deleted region in the region-deleted image 161 is 40% of the total image area, the area of ​​the unmodified region is 60%, so the truthfulness of the image is set to 60. Alternatively, instead of extracting the deleted region, a neural network model trained for object detection may be used, and the truthfulness of the image may be calculated from the ratio of the number of objects detected in the test data 160 to the number of objects detected in the region-deleted image 161.

[0027] Figure 4(c) shows an example of performing the truthfulness calculation process in step S142 of the processing flow in Figure 3 using parameter values. If model C108 is selected in step S140, parameter values ​​172 that can be set for model C108 are obtained. The obtained parameter values ​​172 are values ​​that affect the denoised image 171, which is the result of inputting test data 170 into model C108 and generating a denoised image 171. For example, it is an adjustment parameter for the denoising intensity. The value of this parameter value 172 can be used as the truthfulness value. Alternatively, as another method of calculating truthfulness, the truthfulness may be calculated from the similarity of the denoised image 171 generated with parameter value 172, based on the truthfulness obtained from the similarity between the denoised image generated with the upper and lower limits of the settable parameters and the test data 170. The obtained truthfulness and parameter value 172 may be recorded as truthfulness evaluation information 109. The specific details will be described later.

[0028] <text-to-image> Figures 5(a) and 5(b) are explanatory diagrams of a trained model that performs text-to-image processing. Refer to Figure 5(a) to explain model 180. This model performs text-to-image processing and specifically takes Txt181, which is a string of characters such as alphanumeric characters or Japanese, as input and outputs image 182 based on the input string information. This model 180 may be a NN model that combines a Text Encoder and an Image Decoder based on a Transformer, or it may be a Diffusion-based NN model.

[0029] Figure 5(b) shows an example of the truthfulness calculation process performed in step S142 of the processing flow in Figure 3. In step S140, model 180 is selected, and in step S141, TestData184 is obtained as test data. At this time, TestData184 is prompt information such as "white dog", and the correct answer information, the correct image 183 of a "white dog", is attached. In step S142, TestData184 is input to model 180 and a dog image 185 is generated.

[0030] The similarity between the generated dog image 185 and the ground truth image 183 is calculated. The similarity may be calculated using SSIM, as described above. After calculating the similarity, the truth score is calculated based on the calculated similarity. Alternatively, as a method for calculating similarity, object detection may be performed on the generated dog image 185 and the ground truth image 183, and the similarity may be calculated based on the detection results. Object detection may be performed using a CNN-based NN model trained to classify objects in the image by type.

[0031] <Description of Hardware Configuration> Figure 6 is a block diagram showing the hardware configuration of an information processing system (or information processing device) according to the first embodiment. The CPU 301 controls various devices connected to the bus 302 and performs information processing. CPU is an abbreviation for Central Processing Unit. The ROM 303 stores the BIOS program and boot program. ROM is an abbreviation for Read Only Memory.

[0032] RAM 304 is used as the main memory of the CPU 301. RAM stands for Random Access Memory. External memory 305 stores the programs processed by the information processing device 103. The input unit 306 is a keyboard or mouse and performs processing related to information input. The display unit 307 outputs the calculation results of the information processing device 103 to the display device according to instructions from the CPU 301. The display device can be of any type, such as a liquid crystal display, projector, or LED indicator. LED stands for Light Emitter Diode.

[0033] Bus 302 is connected in a manner that enables communication between the CPU 301, RAM 304, ROM 303, and external memory 305.

[0034] I / F308 is an interface that facilitates information communication over a network, enabling communication between the information processing system 101 and external systems. The communication interface can be Ethernet, USB, serial communication, wireless communication, or any other type.

[0035] <Functional Configuration> The functional configuration of the information processing system according to this embodiment will be explained using Figure 7. Figure 7 is a block diagram showing the module configuration of the system according to the first embodiment.

[0036] The information processing system 101 comprises an information processing device 103, a storage unit 104, a truth-of-fact request information acquisition unit 202, and a model provision unit 206. The truth-of-fact request information acquisition unit 202 is an information acquisition device for inputting the truth-of-fact requested by the user 102 to the information processing device 103. The model provision unit 206 is an output device for providing a trained model capable of generating data that corresponds to the truth-of-fact requested by the user 102. The information processing device 103 comprises a truth-of-fact evaluation information acquisition unit 203, a model selection unit 204, and a model acquisition unit 205. The storage unit 104 comprises an input / output information control unit 207, a recording unit 208, and an information acquisition unit 209.

