Generative image value evaluation method and system for data market

Abstract

The invention discloses a data market-oriented generative image value evaluation method and system. According to the method, value evaluation is divided into three scenes according to a use mode of a generative model: for a single image with an unknown generative model, value evaluation is carried out by quantifying original image contribution and market demands; for unconditional batch image generation of the diffusion model, value distribution is carried out by calculating the total utility of a generation set and combining a Shapley value method of Monte Carlo sampling, and then market demand value evaluation is carried out; for conditional generation images of the diffusion model, an optimization model with maximization of the intermediary income as a target and effectiveness of a buyer as a constraint is constructed, and a genetic algorithm is used for solving, so that optimal value evaluation of a customized image set is realized. According to the method, the problems of complex value source, difficulty in quantification and high-efficiency value evaluation of the generative image are solved, contributions of all parties can be accurately quantified, and the transaction efficiency of the data market is improved.

Description

[0001] This invention relates to the intersection of data market valuation and artificial intelligence technology, and in particular to a method and system for efficiently evaluating the value of images generated by generative artificial intelligence models (such as diffusion models) in different scenarios within the data market. Background Technology

[0002] A data marketplace is a platform that trades data resources and AI model services, acting as an intermediary between data providers and data buyers / users. The origins of data marketplaces can be traced back to 2008. With the development of big data and artificial intelligence technologies, the products traded in the market have expanded from traditional raw data to machine learning models. A typical data marketplace usually consists of three parties: buyers, intermediaries, and sellers. Sellers provide data of varying quality and sell it to intermediaries in exchange for payment. Intermediaries create data products and assess their value. Buyers select the data products they want and pay for them.

[0003] Generative images refer to novel images created by artificial intelligence models using deep learning techniques. The emergence of generative adversarial networks (GANs) and diffusion models has enabled generative images to possess good quality, diversity, and controllability. Due to these advantages, generative images are now widely used in fields such as gaming and social media, and have become a new product type with great potential in the data market. For example, in image marketplaces such as ArtBlocks, Getty Images, and Shutterstock, generative images are among the main art pieces traded. In 2016, the AI-generated work "GCHQ" sold for $8,000 at a Google charity auction. Christie's auction house sold "Portrait of Edmond Belamy," created using a generative adversarial network, for $432,500 in 2018, and the NFT work "Everyday: The First 5000 Days" for $69.35 million in 2021. In 2024, Sotheby's auction house sold the AI ​​work "God of AI" for $1,084,800.

[0004] In the current state of research on valuation in the data market, there is no method to solve the evaluation problem of new data products such as generative images. Although generative images have many advantages, there are still many challenges in valuing them. What influences the price of generative images? This is the primary and fundamental challenge in constructing a generative image valuation mechanism. The valuation of traditional data products (such as standardized datasets) usually relies on factors such as the quality and size of the dataset and market supply and demand. Generative images also need to consider the influence of supply and demand, but unlike traditional datasets, they are generated by generative models trained on original image datasets, so the influence of the generative model itself and the original image dataset must be considered. When the generative model is used in different ways, the generated scenarios will also be different, and the factors affecting the price of generative images will also be different in different scenarios. Therefore, the primary challenge is to identify the factors affecting the price of generative images. The second challenge is to identify and quantify the key value-influencing factors for the distinctly different scenarios in generative images, and then construct a corresponding valuation model. The valuation models of traditional products are often based on cost or simple supply and demand curves, which cannot be applied to the valuation of generative images because they cannot capture the complex value composition of generative images. There are two main difficulties in modeling the factors that influence the price of generated images: First, how to find or design suitable, calculable metrics for each key factor. Second, how to integrate these metrics into a unified value assessment model to determine the final price of the generated image.

[0005] Therefore, there is an urgent need in this field for a value assessment technology that can adapt to various generation scenarios, quantify value contribution, and conform to market rules. Summary of the Invention

[0006] By examining existing technologies and in order to overcome their shortcomings and better assess the value of generative images, this invention provides a method and system for assessing the value of generative images for the data market. In this invention, scenarios are divided into three categories based on the different usage modes of the generative model: Unknown generative model scenario: The intermediary uses a pre-trained generative model to generate a new image based on a purchased original image. Unconditional generative model batch generation scenario: The intermediary uses a diffusion model to generate multiple new images in batches based on purchased original images. Conditional generative model batch generation scenario: The intermediary uses a conditional diffusion model based on purchased original images to generate multiple new images according to conditions proposed by the buyer. The purpose of this invention is to solve the comprehensive value assessment problem from single images to batch generated images, and from unconditionally generated images to conditionally customized images.

