Method and system for determining contribution to training of generative AI model and generated product

Abstract

The present disclosure provides a solution for determining the degree of contribution of each copyright protected asset owner to training a generative AI model. In addition, the present disclosure also provides a solution for identifying a particular asset that contributes most to the generation of a particular generated product by the generative AI model. This is performed in accordance with metric dependencies relating to meta-features that make up the asset and the generated product. By determining the distribution of the contribution of the owner to the generative AI model and the specifically generated product, the contribution of the owner can be attributed. This identification is embodied in a variety of ways, for example, in a copyright attribution way or in a way that distributes revenues obtained as a result of the use of a generative AI model according to a certain financial pattern. Therefore, through the solution of the present disclosure, use of copyright protected assets during training of the generative AI model can be standardized.

Description

Technical Field

[0001] This disclosure relates to the field of generative artificial intelligence (AI) models, and more particularly to a method and system for determining the contribution of a generative AI model to training and the generated product. Background Technology

[0002] In related technologies, generative AI models are typically trained and then used to generate products. Summary of the Invention

[0003] This disclosure provides a solution for determining the contribution of each copyrighted asset owner to the training of a generative AI model. Furthermore, this disclosure provides a solution for identifying the specific asset that contributes most to the generation of a particular generative product by the generative AI model. This is performed based on a metric of relevance related to the meta-features constituting the asset and the generated product. By determining the distribution of the owner's contribution to the generative AI model and the specific generative product, the owner's contribution can be attributed. This identification can be reflected in various ways, such as by copyright attribution or by distributing revenue derived from the use of the generative AI model according to a certain financial model. Therefore, the solution of this disclosure can regulate the use of copyrighted assets in training generative AI models.

[0004] Therefore, one aspect of this disclosure provides a method for determining the distribution of contributions by asset owners to products generated by a generative artificial intelligence (AI) model, wherein the generated products are, for example, images generated by a generative AI platform. The method includes receiving multiple assets that have contributed to or are intended for training the generative AI model; that is, these multiple assets may include assets already involved in training the generative AI model or assets currently received by the generative AI model for training the model. Assets can be any media assets eligible for training the generative AI model. For example, assets can be code, scripts, images, videos, audio files, literary works, engineering or architectural designs, technological innovations, or any other copyrighted assets. Each of the multiple assets may be attributed to at least one owner of the copyright or any other proprietary right or intellectual property rights of that asset (assets may also include assets without proprietary rights, such as various images, text, audio, video, etc., whose copyrights have expired). The method further includes analyzing the multiple assets to extract meta-features from each asset. Meta-features may include asset tags, or, in the case of graphic assets, asset captions that can be associated with the asset or obtained through a multimodal language model, entity recognition in the text associated with the asset, the type of objects displayed in the asset, the style of objects displayed in the graphic asset, etc. Building upon the extraction of asset meta-features, the method further includes generating an asset dataset for each of multiple assets. This asset dataset includes owner data indicating the asset owner and the aforementioned meta-features. Therefore, the asset dataset can identify the asset owner and includes data characterizing a specific asset. The method further includes processing multiple asset datasets to determine contribution distribution data, which indicates the distribution of contributions by asset owners to the training of the generative AI model. That is, the contribution distribution defines the degree of contribution of each asset owner to the generative AI model. It should be noted that in some cases, the sum of all contributions may be less than 100% of the generative AI model's contribution, for example, when contributions also include assets whose copyrights have expired. The contribution distribution data includes a contribution score for each asset owner; therefore, the contribution distribution is determined based on the relative score of each asset owner relative to the total score of all asset owners. It is important to note that the term "score" should be interpreted as any comparable value that can be used to determine each asset owner's relative contribution to the overall model training. The method further includes output contribution distribution data, which includes contribution distribution data used, for example, to allow the payment distribution of products generated by the generative AI model to be executed based on the contribution distribution.

[0005] It should be noted that any combination of the described embodiments related to any aspect of this disclosure is applicable. In other words, any aspect of this disclosure can be defined by any combination of the described embodiments.

[0006] In some embodiments, the method further includes filtering a plurality of received assets, the filtering including excluding assets that are not qualified to be trained on a generative AI model. This includes assets that fall within the definition of unsafe working environments (NSFW), such as offensive images, nude images, etc.

[0007] In some embodiments, the method may include filtering assets without relevant rights to obtain a generated product free of any third-party contributions. For example, this may include assets whose copyrights have expired, assets that have been released to the public for free, etc.