[0037] The truthfulness request information acquisition unit 202 acquires truthfulness request information representing the truthfulness request input to the information processing system 101 (request acquisition). The truthfulness evaluation information acquisition unit 203 acquires truthfulness evaluation information from the storage unit 104. The truthfulness evaluation information is a management table that associates models with truthfulness, but details will be described later. The model selection unit 204 selects a trained model based on the input truthfulness request information and the acquired truthfulness evaluation information. The model acquisition unit 205 acquires the corresponding trained model from the storage unit 104 based on the model information selected by the model selection unit 204. Subsequently, the acquired trained model is provided to the user through the model provision unit 206. Specific details will be described later.

[0038] The holding unit 104 is connected to the information processing device 103 and receives requests from the information processing device 103 as input. In response to the requests, it outputs truthfulness evaluation information or a trained model. The input / output information control unit 207 controls the input and output of information between the information processing device 103 and the holding unit 208. The recording unit 208 records multiple trained models and a table of truthfulness calculated for each model as truthfulness evaluation information. The specific details will be described later.

[0039] In the example shown in Figure 7, the information processing device 103, the data storage unit 104, the truthfulness request information acquisition unit 202, and the model providing unit 206 are depicted as separate components. However, the example is not limited to this one, and the information processing device 103 may be configured to include at least some or all of the data storage unit 104, the truthfulness request information acquisition unit 202, and the model providing unit 206.

[0040] <Processing procedure and detailed processing method> Next, the processing procedure and detailed processing method of the information processing device 103 according to this embodiment will be explained with reference to Figure 8. Figure 8 is a flowchart showing the overall flow of processing performed by the information processing device 103 shown in Figure 7.

[0041] In step S401, the truthfulness request information acquisition unit 202 acquires truthfulness request information 105, which is a request from user 102 that has been input to the information processing system 101 via the input unit 306. In this embodiment, the truthfulness request information 105 is the truthfulness input by user 102, but is not limited to this. For example, it may be acquired from truthfulness input by user 102 in the past. Also, the requested truthfulness may be a specific numerical value, or a numerical value that can be specified within a range including upper and lower limits.

[0042] In step S402, the truthfulness evaluation information acquisition unit 203 acquires truthfulness evaluation information 109 from the storage unit 104. The storage unit 104 receives a request from the truthfulness evaluation information acquisition unit 203 to output truthfulness evaluation information 109 via the input / output information control unit 207. Based on the request received by the input / output information control unit 207, the information acquisition unit 209 acquires truthfulness evaluation information 109 from the recording unit 208 and transmits it to the input / output information control unit 207. The input / output information control unit 207 transmits the received truthfulness evaluation information 109 to the truthfulness evaluation information acquisition unit 203 in the information processing device 103. In this embodiment, the truthfulness evaluation information 109 is stored in the storage unit 104 in the form of a management table that associates multiple different trained models recorded in the storage unit 104 with the truthfulness of each model.

[0043] In step S403, the model selection unit 204 selects a trained model that satisfies the user's request based on the truthfulness request information 105 and the truthfulness evaluation information 109. If there are multiple trained models in the truthfulness evaluation information 109 that have the same truthfulness as the truthfulness request information 105, all trained models with the same truthfulness are selected. If there are no trained models in the truthfulness evaluation information 109 that have the same truthfulness as the truthfulness request information 105, a trained model recorded in the management table of the truthfulness evaluation information 109 with a truthfulness value higher than the truthfulness request information 105 is selected. In this case, the trained model recorded in the management table of the truthfulness evaluation information 109 with the closest and highest truthfulness value may be selected. Alternatively, if no trained model with the same truth value as the truth value included in the truth value request information 105 exists in the truth value evaluation information 109, a trained model recorded in the management table of the truth value evaluation information 109 may be selected as having the truth value closest to the truth value included in the truth value request information 105, or a truth value where the difference is less than or equal to a predetermined value.

[0044] Alternatively, by obtaining the truthfulness and threshold in step S401 as truthfulness request information 105, a trained model that matches the truthfulness and threshold range requested by user 102 may be selected from truthfulness evaluation information 109.