[0007] To achieve the above objectives, a generative image value assessment method and system for the data market is characterized by dividing the value assessment into three scenarios based on the usage patterns of the generative model, and constructing value assessment models for each scenario. This includes:

[0008] Scenario 1: Unknown Generation Model. Since the generation model is unknown, the core influencing factors on the value of the generated image are the influence of the original image and external market demand. The price impact of the original image on the generated image is modeled and quantified using similarity indices (SSIM and LPIPS). The demand formula from transformational economics is used to model the relationship between market demand and price of the generated image. A value assessment formula for the generated image in this scenario is proposed, combining the influence of these two factors.

[0009] Scenario 2: Batch Generation Scenario with Unconditional Generative Model. The generative model is known, and the output is an image set, making value assessment more complex. Besides the influence of the original images and external market demand, the diversity quality among multiple images within a batch must also be considered. Since the generative model is known, the influence of the original images is no longer measured by similarity; instead, the association probability between generated and original images is quantified using the principle of a diffusion model. For measuring batch image diversity, we choose the LPIPS index to calculate the pairwise differences between images of a single category, reflecting the diversity quality of the image set, and the IS index to directly calculate the diversity quality of multi-category image sets. Since image quality is evaluated from a set perspective, a Shapley value-based valuation method is proposed to determine the price of each generated image. The Shapley value is defined as the marginal contribution of a single generated image to the overall set, and the final price of each generated image is given by combining the market demand and price function modeled in the first scenario.

[0010] Scenario 3: Batch generation scenario using a conditional generation model. The price influencing factor shifts from market demand to the specific needs of buyers. Image quality, besides considering the value derived from the original image, also needs to consider "the extent to which the image satisfies the buyer's intent," i.e., the alignment between the generated image and the text prompts provided by the buyer. Due to high costs, it is assumed that the intermediary directly sells customized image sets. For the influence of the original image, the FID metric is used to measure the similarity between the entire generated image set and the original image. For the measurement of image-text alignment, the CLIP Score metric is chosen. The quality score of the generated image set is quantified based on these two metrics. Then, based on the quality score, the buyer's willingness-to-pay function and utility function are designed from the perspective of specific buyer needs. An optimization problem maximizing intermediary revenue is constructed while ensuring that buyer utility is non-negative and maximized. A genetic algorithm is used to solve the optimization problem, achieving the value assessment of the generated image set.

[0011] Embodiments of the present invention will become apparent from the following description, or may be learned by practice of the invention. Attached Figure Description

[0012] The accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below.

[0013] Figure 1 This is a schematic diagram illustrating the overall scenario and value assessment of the method described in this invention;

[0014] Figure 2 This is a schematic diagram of the first scenario (value assessment of a single generated image) in an embodiment of the present invention;

[0015] Figure 3 This is a schematic diagram of the second scenario (unconditional batch generation of image value assessment) in an embodiment of the present invention;

[0016] Figure 4 This is a schematic diagram of the third scenario (value assessment of conditionally generated image set) in an embodiment of the present invention;

[0017] Figure 5 System architecture block diagram for implementing the method described in this invention; Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Those skilled in the art should understand that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0019] This invention follows Figure 1 The system demonstrates a scene recognition and value assessment framework. It first receives a generation request, then identifies the scene to which it belongs (Scene 1, 2, or 3), calls the corresponding value assessment method for processing, and finally outputs the price.

[0020] Example 1: First Scenario Application (Value Assessment of a Single Generated Image with an Unknown Generation Model)

[0021] This embodiment corresponds to Figure 1 For a detailed illustration of "Scene 1", please refer to [link / reference]. Figure 2 .

[0022] Suppose the seller provides the original image dataset, and the intermediary uses a pre-trained, unknown generative model to generate a new image based on the original image dataset provided by the seller. Let the original image dataset be I = {i1, i2, ..., i...}. n The set of prices corresponding to these objects is P = {p1, p2, ..., p}. n}, where image i j The price of ∈I is p j∈P. Let the image generated by the intermediary be... On the one hand, the price contribution of the original image to the generated image is measured based on the similarity between the generated image and the original image. On the other hand, as a commodity, the price of the generated image is also influenced by market demand according to economic principles. Therefore, a value assessment function for a single generated image is constructed based on image content and market demand. Specific steps include:

[0023] Step 1: Calculate the content-based value assessment of a single generated image using the price of the original image.