[0008] In some embodiments, the method further includes filtering a plurality of received assets, the filtering comprising identifying one or more first assets that are identical to or have a similarity to the second asset above a selected threshold, and excluding the second asset. For example, the second asset may be an asset that is already part of the generative AI model and its training, and therefore cannot be included again when determining the contribution distribution.

[0009] In some embodiments of the method, the plurality of assets are graphic assets, and the generated product is a graphic product, i.e., a picture or image.

[0010] In some embodiments of this method, meta-features include labels for graphic assets, namely descriptive annotations defining an image or a portion thereof, explanatory text associated with the graphic asset, the style of the graphic asset, objects in the graphic asset, or any combination thereof.

[0011] In some embodiments of the method, the plurality of graphic assets include images, drawings, photographs, or any combination thereof.

[0012] In some embodiments of the method, processing multiple asset datasets includes assigning a quality score to each asset dataset, which indicates the quality of the asset's contribution to the training of the generative AI model, wherein each asset owner's contribution score is determined by the number of assets and their quality scores. That is, among other factors, each asset owner's contribution score is determined by the sum of the quality scores of all the owner's assets. The quality score of an asset is determined by at least one or any combination of the following: (1) crowdsourcing, i.e., ratings of the asset owner's users or specific assets; (2) using an AI model, such as Neural Image Evaluation (NIMA) or Aesthetic Visual Analysis (AVA); (3) comparing the asset to a database of quality-verified assets, wherein the similarity to one or more comparable assets among the quality-verified assets indicates the asset's quality score. For example, a relatively high similarity of an asset to a set of quality-verified assets indicates that the asset has a relatively high quality score.

[0013] In some embodiments of this method, each asset dataset includes a time factor that indicates at least one of the following: the time elapsed since the asset was incorporated into the training of the generative AI model, the term of the copyright or other intellectual property rights, or a combination thereof. Contribution scores may be affected by the time factor. For example, assets that have been incorporated into the generative AI model for a relatively long time compared to other assets may result in an increased contribution score for the owner. In another example, assets whose copyrights are about to expire may result in a decreased contribution score attributable to the owner.

[0014] In some embodiments, the method further includes determining, in response to a guidance prompt, the contribution of one or more asset owners to a specific generated product asset generated by the generative AI model. The guidance prompt may include written descriptions and / or image input. Determining the contribution of one or more asset owners to a specific generated product asset generated by the generative AI model in response to the guidance prompt includes: (i) extracting generated product meta-features; (ii) identifying matching assets from the plurality of assets, wherein one or more asset meta-features of the matching asset have a correlation higher than a selected threshold with one or more generated product meta-features; and (iii) defining a specific contribution score for the matching asset based on the correlation. The contribution distribution output data further includes the specific contribution score; therefore, the contribution data may include two different scores: a general contribution score and a specific contribution score.

[0015] In some embodiments of the method, the plurality of assets are graphic assets, and the generated product is a graphic product, wherein the meta-features of the generated product include the tags of the generated graphic assets, namely, descriptive annotations defining the image or portions thereof, explanatory text associated with the graphic assets, the style of the graphic assets, objects in the graphic assets, or any combination thereof.

[0016] In some embodiments of the method, the determination further includes calculating the similarity between descriptive text belonging to the asset and guidance prompts used for the identification.

[0017] In some embodiments of the method, the determination further includes calculating the similarity between a tag belonging to the asset and a contextual analysis of the guidance prompt used for the identification. The contextual analysis of the guidance prompt is an analysis of input text inserted by the product's generator / creator. This analysis aims to interpret whether the input text is related to a particular tag associated with the graphic asset.

[0018] In some embodiments of the method, the determination further includes calculating the similarity between a tag belonging to the asset and a graphical analysis of the image guidance cues used for the identification. The graphical analysis of the guidance cues is an analysis of an input image inserted by the product's generator / creator. This analysis aims to interpret whether the input image is related to a certain tag associated with the graphical asset.

[0019] In some embodiments of this method, a selected threshold is defined to obtain a selected, finite number of matching assets whose relevance meets specific conditions, such as having the highest relevance. That is, assets are ranked according to their relevance (one or more meta-features or total relevance), and matching assets are the number of assets with the highest relevance. In some embodiments, matching assets may include the highest-relevance asset for each meta-feature. The number of matching assets may be constant or updated as needed. In some other embodiments, matching assets may be assets whose relevance exceeds a selected threshold.

[0020] In some embodiments of the method, the output is triggered in response to a generated product produced by a generative AI model.

[0021] In some embodiments of the method, the output includes allocating payments associated with the generated products produced by the generative AI model based on contribution scores.

[0022] In some embodiments of the method, the output includes distribution parameters that determine the attribution of value associated with the generated product generated by the generative AI model based on contribution scores.