[0045] In step S404, the model acquisition unit 205 acquires the trained model selected in step S403 from the storage unit 104. The storage unit 104 receives an input / output information control unit 207 from the model acquisition unit 205 to acquire the trained model. Based on the request received by the input / output information control unit 207, the information acquisition unit 209 acquires the requested trained model from the recording unit 208 and transmits it to the input / output information control unit 207. The input / output information control unit 207 transmits the received trained model to the model acquisition unit 205 in the information processing device 103.

[0046] In step S405, the model provision unit 206 provides the trained model acquired in step S404 to the user 102. When providing the trained model, it may be provided to the user 102 in an order of priority based on truth value. For example, trained models with truth values ​​close to the truth value requested by the user 102 may be provided to the user 102 preferentially, or they may be sorted in ascending or descending order of truth value and provided to the user 102.

[0047] <Effects> As described above, according to this embodiment, the user will be able to select an appropriate pre-trained model based on its truthfulness from among several different pre-trained models.

[0048] <Example 1> In the first embodiment, an example was described in which a trained model is selected from a management table that associates multiple trained models with the truth value for each model, which is recorded as truth value evaluation information. However, a trained model may have multiple parameters that affect the data it generates, and the truth value may change as the generated data changes depending on the parameter values. Therefore, in addition to multiple trained models and the truth value for each model, a management table may be prepared that associates the parameters to be changed with their parameter values, and the model may be selected from this management table. The specific details will be explained using Figures 8 and 9. Note that steps S401, S403, and S404 in Figure 8 are the same as the contents of the processes described above, so their explanation will be omitted.

[0049] Next, Figure 9 shows an example of truthfulness evaluation information related to this modified example. In step S402 of Figure 8, the truthfulness evaluation information acquisition unit 203 acquires truthfulness evaluation information 501. This truthfulness evaluation information 501 is a management table that associates the truthfulness of each model with the modifiable parameters and their values ​​for each model for multiple trained models recorded in the storage unit 104. The modifiable parameters may include adjustment parameters for noise reduction intensity or the CFG (Classifier-Free Guidance) scale.

[0050] For example, a value of 0.1 for the noise reduction intensity adjustment parameter could correspond to a truth score of 95, and a value of 0.2 for a truth score of 91. Furthermore, the CFG scale is a numerical value that specifies how faithfully the generated image is to the prompt. You can control this by setting a large CFG scale value to strongly reflect the prompt, or a small CFG scale value to prioritize image quality. For example, a CFG scale=1 could correspond to a truth score of 95, and a CFG scale=2 could correspond to a truth score of 80.

[0051] In step S405, the model provision unit 206 provides the user 102 with the trained model acquired in step S404 and the model-related parameters to be changed and their values, based on the truthfulness evaluation information 501 acquired in step S402.

[0052] This allows user 102 to select a pre-trained model from among several different pre-trained models that can generate data with the desired truthfulness level, based on the parameters to be changed and their values. Therefore, after selecting a pre-trained model, user 102 can generate data with the desired truthfulness level without having to search for parameters.

[0053] <Modification 2> Furthermore, in the first embodiment, it was assumed that a management table associating multiple different trained models with the truth value for each model was pre-recorded in the storage unit 104 as truth value evaluation information 109, and the trained model was selected based on this assumption. However, the truth value recorded in advance is a value calculated using test data, and not a truth value calculated based on the target information (original data) that the user 102 actually wants to change. Therefore, even after selecting a trained model and actually inputting the target information to be changed to obtain the generated result, it is possible that the truth value may not be what the user 102 requested. Thus, in this modified example, a method for calculating the truth value based on the target information that the user 102 actually wants to change and selecting a trained model will be described. The specific details will be explained using Figures 8, 10, and 11. Note that steps S402 to S405 in Figure 8 are the same as the contents of each process described above, so the explanation will be omitted.

[0054] In step S401 of Figure 8, the truthfulness request information acquisition unit 202 acquires the truthfulness requested by the user 102 as truthfulness request information 105, and the information to be changed 601 as shown in Figure 11. In this embodiment, the information to be changed 601 is an image, but is not limited to this. For example, it may be audio data or text data.

[0055] Next, Figure 10 is a flowchart showing the procedure for calculating truthfulness evaluation information according to this modified example. This flowchart shows the flow of truthfulness calculation for truthfulness evaluation information 602 acquired by the information acquisition unit 209 of the holding unit 104. Steps S140 and steps S142 to S144 are the same as the processes described with reference to Figure 3, so their explanation is omitted.