[0024] Step 2: Calculate the value assessment of the generated image based on market demand.

[0025] Step 3: Final value assessment.

[0026] The following will explain steps 1 to 3 in detail:

[0027] The price of the generated image can be viewed as a weighted sum of the prices of the original images, where the weights represent the contribution of the original images, i.e., the similarity between the content of the generated and original images. We use the SSIM metric and LPIPS to calculate the similarity. A higher similarity indicates a greater contribution of the original image to the generated image. The closer the SSIM value is to 1, the more similar the two images are. The lower the LPIPS value, the more similar the two images are.

[0028] Given a generated image and the original image i j Let the weights of the two be w. j , here w j The value of w ranges from 0 to 1. The more similar two images are, the greater their weight. Therefore, the SSIM metric can be directly used for weighting, i.e., w j =SSIM j The LPIPS metric needs to be mapped to weights, i.e., w j =1-LPIPS j . Then i j right The price impact can be expressed as the product of the weight and the price: w j p j Note that there are multiple original images, so a simple way is to sum all the original image pairs in I. Price contribution: However, there may be similar content between the original images. The above calculation method will lead to repeated calculation of these similar content pairs. The contribution of the original image dataset I led to an overestimation of its value. Borrowing the inclusion-exclusion principle from set theory, we can more accurately calculate the value of the original image dataset I. The price contribution. The inclusion-exclusion principle is used to calculate the price contribution of multiple sets {A1, A2, ..., A...}. j Size of the union of}

[0029]

[0030] Replacing the sets in the inclusion-exclusion principle with images, the first term in the above formula calculates the price contribution of all original images to the generated image, i.e. The second term corresponds to the similar parts of the two original images in the generated image. The price impact. For example, for two original images i j and i k They and The similarity is w j and w k Let i j and i k The similar parts of the image content are i jk i j and i k The similarity is w jk . Then i jk and The similarity can be represented as w jk min(w j ,w k ), its upper limit is min(w) j ,w k For example, if with i j The similarity is 0.6. with i k The similarity is 0.4, then with i jk The similarity has an upper limit of 0.4. Combining this similarity with the price of the original image, we can obtain its effect on the generated image. The price impact, i.e., w jk w l p l , where i l =argmin l＝j,k w l The third term in the formula corresponds to the similar parts of the three original images. The price impact. Similarly, each term of the inclusion-exclusion principle corresponds to a specific number of similar parts of the original image. Price impact.

[0031] The price formula for the original image to the generated image can be accurately derived based on the inclusion-exclusion principle, but the computational complexity is exponentially related to the number of original images. Specifically, the complete expansion of formula (1) contains 2 n-1 subsets, therefore the time complexity is O(2^3). n To reduce computational complexity, only the first two terms of the inclusion-exclusion principle are used to approximate the impact of the original image on the price of the generated image:

[0032]

[0033] The first term calculates the similarity price contribution of the n original images to the generated image, while the second term calculates the impact of the pairwise similarity between the n original images on the price of the generated image. The total number of original image pairs is... Because the first and second terms have different computational magnitudes, the price contribution of the original image might be excessively reduced, resulting in an undervalued price. Therefore, a penalty coefficient is set for the second term. Balancing the difference in magnitude between the first two terms, we obtain the following formula:

[0034]

[0035] We introduce a price function based on demand. Because intermediaries in the data market have the power to assess value, the higher the demand for an image, the higher the price they will set. Therefore, we choose a positive price function:

[0036] P Q =c+d·Q(d>0) (4)

[0037] Where c is the price of the generated image under conditions of no change in demand intensity. d reflects the degree of influence of changes in demand on the price of the generated image; it is a constant set by intermediaries based on market characteristics and image properties. Q is the demand intensity for each generated image, measured by market demand.