[0023] In some embodiments, the allocation of payments associated with products generated by the generative AI model is further based on specific contribution scores. That is, payments are allocated according to both general contribution scores and specific contribution scores. The weighting factors defining the influence of each score on payment allocation can be constants or can vary based on specific parameters for each generation of products.

[0024] In some embodiments of this method, the contribution score is determined based on at least one of the following parameters: the number of assets, a promotion factor for owners particularly relevant to the generative AI model (e.g., according to a contract or any other agreement), a variety score indicating the uniqueness of assets in the database (i.e., the rarity of that type of asset in the database of assets used to train the generative AI model), a community score indicating users' assessment of the asset's value, a diffusion factor indicating the asset's virality in the network, and the age of each asset (i.e., how long any one of the owner's assets has been part of the training set for the generative AI model). This can be calculated as the average age of the owner's multiple assets.

[0025] In some embodiments, the method further includes calculating a relative contribution parameter that indicates the relative contribution of copyrighted assets to model training compared to uncopyrighted assets (i.e., assets whose copyrights have expired). For example, X% of the assets used to train the model may be copyrighted assets, while 100-X% of the assets used to train the model may be uncopyrighted assets. The distributed output data includes this relative contribution parameter. Therefore, if a financial model is applied based on the distributed output data, the allocation of revenue from using the generative AI model can also be determined based on the relative contribution parameter.

[0026] Another aspect of this disclosure provides a system for determining the distribution of contributions by asset owners to products generated by generative artificial intelligence (AI) models. The system includes: At least one processing circuit, i.e., at least one processing unit or processor. The at least one processing circuit may be cloud-based, centralized, or distributed. The at least one processing circuit is configured to: receive multiple assets that contribute to or are intended for training a generative AI model, wherein each of these multiple assets belongs to at least one owner who owns the copyright to that asset; analyze the multiple assets to extract meta-features from each asset, the meta-features including the asset's tag, or, in the case of a graphic asset, the asset's descriptive text, the type of objects displayed in the asset, the style of the objects displayed in the asset, etc.; generate an asset dataset for each of the multiple assets, the asset dataset including owner data indicating the asset owner and the meta-features; process the multiple asset datasets to determine contribution distribution data indicating the contribution distribution of asset owners to the training of the generative AI model, wherein the contribution distribution data includes a contribution rating for each asset owner, and therefore the contribution distribution is determined based on a relative rating of each asset owner relative to the total rating of all asset owners; and output contribution distribution output data, which includes contribution distribution data.

[0027] In some embodiments of the system, the processing circuitry is further configured to filter multiple received assets. This filtering includes excluding assets that are not qualified to be trained on a generative AI model.

[0028] In some embodiments of the system, the processing circuitry is further configured to filter multiple received assets, the filtering including identifying and excluding assets that are identical to or have a similarity to previous assets that is above a selected threshold.

[0029] In some embodiments of the system, the plurality of assets are graphic assets, and the generated product is a graphic product.

[0030] In some embodiments of this system, meta-features include labels for graphic assets, descriptive text associated with the graphic assets, styles of the graphic assets, objects within the graphic assets, or any combination thereof. Meta-features can be derived from a multimodal language model applied to the graphic assets. For example, by using a multimodal language model, descriptive text, text, or words can be obtained from an image.

[0031] In some embodiments of the system, the plurality of graphic assets include images, drawings, photographs, or any combination thereof.

[0032] In some embodiments of the system, the processing circuitry is further configured to determine, in response to a guidance prompt, the contribution of one or more asset owners to a specific generated product asset generated by the generative AI model. This guidance prompt may include written descriptions and / or image input. Determining the contribution of one or more asset owners to a specific generated product asset generated by the generative AI model in response to the guidance prompt includes: extracting generated product meta-features; identifying matching assets from the plurality of assets, wherein one or more asset meta-features of the matching asset have a correlation higher than a selected threshold with one or more generated product meta-features; and defining a specific contribution score for the matching asset based on the correlation. Therefore, the contribution distribution output data further includes the specific contribution score.

[0033] In some embodiments of the system, the plurality of assets are graphic assets, and the generated product is a graphic product, wherein the meta-features of the generated product include the tags of the generated graphic assets, the descriptive text associated with the graphic assets, the style of the graphic assets, objects in the graphic assets, or any combination thereof.

[0034] In some embodiments of the system, the determination further includes calculating the similarity between descriptive text belonging to an asset and guidance prompts used for the identification.

[0035] In some embodiments of the system, the determination further includes calculating the similarity between tags belonging to an asset and contextual analysis of guidance prompts used for the identification.