[0056] In step S601 of Figure 10, the information acquisition unit 209 acquires the change target information 601 input to the input / output information control unit 207 and multiple trained models, model A106, model B107, and model C108, which are recorded in the recording unit 208.

[0057] In step S141, the information processing device 103 generates an image using the trained model selected in step S140. In this embodiment, the information to be changed 601 acquired in step S601 is input to the trained model to generate the image. The specific details will be explained with reference to Figure 11.

[0058] Figure 11 is an explanatory diagram illustrating an example of the processing from step S140 to step S143. The information to be changed 601 acquired in step S601 is input to models A106, B107, and C108, respectively, to generate a style conversion image 603, a region deletion image 604, and a noise reduction image 605. The truth scores A606, B607, and C608 are calculated based on the generated images and the information to be changed 601, and recorded in the management table as truth score evaluation information 602. Similarly, the generated style conversion image 603, region deletion image 604, and noise reduction image 605 are also recorded in the management table in association with the generated models.

[0059] This makes it possible to calculate truthfulness based on the information that user 102 actually wants to change, rather than test data, allowing for the selection of a trained model that better satisfies user 102's requirements.

[0060] (Second embodiment) In the first embodiment, a method for selecting a trained model was described in an environment where the information processing device and the storage unit are provided within the same information processing system, and multiple trained models that are candidates for selection are also recorded in the information processing system. In contrast, this embodiment describes a method for selecting a trained model in an environment where multiple trained models that are candidates are recorded in a second information processing device separate from the first information processing device in the information processing system. Details of the second information processing device and the selection method will be described later. Note that the hardware configuration in the second embodiment is the same as in Figure 6, so the description will be omitted.

[0061] <Usage form> First, the usage configuration according to the second embodiment will be explained with reference to Figure 12. The information processing system 101 according to this embodiment includes an information processing device 103, which communicates with N holding devices as a second information processing device. For the purposes of this explanation, these will be referred to as holding device A701, holding device B702, and holding device N703. The information processing system 101, the first information processing device 103, the truthfulness request information 105, and the selection model 110 have been explained with reference to Figure 1, so a detailed explanation will be omitted.

[0062] Each of the holding devices A701, B702, and N703 has a management table that records multiple different trained models and the truth value for each model in association with them. Based on the input truth value request information 105, a selected model 110 is selected from at least one of the holding devices A701, B702, and N703.

[0063] <Functional Configuration> Referring to Figure 13, the configuration of the information processing system and holding device according to the second embodiment will be described. Figure 13 is a block diagram of the module configuration of the system according to the second embodiment, in the case where there are N holding devices. N holding devices, here, holding device A701, holding device B702, and holding device N703 are connected via the input / output information control unit 1301 of the information processing system 101. Holding device A701 is the same as the configuration of the holding unit 104 described with reference to Figure 7, so its description is omitted. Also, the input / output information control unit 1303, input / output information control unit 1306, recording unit 1304, recording unit 1307, information acquisition unit 1305, and information acquisition unit 1308 in holding device B702 and holding device N703 are the same as those in holding device A701, so their description is omitted. The specific processing details will be described later.

[0064] In the example shown in Figure 13, the information processing device 103, the truthfulness request information acquisition unit 202, the model provisioning unit 206, and the input / output information control unit 1301 are depicted as separate components. However, the example is not limited to this one, and the information processing device 103 may be configured to include at least some or all of the truthfulness request information acquisition unit 202, the model provisioning unit 206, and the input / output information control unit 1301.

[0065] <Processing procedure and detailed processing method> Next, the processing procedure and detailed processing method of the information processing device 103 according to this embodiment will be explained using Figures 8 and 13. Note that the processing steps S401 and S403 in Figure 8 are the same as those described above, so their explanation will be omitted.

[0066] In step S402 of Figure 8, the truthfulness evaluation information acquisition unit 203 requests the input / output information control unit 1301 in the information processing system 101 to acquire truthfulness evaluation information from an information processing device other than the information processing device 103. For example, based on the request, the input / output information control unit 1301 acquires truthfulness evaluation information from the holding device A701 and the holding device B702, respectively, and transmits it to the truthfulness evaluation information acquisition unit 203. During acquisition, communication with the holding device A701 and the holding device B702 may be performed from either device, or communication may be performed in parallel.