[0038] Based on the above discussion, the price of a single generated image consists of two parts: a price based on the image content and a price based on market demand assessment. To offset the initial cost of purchasing the original images, the price of each generated image needs to be higher than the total price of all the original images. Therefore, the price in the first part is taken as the maximum value between the price of the first part and the price of each image in the original image dataset. The final generated image... price for:

[0039]

[0040] Where P I Partially, it requires calculating all original image pairs, with a time complexity of O(n^2). 2 ).

[0041] Example 2: Second Scenario Application (Unconditional Batch Generation Image Value Assessment Based on Diffusion Model)

[0042] This embodiment corresponds to Figure 1 The core illustration of "Scenario Two" can be found in [link to relevant documentation]. Figure 3 .

[0043] The intermediary still uses the original image dataset to generate images, employing a diffusion model for generation, and lets the generated image set be... The diffusion model adds noise forward to the original image and then denoises it backward to generate the final image. When generating images in intermediate batches, the images generated in the same batch originate from the same noise distribution, but the initial noise for each image in the batch is independently and randomly sampled. This independently sampled noise ensures the diversity of images within the same batch. In multi-image generation, the quality of the generative model can be measured by the IS metric, which evaluates the quality and diversity of images generated by the generative model. Its core idea is that high-quality images should be explicitly identifiable by a classifier, and the generated images should cover multiple categories. Therefore, in scenarios involving the generation of multiple images, the value of these images is not only related to the original image but also to other generated images; evaluating the value of these images requires considering the entire generated set rather than individual images. Therefore, by combining the contribution of the original image set with the diversity of the generated image set, a utility function for the generated image set can be constructed to measure the value of the entire generated set.

[0044] Shapley values ​​ensure a fair distribution of total revenue among participants; similarly, they can be used to allocate the total value of the generated set to individual images. Shapley values ​​guarantee that the price allocated to each generated image reflects its actual contribution. As a commodity, the value assessment of images generated by the diffusion model also needs to incorporate market demand factors; ultimately, the price of each generated image is the sum of the price allocated by the Shapley value and the market demand value assessment. Specific steps include:

[0045] Step 1: Construct a utility function for the generated image set using the utility contribution of the original image and the utility contribution of the diversity of the generated images.

[0046] Step 2: Calculate the Shapley value for a single generated image using the constructed utility function.

[0047] Step 3: Calculate the value assessment of the generated image based on market demand.

[0048] Step 4: Final value assessment.

[0049] The following will explain steps 1 to 4 in detail:

[0050] Calculating the Shapley value of a single generated image as its price requires first defining a utility function for the image set. Since a diffusion model is used to generate images, the image generation process can be described using a Markov chain. The diffusion model associates generated images with original images through forward and backward processes. For example: generated image data The relationship between the noise and the original image data x0 can be described using probability. The noise added in the forward propagation of the diffusion model implies information from the original image. In the backward propagation, the noise predicted by the model ∈ θ (x t The noise residual (t) should be consistent with the noise added from the original image x0 during the forward process. Therefore, the noise residual is defined as the difference between the generated path noise and the theoretical noise:

[0051] Δ∈=||∈ θ (x t ,t)-∈|| 2 (6)

[0052] Where ∈ t x is generated during the forward process t The actual noise added at that time. By comparing the difference between the predicted noise and the theoretical noise, the association probability between the generated image and the original image can be indirectly estimated. The smaller the residual, the higher the probability of association between the generated image and the original image. The higher the probability from x0:

[0053] Therefore, the association probability between the generated image and the original image has a general relationship as shown in the formula above. The association probability can be used to directly measure the price contribution of the original image to the generated image, without needing to use methods that statistically analyze the content similarity between the two. Specifically, for the original image i... j and generating images Using π jk This indicates the probability of association between the two.

[0054] The correlation probability is calculated as follows. During the noise addition and denoising process, the actual noise added to the original image and the predicted noise for the generated image are obtained by sampling time steps. If multiple time steps are sampled uniformly, the mean squared error residual of the noise is calculated, and then the residual is divided by the number of time steps to obtain the average residual. If only one time step is sampled, the noise residual of that time step is directly calculated. Then, softmax is used to convert the residual into a probability distribution. The smaller the residual, the greater the correlation probability between the generated image and the original image.

[0055] Based on the association probability, the utility contribution of the original image is defined. In addition, the IS and LPIPS metrics are used to measure the diversity of the generated image set, and the utility contribution of generated image diversity is defined accordingly. Combining these two factors, the final comprehensive utility function is obtained.