[0036] In some embodiments of the system, the determination further includes calculating the similarity between tags belonging to assets and graphical analysis of image guidance cues used for the identification.

[0037] In some embodiments of the system, a selected threshold is defined to obtain a limited number of matching assets whose relevance meets specific conditions, such as having the highest relevance. That is, assets are ranked according to their relevance (one or more meta-features or total relevance), and the matching assets are the number of assets with the highest relevance.

[0038] In some embodiments of the system, the output is triggered in response to a product generated by a generative AI model.

[0039] In some embodiments of the system, the output includes allocating payments associated with the generated products produced by the generative AI model based on contribution scores.

[0040] In some embodiments of the system, the output includes distribution parameters that determine the attribution of value associated with the generated products generated by the generative AI model based on contribution scores.

[0041] In some embodiments of the system, the contribution score is determined based on at least one of the following parameters: the number of assets, a promotion factor for the owner particularly relevant to the generative AI model (e.g., according to a contract or any other condition), a group score indicating users' assessment of the asset's value, a diffusion factor indicating the asset's diffusion within the network, and the age of each asset (i.e., how long any one of the owner's assets has been part of the training set for the generative AI model). This can be calculated as the average age of the owner's multiple assets.

[0042] In some embodiments of the system, processing multiple asset datasets includes assigning a quality score to each asset dataset, the quality score indicating the quality of the asset's contribution to the training of a generative AI model, wherein the contribution score for each asset owner is determined by the number of assets and their quality scores. The quality score of an asset is determined by at least one or any combination of the following: (1) crowdsourcing, i.e., ratings of a specific asset by a user of the asset owner; (2) using an AI model, such as Neural Image Evaluation (NIMA) or Aesthetic Visual Analysis (AVA); (3) comparing the asset to quality-verified assets stored in a database, wherein the similarity to one or more comparable assets among the quality-verified assets indicates the asset's quality score.

[0043] In some embodiments of the system, each asset dataset includes a time factor that indicates at least one of the following: the time elapsed since the asset was incorporated into the training of the generative AI model, the copyright term, or a combination thereof. Contribution scores are influenced by the time factor.

[0044] In some embodiments of the system, at least one processing circuit is further configured to calculate a relative contribution parameter, which indicates the relative contribution of copyrighted assets to model training compared to uncopyrighted assets (i.e., assets whose copyright terms have expired). The distributed output data includes this relative contribution parameter.

[0045] Example Based on various aspects of this disclosure, the following are optional embodiments and combinations thereof.

[0046] 1. A method for determining the contribution to a generated product produced by a generative AI model, comprising: Receive multiple assets that contribute to the training of a generative AI model or are intended to be used to train a generative AI model, wherein each of the multiple assets belongs to at least one owner who owns the copyright to the asset. The plurality of assets are analyzed to extract meta-features from each asset, and an asset dataset is generated for each of the plurality of assets, the asset dataset including owner data indicating the asset owner and the meta-features; Multiple asset datasets are processed to determine contribution distribution data, which indicates the distribution of asset owners' contributions to the training of the generative AI model, wherein the contribution distribution data includes contribution scores for each asset owner; Output contribution distribution data, which includes contribution distribution data.

[0047] 2. The method according to Embodiment 1 includes filtering a plurality of received assets, the filtering including excluding assets that are not qualified to train a generative AI model.

[0048] 3. The method according to Example 2, wherein the filtering further includes identifying a first asset that is the same as or has a similarity to the second asset above a defined threshold, and excluding the second asset.

[0049] 4. The method according to any one of Examples 1-3, wherein the plurality of assets are graphic assets and the generated product is a graphic product.

[0050] 5. The method according to Example 4, wherein the meta-features include a label of a graphic asset, descriptive text associated with the graphic asset, a style of the graphic asset, an object in the graphic asset, or any combination thereof.

[0051] 6. The method according to embodiment 4 or 5, wherein the plurality of graphic assets includes images, drawings, photographs or any combination thereof.

[0052] 7. The method according to any one of Examples 1-6 further includes determining, in response to guidance prompts, the contribution of one or more asset owners to a specific generated product asset generated by the generative AI model, said determination comprising: Extract and generate product meta-features; Identify matching assets from the plurality of assets, wherein one or more asset meta-features of the matching asset have a correlation with one or more generated product meta-features that is higher than a selected threshold; Define a specific contribution score for matched assets based on relevance; The contribution distribution output data further includes the specific contribution score attributed to the asset owner.