[0067] Similarly, the holding device A701 and the holding device B702 also include an input / output information control unit 207 and an input / output information control unit 1303, respectively. In the holding device A701, based on a received request, the information acquisition unit 209 acquires truthfulness evaluation information from the recording unit 208 and transmits it to the input / output information control unit 207. Similarly, in the holding device B702, based on a received request, the information acquisition unit 1305 acquires truthfulness evaluation information from the recording unit 1304 and transmits it to the input / output information control unit 1303.

[0068] The truthfulness evaluation information recorded in the holding device A701 will be referred to as truthfulness evaluation information A901, and the truthfulness evaluation information recorded in the holding device B702 will be referred to as truthfulness evaluation information B902. Details will be explained in section 149. The input / output information control unit 207 transmits truthfulness evaluation information A901 to the truthfulness evaluation information acquisition unit 203 in the information processing device 103. Similarly, the input / output information control unit 1303 transmits truthfulness evaluation information B902 to the truthfulness evaluation information acquisition unit 203 in the information processing device 103.

[0069] Here, Figures 14(a) and 14(b) show an example of truthfulness evaluation information acquired by the truthfulness evaluation information acquisition unit 203 from multiple holding devices in step S402. The truthfulness evaluation information acquired from holding device A701 is denoted as truthfulness evaluation information A901, and the truthfulness evaluation information acquired from holding device B702 is denoted as truthfulness evaluation information B902. Truthfulness evaluation information A901 is a management table that associates multiple different trained models with the truthfulness of each model. Truthfulness evaluation information B902 is a management table that further associates the changeable parameters and values ​​for each model.

[0070] Furthermore, when truthfulness evaluation information A901 and truthfulness evaluation information B902 are acquired by the information acquisition unit 209 and the information acquisition unit 1305, respectively, the specific information of the holding device may be associated with the truthfulness evaluation information and recorded in the management table. The specific information of the holding device is, for example, unique information such as the device name, the name of the company providing the device, and the IP address. In this embodiment, the device name is recorded.

[0071] In step S404, the model acquisition unit 205 requests the input / output information control unit 1301 to acquire the trained model selected in step S403 from the holding device. Based on the request, the input / output information control unit 1301 acquires the trained model from at least one of the holding devices A701, B702, and N703 and transmits it to the model acquisition unit 205. During acquisition, communication with holding device A701 and holding device B702 may be performed in either order, and communication may be prioritized according to the file size of the trained model. Communication may also be performed in parallel. Holding device A701 and holding device B702 receive information about the trained model requested by the model acquisition unit 205 of the information processing device 103 via the input / output information control unit 207 and the input / output information control unit 1303, respectively.

[0072] In the holding device A701, based on a request received by the input / output information control unit 207, the information acquisition unit 209 acquires a trained model from the recording unit 208 and transmits it to the input / output information control unit 207. The input / output information control unit 207 transmits the received trained model to the model acquisition unit 205 in the information processing device 103. Similarly, in the holding device B701, based on a request received by the input / output information control unit 1303, the information acquisition unit 1305 acquires a trained model from the recording unit 1304 and transmits it to the input / output information control unit 1303. The input / output information control unit 1303 transmits the received trained model to the model acquisition unit 205 in the information processing device 103.

[0073] In step S405, the model provision unit 206 provides the trained model acquired in step S404 to the user 102. When providing the trained model, the unit may determine a priority based on the system information (specific information that identifies a system or device) included in the truthfulness evaluation information, and then sort the trained models according to the priority before providing them to the user 102. For example, the models may be sorted in ascending or descending order by system name (or device name) and provided to the user 102, or system names that the user 102 has selected in the past may be prioritized and displayed.

[0074] <Effects> As explained above, according to this embodiment, since it is not necessary to store multiple trained models in the information processing device, it becomes possible to configure the system with less memory.

[0075] (Third embodiment) In the first and second embodiments, the selection method was described when the output destination of the selected trained model is the model provision unit 206 of the information processing system 101. In this embodiment, the selection method of the trained model is described when the output destination is the display unit of the information processing system 101. Specifically, a display UI system is provided that displays truth-of-fact request information regarding the truth-of-fact degree requested by the user 102, truth-of-fact evaluation information obtained from the storage unit 104, and the trained model selected by the model selection unit 204 to the user 102. A detailed explanation of the displayed content will be given later.

[0076] The usage configuration in the third embodiment is the same as in Figure 1, so the explanation is omitted. Also, the hardware configuration in the third embodiment is the same as in Figure 6, so the explanation is omitted.