[0056] The utility contribution of the original image. The generated image is approximated. With the original image i j The probability of association π jk . Then i j right The price effect can be expressed as the product of probability and price: π jk p j Combine all generated images with i j By adding the association probabilities, we can obtain i. j For the generated image set Price contribution: By further summing the price contributions of all the original images in I, we can obtain the original image dataset I paired with the generated image set. Price contribution:

[0057]

[0058] The utility contribution of generated image diversity. The IS metric is used to measure the diversity of a multi-class generated image set, and it is a commonly used indicator to evaluate the quality of generated images. A higher value indicates better quality and a more balanced class distribution. If the generated image set is single-class, the LPIPS metric is used to measure diversity. A larger LPIPS distance indicates greater perceptual difference between images, and thus higher image set diversity. For the IS metric, the generated image set is first used... Calculate the IS score, then use the function AF(·) = ID / (IS+1) to map the score to the range of 0-1, and obtain the result. The diversity score is calculated by summing the pairwise LPIPS distances between generated images and then averaging the distances. The output range of LPIPS values ​​is typically [0,1], therefore the diversity score can be directly obtained from the average LPIPS distance. on the other hand, This is the average price of the original image set I. The product of the average price and the number of generated images. This is the base cost for the entire generated image set I. Combining the diversity score with the base cost yields the utility based on the generated image set:

[0059]

[0060] The utility function for generating the image set. Based on the above two utilities, the final combined utility function is obtained through weighting:

[0061]

[0062] Here, α and β satisfy α,β∈(0,1),α+β=1, and they measure the relative importance of the two utility generation. The specific weight values ​​are adjusted by the intermediary according to the actual situation.

[0063] The utility function U(·) of the generated image set is used to measure the value of the generated image set. Based on this, the Shapley value is introduced to measure the contribution of each generated image in the set to the set, that is, to determine the price of a single generated image. Specifically, a single generated image... The Shapley values ​​are shown below:

[0064]

[0065] Based on the above idea, the Monte Carlo method is used to calculate the Shapley value of the generated image. Specifically, let m s It represents the number of sampling permutations. It is a permutation of π t The generated image with position index j is selected. Random permutations of the generated images are sampled, and then each permutation is traversed from beginning to end, calculating the marginal contribution of each generated image (lines 3-6). With a sufficiently large sample set, the final estimated Shapley value is the average of all calculated marginal contributions in the sample (lines 7-8). This Monte Carlo method can provide an unbiased estimate of the Shapley value. Previous studies have shown that, given the error bound ∈ and the confidence level 1-δ, if the sample size... but Where r represents the range of values ​​that the utility function can take.

[0066]

[0067] The price of an unconditionally generated image based on a diffusion model consists of two parts: a price component assigned by the Shapley value and a market demand value assessment component (calculated in the same way as in the first scenario). The final generated image price for:

[0068]

[0069] Example 3: Third Scenario Application (Conditional Generative Image Value Assessment Based on Diffusion Model)

[0070] This embodiment corresponds to Figure 1 For the optimization model and solution process of "Scenario 3" in the example, please refer to [link / reference]. Figure 4 .

[0071] Compared to unconditional image generation, training conditional generation models is more complex and computationally expensive because it requires learning complex relationships such as text-image alignment. Therefore, in such scenarios, it's assumed that an intermediary surveys buyer demand in the market (e.g., there are many buyers in the market who want to purchase images of the "cat" type) and runs a diffusion model to generate multiple sets of conditional images. The agent requires the buyer to... Select and purchase an image set Rather than individual images. If buyers purchase only a single image, it could result in high intermediary costs and meager profits.

[0072] Three key factors are considered: the buyer, the intermediary, and the quality of the generated image set. In such a complex image generation scenario, the buyer's suggestions directly determine the output, and the generated images must meet specific requirements. Therefore, value assessment must consider the buyer's involvement. It is assumed that the intermediary remains neutral and does not charge any costs, meaning the total revenue from the sale of the generated image set is distributed fairly to the original image owners. In practice, the intermediary can use a portion of the total revenue for building the generation model and other costs, which does not affect the mechanism design. Therefore, a value assessment mechanism is designed from a dual perspective, aiming to maximize intermediary revenue while ensuring that the images meet the quality expectations of different buyers, thereby achieving non-negative utility. Specific steps include:

[0073] Step 1: Evaluate the quality of multiple generated image sets.