[0053] 8. The method according to Example 7, wherein the plurality of assets are graphic assets, the generated product is a graphic product, and wherein the meta-features of the generated product include tags of the generated graphic assets, descriptive text associated with the graphic assets, style of the graphic assets, objects in the graphic assets, or any combination thereof.

[0054] 9. The method according to embodiment 7 or 8, wherein the determination further includes calculating the similarity between descriptive text belonging to an asset and guidance prompts used for the identification.

[0055] 10. The method according to any one of Examples 7-9, wherein the determination further includes calculating the similarity between a tag belonging to an asset and a contextual analysis of the guidance prompt used for the identification.

[0056] 11. The method according to any one of Examples 7-10, wherein the determination further includes calculating the similarity between a tag belonging to an asset and a graphical analysis of an image guidance cues used for the identification.

[0057] 12. The method according to any one of Examples 7-11, wherein a selected threshold is defined to obtain a selected finite number of matching assets that have the highest relevance or have a relevance that satisfies a specific condition.

[0058] 13. The method according to any one of embodiments 1-12, wherein the output is triggered in response to a generated product generated by a generative AI model.

[0059] 14. The method according to Example 13, wherein the output includes allocating payments associated with the generated products generated by the generative AI model based on contribution scores.

[0060] 15. The method according to any one of Examples 1-14, wherein the contribution score is determined based on at least one of the following parameters: the number of assets, the age of each asset, and a group score indicating the user's assessment of the asset's value.

[0061] 16. A system for determining the contribution to a generated product produced by a generative AI model, comprising: At least one processing circuit is configured to: Receive multiple assets that contribute to the training of a generative AI model or are intended to be used to train a generative AI model, wherein each of these multiple assets belongs to at least one owner who owns the copyright to the asset. The plurality of assets are analyzed to extract meta-features from each asset, and an asset dataset is generated for each of the plurality of assets, the asset dataset including owner data indicating the asset owner and the meta-features; Multiple asset datasets are processed to determine contribution distribution data, which indicates the distribution of asset owners' contributions to the training of a generative AI model. This contribution distribution data includes contribution scores for each asset owner. Output contribution distribution data, which includes contribution distribution data.

[0062] 17. The system according to embodiment 16, wherein the processing circuitry is further configured to filter a plurality of received assets, the filtering including excluding assets that are not qualified to train a generative AI model.

[0063] 18. The system according to Example 17, wherein the filtering further includes identifying a first asset that is identical to or has a similarity to the second asset above a defined threshold, and excluding the second asset.

[0064] 19. The system according to any one of embodiments 16-18, wherein the plurality of assets are graphic assets and the generated product is a graphic product.

[0065] 20. The system according to embodiment 19, wherein the meta-features include tags of a graphic asset, descriptive text associated with the graphic asset, style of the graphic asset, objects in the graphic asset, or any combination thereof.

[0066] 21. The system according to embodiment 19 or 20, wherein the plurality of graphic assets includes images, drawings, photographs or any combination thereof.

[0067] 22. The system according to any one of embodiments 16-21, wherein the processing circuitry is further configured to determine, in response to a guidance prompt, the contribution of one or more asset owners to a specific generated product asset generated by a generative AI model, the determination comprising: Extract and generate product meta-features; Identify matching assets from the plurality of assets, wherein one or more asset meta-features of the matching asset have a correlation with one or more generated product meta-features that is higher than a selected threshold; Define a specific contribution score for matched assets based on relevance; The contribution distribution output data further includes the specific contribution score attributed to the asset owner.

[0068] 23. The system according to embodiment 22, wherein the plurality of assets are graphic assets, the generated product is a graphic product, and wherein the meta-features of the generated product include tags of the generated graphic assets, descriptive text associated with the graphic assets, the style of the graphic assets, objects in the graphic assets, or any combination thereof.

[0069] 24. The system according to embodiment 22 or 23, wherein the determination further includes calculating the similarity between descriptive text belonging to an asset and guidance prompts used for the identification.

[0070] 25. The system according to any one of embodiments 22-24, wherein the determination further includes calculating the similarity between a tag belonging to an asset and a contextual analysis of guidance prompts used for the identification.

[0071] 26. The system according to any one of embodiments 22-25, wherein the determination further includes calculating the similarity between a tag belonging to an asset and a graphical analysis of an image guidance prompt used for the identification.

[0072] 27. The system according to any one of Examples 22-26, wherein a selected threshold is defined to obtain a selected finite number of matching assets that have the highest relevance or have a relevance that satisfies a specific condition.

[0073] 28. The system according to any one of embodiments 16-27, wherein the output is triggered in response to a generated product generated by a generative AI model.