[0077] <Functional Configuration> The functional configuration of the information processing device in the third embodiment will be described with reference to Figure 15. Figure 15 is a block diagram showing the module configuration of the information processing system according to the third embodiment.

[0078] In addition to the components described with reference to Figure 7, the information processing device 103 includes a model information acquisition unit 1501. The model information acquisition unit 1501 acquires information related to the trained model, as well as the trained model selected by the model selection unit 204, from the storage unit 104. Details of the acquired information will be described later.

[0079] The display unit 307 provides the user 102 with the trained model acquired by the model acquisition unit 205, and information related to the trained model acquired by the model information acquisition unit 1501, through a display UI system. The specific display method and content will be described later.

[0080] In the example shown in Figure 15, the information processing device 103, the storage unit 104, the truthfulness request information acquisition unit 202, and the display unit 307 are depicted as separate components. However, the example is not limited to this one, and the information processing device 103 may be configured to include at least some or all of the storage unit 104, the truthfulness request information acquisition unit 202, and the display unit 307.

[0081] <Processing procedure and detailed processing method> The processing procedure and detailed processing method of the information processing device 103 according to this embodiment will be explained with reference to Figures 16 and 17. Figure 16 is a flowchart showing the overall processing flow in this embodiment. Each process from step S401 to step S404 is the same as in Figure 8, so its explanation will be omitted.

[0082] In step S1101, the model information acquisition unit 1501 acquires information related to the trained model selected in step S403 from the storage unit 104. The information related to the model may be, for example, time information such as the date and time or update date when the corresponding trained model was recorded in the recording unit 208 within the storage unit 104, or it may be information about the person who recorded it.

[0083] In step S1102, the display unit 307 notifies the display UI system of the trained model acquired by the model information acquisition unit 1501 in step S1101, along with information related to the model. The specific contents of the notification will be explained with reference to Figure 17.

[0084] Figure 17 shows an example of a display screen for a display UI system provided to the user 102 via the display unit 307. The display UI system provides information to the user by displaying at least one of the following: truthfulness request information acquired by the truthfulness request information acquisition unit 202, truthfulness evaluation information acquired by the truthfulness evaluation information acquisition unit 203, and a trained model selected by the model selection unit 204. Furthermore, it may also display information related to the trained model acquired by the model information acquisition unit 1501.

[0085] In Figure 17, the selection information display screen 1201 displays at least the information of the trained model selected by the model selection unit 204 in step S403. The truth value for each model may be displayed from the truth value evaluation information 109 acquired in step S402, or checkboxes that notify the user's selection status for each model may be displayed to visualize the user's operation 102. In addition, date and time information and recorder information related to the model acquired by the model information acquisition unit 1501 in step S1101 may be displayed. Furthermore, a toggle button may be displayed that allows the user 102 to rearrange the displayed information in any order of priority.

[0086] The truth-related information display screen 1202 displays the information to be changed (original data) included in the truth-of-fact request information 105 acquired in step S401, and the generated result images for each model included in the truth-of-fact evaluation information 109 acquired in step S402. In this embodiment, an example of displaying the information to be changed 601 and the style conversion image 603, which is the generated result information of the selected model A106, is shown. The information to be displayed may also be text or audio data, and is not limited to images.

[0087] The truthfulness request information input screen 1203 is an input form screen for truthfulness request information 105 that the user 102 inputs via the input unit 306. The input form screen allows the user to input at least the requested truthfulness, the parameters to be changed (e.g., noise reduction intensity adjustment parameters, CFG Scale, etc.), and the information to be changed via the screen. Input may be freely accepted by the user 102 in a text field, or a pre-configured list may be displayed in a dropdown format for the user 102 to select from.

[0088] The selection execution instruction screen 1204 includes at least a selection execution button 1701 for the information processing device 103 to execute the selection of a trained model, and a cancel button 1702 for interrupting the execution process. It may also include a button for downloading the trained model selected on the selection information display screen 1201 to any storage or server specified by the user 102.

[0089] <Effects> As described above, according to this embodiment, users can select a trained model capable of generating data with the desired level of truthfulness while visually confirming it on the display UI system, thereby improving usability when selecting a model.

[0090] The disclosures herein include the following information processing systems, methods for controlling information processing systems, and programs.