[0074] Step 2: Define the buyer's utility function based on the buyer's willingness to pay.

[0075] Step 3: Based on maximizing intermediary revenue, the value assessment strategy is formulated as an optimization problem.

[0076] Step 4: Solve the optimization problem to obtain the price of the generated image set.

[0077] The following will explain steps 1 to 4 in detail:

[0078] Quality evaluation of the generated image set. Given a set of generated images. The quality is evaluated using two metrics: FID and CLIP Score. Let q1 represent the quality result obtained by the CLIP Score metric, and q2 represent the quality result obtained by the FID metric, where q1,q2∈(0,1). The CLIP Score ranges from [-1,1], where 1 indicates the highest similarity between the condition and the image (perfect match), 0 indicates no similarity between the condition and the image, and -1 indicates the lowest similarity between the condition and the image. Specifically, the calculation... The CLIP Score of each generated image is calculated, and the average value is used as the score for the entire set. Let its score be . Then use linear mapping to CLIP avg Mapped to the range of 0-1.

[0079] The FID metric theoretically ranges from [0, +∞). A smaller FID value indicates a better generative model and higher image quality. Using an exponential decay function, the FID metric is mapped to a value between 0 and 1; a higher value represents better image quality.

[0080]

[0081] Where ρ is a parameter controlling the decay rate, ρ>0. The larger the value of ρ, the faster the decay rate. In practical applications, different datasets and models will result in FID values ​​falling into different ranges, therefore the value of ρ should also be different. For example, if on a certain dataset, an FID value less than 30 indicates that the generated image set is of good quality, then when FID equals 30, Taking a passing score of 0.6, we can calculate ρ≈0.017 according to the formula.

[0082] The intermediary surveyed buyers in the market regarding their requirements for generated images, and then analyzed the prompts for L different buyers {B1, B2, ..., B...} with the same requirements. L Provides K sets of generated images For generating image sets A quality score is given using FID and CLIP Score.

[0083] A buyer's utility reflects their willingness to purchase the generated image set. Buyers make their decisions based on their specific requirements and the price of the set. According to economic theory, utility is defined as a function of willingness to pay minus price. A function that integrates the concepts of saturated quality and retention quality is used to define the buyer's willingness to pay, assuming that: when product quality is below a certain lower threshold (indicating that the consumer will not consider purchasing the product), a particular consumer's willingness to pay is zero; when quality exceeds a certain upper threshold, the willingness to pay remains constant. These two thresholds are called the saturated quality level and the retention quality level, respectively. Using this model, let's assume buyer B... l For quality score Each has a marginal willingness to pay Retain quality level and saturation mass level Buyer B l Willing to generate image sets mass fraction Payment

[0084]

[0085] Buyer B l mass fraction Willingness to pay and The calculation method is the same. Buyer B l The total willingness to pay is the average of the willingness to pay across all quality dimensions. Therefore, buyer B l Buy Its utility is: To generate an image set The price.

[0086] In the data market, intermediaries determine the quantity and price of generated image sets, maximizing revenue by providing image sets with varying quality scores. Therefore, the valuation strategy can be formulated as a revenue maximization problem:

[0087]

[0088] The intermediary's income, R, is the sum of all payments made by the buyers. (Binary variable x) l,k Buyer B l Select Set The first constraint in the above formula states that each buyer can purchase at most one set. The second constraint requires that the buyer's utility after the purchase be non-negative to enforce rational decision-making, thereby preventing them from suffering losses. If the utility is negative, the purchase is invalid. The third constraint guarantees that buyer B... l Choose to purchase the set that will bring the greatest utility. This valuation problem is essentially a classic nonlinear constrained optimization problem. Therefore, a genetic algorithm is used to find an approximate solution.