[0074] 29. The system according to Example 28, wherein the output includes allocating payments associated with the generated products generated by the generative AI model based on contribution scores.

[0075] 30. The system according to any one of Examples 16-29, wherein the contribution score is determined based on at least one of the following parameters: the number of assets, the age of each asset, and a group score indicating the user's assessment of the asset's value.

[0076] 31. The system according to any one of embodiments 16-30, wherein the at least one processing circuit is further configured to calculate a relative contribution parameter indicating the relative contribution of a copyrighted asset to the training of a generative AI model between copyrighted and uncopyrighted assets; wherein the distributed output data includes the relative contribution parameter.

[0077] 32. The system of claim 28, wherein the output includes distribution parameters that determine the attribution of value associated with the generated product generated by the generative AI model based on contribution scores.

[0078] 33. The method of claim 13, wherein the output includes a distribution parameter that determines the attribution of value associated with the generated product generated by the generative AI model based on the contribution score. Attached Figure Description

[0079] To better understand the subject matter disclosed herein and to illustrate how it can be implemented in practice, embodiments of this disclosure will now be described by way of non-limiting example only, with reference to the accompanying drawings.

[0080] Figure 1A-1DThis is a flowchart of a method for determining the contribution of a generative AI model to the generation of generated products and allocating the revenue obtained from the generation of the generated products, according to one aspect of this disclosure. Detailed Implementation

[0081] The following figures are provided to illustrate embodiments and implementations of this disclosure.

[0082] Figure 1A-1D This is a flowchart illustrating a method for determining the contribution of a generative AI model to the generated product. (By...) Figure 1A-1D The example shown allows attribution based on the relative contributions of asset owners who contribute to the training of the generative AI model and, consequently, to the generation of the generative products produced by the model. If the use of a particular model is based on a financial model, then the method can allocate revenue generated from using the generative AI model to produce products based on the relative contributions of owners of assets typically used to train the model, or based on a specific correlation between an asset and the generative AI model's generation of a particular product.

[0083] Although Figure 1A-1D The example focuses on generative AI models that generate graphical products, but in a similar way, the method can be adapted to different generative AI models as necessary.

[0084] Figure 1A The example illustrates the steps involved in establishing a new asset owner. It should be noted that while this example provides a comprehensive and holistic explanation of the process, not all steps in the diagram are necessary to perform the methods according to this disclosure.

[0085] therefore, Figure 1A The example method includes a registration process for new asset owners. In this example, the asset owner is referred to as an artist, but it should be noted that the asset owner can be different from the artist who created the asset. After registration, the asset owner enters multiple asset files, each including the asset and optional data associated with it, such as the asset's description, ownership distribution, copyright term, descriptive text, or any other metadata associated with the asset. The entered assets are then filtered by a "job not safe" filter, which is configured to identify offensive assets or assets unsuitable for training generative AI models. Additionally, assets are filtered by a duplicate filter, which is configured to identify assets that have already been entered into the generative AI model by the same owner or another owner. Both filters can trigger alerts or notifications to the owner indicating the filtered assets and the reasons for their rejection.

[0086] Assets that pass the filtering process then proceed to the metadata and feature extraction process. An asset feature dataset is generated for each asset, comprising a list of descriptive and quantitative features. This dataset may include descriptive text or tags associated with the asset. The feature extraction process from images includes an image-to-text process, where an AI model is applied to the image, and the model's output generates descriptive text for the image. The descriptive text may include descriptions of the content displayed in the image and the image's style, such as the style of a specific artist, era, or image environment. Some assets may already be tagged upon receipt, and the extraction process extracts these tags. Further information extracted from the image is the copyright information associated with it.

[0087] Figure 1A The text further illustrates the steps a user takes to generate new products using a generative AI model, whether these are graphical products (e.g., images or pictures), audio products, coded products, or text-based products. The user inputs guidance input, which describes the product expected from the generative AI model. This input can be text prompts, images, pictures, audio files, or any combination thereof. The generative AI model then outputs the generated product and calculates ownership among the owners.

[0088] Figure 1B An example illustrates the attribution assessment process for the generated products. The guiding input and the generated products (referred to as generated assets in the diagram) are analyzed to extract meta-features that can be compared with the asset dataset used to train the generative AI model.

[0089] Figure 1C An example illustrates the extraction of metadata. This diagram illustrates the process of extracting features from the generated assets and optional guidance inputs (including images, descriptive text, and labels). It should be noted that the guidance inputs may include only some of their components; this method does not require all components.