[0091] (Item 1) A storage means that stores evaluation information by associating a trained model with information indicating the degree to which the generated results produced by the trained model change relative to the original data, A request acquisition means for acquiring request information representing the request regarding the degree of change, A selection means that selects at least one trained model from among a plurality of trained models held in the holding means based on the request information and the evaluation information, An information processing system characterized by comprising the following features.

[0092] (Item 2) The information processing system according to item 1, characterized in that the requested information includes information indicating the degree of change entered by the user.

[0093] (Item 3) The information processing system according to item 2, characterized in that the request information further includes raw data to be input to the trained model.

[0094] (Item 4) The information processing system according to item 3, characterized in that the source data includes images, audio, text, or 3D models.

[0095] (Item 5) The information processing system according to any one of items 2 to 4, characterized in that the request information further includes parameters to be changed to be input to the trained model.

[0096] (Item 6) The information processing system according to item 5, characterized in that the parameter to be changed includes a noise reduction intensity adjustment parameter or a CFG (Classifier-Free Guidance) scale.

[0097] (Item 7) The information processing system according to any one of items 1 to 6, characterized in that the selection means selects a trained model that matches the requested information.

[0098] (Item 8) The information processing system according to item 7, characterized in that, if there are multiple trained models that match the requested information, the selection means selects all trained models that match the requested information.

[0099] (Item 9) The information processing system according to item 7 or 8, characterized in that the selection means selects the trained model that is closest to the requested information if no trained model that matches the requested information exists.

[0100] (Item 10) The information processing system according to any one of items 7 to 9, characterized in that, if no trained model matches the requested information, the selection means selects a trained model with a lower degree of change than the degree of change included in the requested information.

[0101] (Item 11) A model acquisition means for acquiring the at least one trained model selected by the selection means from the holding means, A providing means for providing the user with the at least one trained model acquired by the model acquisition means, An information processing system according to any one of items 1 to 10, further comprising:

[0102] (Item 12) The information processing system according to item 11, characterized in that the providing means determines the priority of the at least one trained model based on the degree of change, and rearranges the at least one trained model based on the priority and provides it to the user.

[0103] (Item 13) The information processing system according to any one of items 1 to 12, characterized in that the holding means further associates at least one of the information indicating the generation result and the parameters to be changed to be input to the trained model and holds them as evaluation information.

[0104] (Item 14) Model information acquisition means for acquiring model information related to the at least one trained model selected by the selection means from the holding means, Control means for displaying the at least one trained model and the model information on a display means, An information processing system according to any one of items 1 to 13, further comprising:

[0105] (Item 15) The information processing system according to item 14, characterized in that the model information includes time information relating to the date and time when the trained model was recorded, or information of the person who recorded the trained model.

[0106] (Item 16) An information processing system including a holding device and an information processing device, The holding device is The system includes a storage means that stores, as evaluation information, a trained model and information indicating the degree to which the generated results produced by the trained model change relative to the original data, in association with each other. The aforementioned information processing device is A request acquisition means for acquiring request information representing the request regarding the degree of change, Information acquisition means for acquiring the aforementioned evaluation information, A selection means that selects at least one trained model from among a plurality of trained models held in the holding means based on the request information and the evaluation information, An information processing system characterized by comprising the following features.

[0107] (Item 17) The information processing system according to item 16, characterized in that the holding means further associates specific information that identifies the holding device with the evaluation information.

[0108] (Item 18) The aforementioned information processing system is It is equipped with multiple holding devices, The holding means of the plurality of holding devices further associates specific information that identifies the holding device and holds it as evaluation information. The selection means selects at least one trained model from among the multiple trained models held in the holding means of the multiple holding devices, based on the request information and the evaluation information. The aforementioned information processing device is The information processing system according to item 16, further comprising a providing means for determining priority based on the specified information and rearranging the at least one trained model selected by the selection means and providing it to the user based on the priority.

[0109] (Item 19) A storage step involves associating a trained model with information indicating the degree of change in the generated results produced by the trained model relative to the original data, and storing this information as evaluation information in a storage means. A request acquisition step involves acquiring request information that represents the requirements regarding the degree of change, A selection step of selecting at least one trained model from among a plurality of trained models held in the holding means, based on the request information and the evaluation information. A control method for an information processing system, characterized by having the following features.

[0110] (Item 20) A program that causes a computer to execute the control method for the information processing system described in item 19.

[0111] (Other embodiments) The present invention can also be realized by supplying a program that implements one or more of the functions of the above-described embodiments to a system or device via a network or storage medium, and by having one or more processors in the computer of that system or device read and execute the program. It can also be realized by a circuit (e.g., an ASIC) that implements one or more functions.