[0089] Genetic algorithms are optimization algorithms based on the principles of natural selection and genetics, searching for optimal approximate solutions by simulating the evolutionary process. This paper applies genetic algorithms to solve the problem of maximizing intermediary income, thereby evaluating the value of generated image sets. Specifically, based on the quality scores of the generated image sets provided by the intermediaries and the surveyed buyer parameters, the willingness to pay for each quality score by all buyers is calculated using the willingness-to-pay function defined by formula (15). and Then, the average value is taken to calculate the buyer's total willingness to pay W for each generated image set. l,k (Lines 1-5) To represent the buyer's maximum total willingness to pay, A population of N price vectors is randomly initialized within the range, each price vector representing a scheme for evaluating the value of a set of K images (lines 7-9). The main evolutionary process of the genetic algorithm (lines 10-21) is explained in detail: Initialize the income of each price vector in the population (lines 11-12). Calculate the price vectors by simulating buyers' purchasing decisions. The income generated (Lines 13-17). Determine if the current income exceeds the historical maximum income (Line 18). Update the optimal price vector and maximum income (Line 19). Generate a new generation of population through basic operations of the genetic algorithm (elite retention, selection, crossover, mutation), initiating the next generation of evolution (Lines 20-21). Unlike the first completely random initialization, the generation of the new population is a directed evolution based on historical experience. This process is repeated multiple times until the preset number of generations M is reached, ultimately returning the optimal price vector and corresponding maximum income found during the entire evolution process.

[0090]

[0091]

[0092] The above embodiments illustrate in detail the specific workflow and technical details of the present invention in different scenarios, fully demonstrating the practicality, effectiveness, and innovation of the present invention. Implementing the generative image value assessment method and system for the data market according to the above embodiments requires the equipment to have basic functionalities.

[0093] The specific embodiments described above are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Any modifications, variations, or equivalent substitutions made based on the technical solutions of this invention without departing from the core principles of this invention shall fall within the scope of the claims of this invention.

Claims

1. A generative image value assessment method and system for the data market, characterized in that, Includes the following steps: Step 1: Identify the use cases of the generative model. The use cases include at least: single generated image value assessment scenario where the generative model is unknown, unconditional batch generated image value assessment scenario based on diffusion model, and conditional generated image value assessment scenario based on diffusion model. Step 2: Based on the identified scene, call the corresponding value assessment model to calculate the price of the generative image or image set.

2. The generative image value assessment method and system for the data market according to claim 1, characterized in that, The value assessment steps for the scenario where the generation model is unknown and the value assessment of a single generated image are as follows: Based on the similarity between the target generated image and the original image, the similarity between the original images, and the price of the original images, the value contribution of the original image set to the generated target is calculated; based on the content of the target generated image, its demand value is predicted through market demand; the value contribution is added to the market demand value to obtain the final price of the target generated image.

3. The generative image value assessment method and system for the data market according to claim 1, characterized in that, The value assessment steps for the unconditional batch generation image value assessment scenario based on the diffusion model include: Based on the generation principle of the diffusion model, the degree of association between the original image set and each image in the generated image set is determined, and the total contribution value of the original image set to the generated image set is calculated. The internal diversity value of the generated image set is calculated. Combining the total contribution value and the diversity index, a total utility function of the generated image set is constructed. Using the Shapley value method, the total value of the set calculated by the total utility function is allocated to each image in the generated image set to obtain the basic value of a single image. The basic value of a single image is added to its corresponding market demand value to obtain the final price of each generated image.

4. The generative image value assessment method and system for the data market according to claim 3, characterized in that, In the step of value allocation using the Shapley value method, the Monte Carlo sampling algorithm is used to approximate the Shapley value.

5. The generative image value assessment method and system for the data market according to claim 1, characterized in that, The value assessment steps for the conditional image generation value assessment scenario based on the diffusion model include: Based on the conditions provided by the buyer, multiple sets of generated images are generated, and the quality of each set of generated images is evaluated to obtain a quality score. Based on the quality score and a preset buyer's willingness to pay function, a buyer utility function is constructed. With the goal of maximizing intermediary income and constrained by the buyer's non-negative utility and the set of buyer's choice maximizing utility, a value assessment optimization model is established. By solving the value assessment optimization model, the optimal price of each set of generated images is determined.

6. The generative image value assessment method and system for the data market according to claim 5, characterized in that, The steps for solving the value assessment optimization model are performed using a genetic algorithm.

7. The generative image value assessment method and system for the data market according to claim 1, characterized in that, The system includes: The first value assessment engine is used to execute the value assessment logic for a single generated image whose generative model is unknown. The second value assessment engine is used to execute the value assessment logic for unconditionally batch-generated images based on the diffusion model; The third value assessment engine is used to execute the conditional generated image value assessment logic based on the diffusion model; The central processing module receives valuation requests, invokes the corresponding valuation engine, and outputs the final price.

Images