[0090] For image input, the image is processed to extract its characteristic features and determine an estimate of the image's descriptive text. For any image input, an image input feature dataset is generated according to the above processing. The generated image is then processed similarly, and its features are extracted. In other words, an AI model capable of generating descriptive text from images is applied. Therefore, the text generated from the input image can be compared with the text in the asset dataset used to train the generative AI model.

[0091] For the descriptive text input, the descriptive text is processed to extract its characteristic features, which can be compared with the features of the asset dataset of the generative model.

[0092] For the tag input, the tags are processed to evaluate their relevance to the image. This evaluation can be based on other users ranking the image tags, or on manual tags from one or more reliable sources or users. Alternatively, AI models can be used to analyze the tags to evaluate their relevance or matching to the image. Tags can also be validated by the results of an AI model that generates descriptive text for the tagged image. In other words, for the tag input, a data quality sanity check is performed to ensure that the tags are relevant to the image.

[0093] A generated image dataset is generated based on meta-features extracted from the guiding input and the generated images. These meta-features are comparable to the meta-features of the assets in the dataset associated with each asset used in the training of the generative AI model.

[0094] Then, according to Figure 1B The process illustrated in the example involves matching the generated image dataset with the most relevant asset dataset based on certain relevance criteria. The matching process is as follows: Figure 1D As shown. For example, for each meta-feature, a relevance score can be calculated to identify the correlation between the meta-features of the generated image and the meta-features of the assets that contributed to the model training. A different weighting factor can be used to calculate the total relevance score for each meta-feature. The total relevance score is calculated based on the sum of all feature relevance scores, with each score multiplied by its weighting factor. Then, the best match between the generated image dataset and the asset dataset can be determined. The relevance condition can be either identifying a certain number of asset datasets with the highest relevance scores, or a defined threshold that identifies any asset dataset with high relevance as a matching asset. For each asset found to match the generated image, an attribution score is calculated. The attribution score determines the evaluated contribution of a particular asset to the generated image. Therefore, there exists a specific distribution of contributions from several assets to the generation of a particular generated image.

[0095] return Figure 1BFor each generated image, two contribution scores are calculated: (1) a general contribution score, which indicates the contribution distribution of all assets used to train the model before a certain point in time (e.g., the point in time when the generated image was generated); and (2) a specific contribution score, which indicates the contribution distribution of the asset most likely to influence the specific generation of the generated image. By combining these two scores according to a selected combination rule, a final contribution score can be determined, indicating the contribution distribution of the assets to the generation of the specific generated image. The combination rule can be based on a weighting factor applied to each score. That is, a higher weighting factor can be applied to one of the scores according to the desired allocation pattern. For example, in some patterns, the weighting factor of the specific contribution score can be higher than that of the general contribution score, thus assigning a higher weight to the specific asset that contributes the most to the generation of the specific image. In other patterns, the weighting factor of the general contribution score can be higher.

[0096] Based on the total contribution score, a financial model can be applied to allocate revenue related to the use of the generative AI model. According to the methodology of this disclosure, the contribution of asset owners to training the generative AI model and / or generating specific images within the generative AI model can be evaluated, and owners can be rewarded based on their relative contributions. Therefore, this disclosure lays the foundation for any financial model applicable to such generative AI models to reward asset owners who contribute to the model.

Claims

1. A method for determining contribution of generative artificial intelligence (AI) model to generated products, comprising: receiving a plurality of assets that contribute to training of the generative AI model or intended for training of the generative AI model, wherein each asset of the plurality of assets is attributed to at least one owner that owns copyright of the asset; analyzing the plurality of assets to extract meta-features from each asset and generate an asset dataset for each asset of the plurality of assets, the asset dataset comprising owner data indicative of an asset owner and the meta-features; processing the asset dataset to determine contribution distribution data indicative of a contribution distribution of the asset owners to the training of the generative AI model, wherein the contribution distribution data comprises a contribution score of each of the asset owners; outputting contribution distribution output data comprising the contribution distribution data.

2. The method of claim 1, further comprising filtering the received plurality of assets, the filtering comprising excluding assets that are not eligible for training of the generative AI model.

3. The method of claim 2, wherein, the filtering further comprising identifying a first asset that is identical or similar to a second asset above a defined threshold and excluding the second asset.

4. The method of claim 1, wherein, the plurality of assets are graphic assets and the generated products are graphic products.

5. The method of claim 4, wherein, the meta-features comprise a label of the graphic asset, a description associated with the graphic asset, a style of the graphic asset, an object in the graphic asset, or any combination thereof; wherein the plurality of graphic assets comprise images, drawings, photos, or any combination thereof.