[0112] The invention is not limited to the embodiments described above, and various modifications and variations are possible without departing from the spirit and scope of the invention. Accordingly, claims are attached to disclose the scope of the invention. [Explanation of Symbols]

[0113] 101: Information processing system, 104: Storage unit, 202: Truthfulness request information acquisition unit, 204: Model information acquisition unit

Claims

1. A storage means that stores evaluation information by associating a trained model with information indicating the degree to which the generated results produced by the trained model change relative to the original data, A request acquisition means for acquiring request information representing the request regarding the degree of change, A selection means that selects at least one trained model from among a plurality of trained models held in the holding means based on the request information and the evaluation information, An information processing system characterized by comprising the following features.

2. The information processing system according to claim 1, characterized in that the requested information includes information indicating the degree of change entered by the user.

3. The information processing system according to claim 2, characterized in that the request information further includes raw data to be input to the trained model.

4. The information processing system according to claim 3, characterized in that the source data includes images, audio, text, or 3D models.

5. The information processing system according to claim 2, characterized in that the request information further includes parameters to be changed to be input to the trained model.

6. The information processing system according to claim 5, characterized in that the parameter to be changed includes a noise reduction intensity adjustment parameter or a CFG (Classifier-Free Guidance) scale.

7. The information processing system according to claim 1, characterized in that the selection means selects a trained model that matches the requested information.

8. The information processing system according to claim 7, characterized in that, if there are multiple trained models that match the requested information, the selection means selects all trained models that match the requested information.

9. The information processing system according to claim 7, characterized in that the selection means selects the trained model that is closest to the requested information if no trained model that matches the requested information exists.

10. The information processing system according to claim 7, characterized in that, if no trained model matches the requested information, the selection means selects a trained model with a lower degree of change than the degree of change included in the requested information.

11. A model acquisition means for acquiring the at least one trained model selected by the selection means from the holding means, A providing means for providing the user with the at least one trained model acquired by the model acquisition means, The information processing system according to claim 1, further comprising the features described above.

12. The information processing system according to claim 11, characterized in that the providing means determines the priority of the at least one trained model based on the degree of change, and rearranges the at least one trained model based on the priority and provides it to the user.

13. The information processing system according to claim 1, characterized in that the holding means further associates at least one of the information indicating the generation result and the parameters to be changed to be input to the trained model and holds them as evaluation information.

14. Model information acquisition means for acquiring model information related to the at least one trained model selected by the selection means from the holding means, Control means for displaying the at least one trained model and the model information on a display means, The information processing system according to claim 1, further comprising the features described above.

15. The information processing system according to claim 14, characterized in that the model information includes time information relating to the date and time when the trained model was recorded, or information of the person who recorded the trained model.

16. An information processing system including a holding device and an information processing device, The holding device is The system includes a storage means that stores, as evaluation information, a trained model and information indicating the degree to which the generated results produced by the trained model change relative to the original data, in association with each other. The aforementioned information processing device is A request acquisition means for acquiring request information representing the request regarding the degree of change, Information acquisition means for acquiring the aforementioned evaluation information, A selection means that selects at least one trained model from among a plurality of trained models held in the holding means based on the request information and the evaluation information, An information processing system characterized by comprising the following features.

17. The information processing system according to claim 16, characterized in that the holding means further associates specific information that identifies the holding device with the evaluation information.

18. The aforementioned information processing system is It is equipped with multiple holding devices, The holding means of the plurality of holding devices further associates specific information that identifies the holding device and holds it as evaluation information. The selection means selects at least one trained model from among the multiple trained models held in the holding means of the multiple holding devices, based on the request information and the evaluation information. The aforementioned information processing device is The information processing system according to claim 16, further comprising a providing means for determining priority based on the specified information, and for rearranging the at least one trained model selected by the selection means and providing it to the user based on the priority.

19. A storage step involves associating a trained model with information indicating the degree of change in the generated results produced by the trained model relative to the original data, and storing this information as evaluation information in a storage means. A request acquisition step involves acquiring request information that represents the requirements regarding the degree of change, A selection step of selecting at least one trained model from among a plurality of trained models held in the holding means, based on the request information and the evaluation information. A control method for an information processing system, characterized by having the following features.

20. A program for causing a computer to execute the control method of the information processing system described in claim 19.

Images