6. The method of claim 1, further comprising determining contribution of one or more of the asset owners to a particular generated product asset generated by the generative AI model in response to a guidance prompt, the determining comprising: extracting generated product meta-features; identifying matching assets from the plurality of assets, one or more asset meta-features of the matching assets are similar to one or more of the generated product meta-features above a selected threshold; defining a particular contribution score for the matching assets according to the similarity; wherein the contribution distribution output data further comprises the particular contribution score attributed to the asset owners.

7. The method of claim 6, wherein the plurality of assets are graphic assets and the generated products are graphic products, and wherein the generated product meta-features comprise a label of a generated graphic asset, a description associated with the graphic asset, a style of the graphic asset, an object in the graphic asset, or any combination thereof; wherein the determining further comprising computing similarity between a description attributed to an asset and the guidance prompt used for the identifying; wherein the determining further comprises computing similarity between a label attributed to an asset and contextual analysis of the guidance prompt used for the identifying.

8. The method of claim 7, wherein the determining further comprises computing similarity between a label attributed to an asset and graphic analysis of an image guidance prompt used for the identifying. wherein The selected threshold is defined to obtain a selected limited number of matching assets having a relevance satisfying a certain condition.

9. The method of claim 1, wherein, The output is triggered in response to a generated product generated by the generative AI model.

10. The method of claim 9, wherein the method further comprises determining a distribution parameter of attribution of a value associated with a generated product generated by the generative AI model according to the contribution score.

11. The method of claim 1, wherein the contribution score is determined according to at least one of: a number of the assets, an age of each of the assets, a crowd score indicative of an evaluation of a user on a value of the assets.

12. A system for determining a contribution to a generated product generated by a generative artificial intelligence (AI) model, comprising: at least one processing circuitry configured to: receive a plurality of assets that contribute to a training of the generative AI model or intended to be used to train the generative AI model, wherein each of the plurality of assets is attributed to at least one owner that owns a copyright of the asset; analyze the plurality of assets to extract meta-features from each of the assets and generate an asset dataset for each of the plurality of assets, the asset dataset comprising owner data indicative of an asset owner and the meta-features; process the asset dataset to determine contribution distribution data indicative of a contribution distribution of the asset owners to the training of the generative AI model, wherein the contribution distribution data comprises a contribution score of each of the asset owners; and output contribution distribution output data comprising the contribution distribution data.

13. The system of claim 12, wherein the plurality of assets are graphic assets and the generated product is a graphic product; wherein, the meta-features comprise a label of the graphic asset, a description associated with the graphic asset, a style of the graphic asset, an object in the graphic asset, or any combination thereof; wherein the plurality of graphic assets comprise an image, a drawing, a photograph, or any combination thereof.

14. The system of claim 12, wherein the processing circuitry is further configured to determine a contribution of one or more of the asset owners to a particular generated product asset generated by the generative AI model in response to a guidance prompt, the determining comprising: extracting generated product meta-features; identifying matching assets from the plurality of assets, one or more asset meta-features of the matching assets having a relevance higher than a selected threshold to one or more of the generated product meta-features; defining a particular contribution score for the matching assets according to the relevance; wherein the contribution distribution output data further comprises the particular contribution score attributed to the asset owners.

15. The system of claim 14, wherein the plurality of assets are graphic assets and the generated product is a graphic product, and wherein the generated product meta-features comprise a label of a generated graphic asset, a description associated with the graphic asset, a style of the graphic asset, an object in the graphic asset, or any combination thereof.

16. The system of claim 14, wherein the determining further comprises computing a similarity between a description attributed to an asset and the guidance cues used for the identification; wherein the determining further comprises computing a similarity between a label attributed to an asset and a contextual analysis of the guidance cues used for the identification; wherein the determining further comprises computing a similarity between a label attributed to an asset and a graphical analysis of the image guidance cues used for the identification.

17. The system of claim 14, wherein the selected threshold is defined to obtain a selected limited number of matching assets having a relevance satisfying a certain condition.

18. The system of claim 12, wherein the output is triggered in response to a generated product generated by the generative AI model; wherein the output comprises a distribution parameter of attribution of a value associated with a generated product generated by the generative AI model according to the contribution score determination.

19. The system of claim 12, wherein the contribution score is determined according to at least one parameter of: a number of the assets, an age of each of the assets, a crowd score indicative of an evaluation of a value of the assets by users.

20. The system of claim 12, wherein the at least one processing circuitry is further configured to compute a relative contribution parameter indicative of a relative contribution of a copyright-protected asset and a non-copyright-protected asset to a training of the generative AI model; wherein the contribution distribution output data comprises the relative contribution parameter.

Images