Method for dynamically constructing and generating meta-universe scene based on generative AI

Abstract

The application discloses a kind of based on generative AI's meta-universe scene dynamic construction and generation method, the method includes: receiving user input information and scene running environment information, the user input information and scene running environment information are multi-modal analysis, obtain scene construction demand;According to the scene construction demand generates scene semantic constraint graph, the scene semantic constraint graph includes scene node and the constraint edge for limiting the space relationship, functional relationship and interactive relationship between scene node, the present application relates to meta-universe interactive technical field.A kind of based on generative AI's meta-universe scene dynamic construction and generation method, by local difference generation mechanism, when user input or environmental event triggers scene change, only variable area or dynamic area is executed generation operation, and the position, number, material, interactive state and business binding relationship of frozen object are kept unchanged, enough reduce computing overhead and improve scene update continuity.

Description

Technical Field

[0001] This invention relates to the field of metaverse interaction, and more specifically, to a method for dynamically constructing and generating metaverse scenes based on generative AI. Background Technology

[0002] With the development of generative AI, 3D modeling, and real-time rendering technologies, metaverse scenes are gradually shifting from manual construction to automatic generation. Existing technologies already offer solutions for constructing metaverse scenes. For example, patent number CN202210704585.X discloses a method and apparatus for constructing metaverse scenes. This method generates an initial virtual scene by determining the metaverse virtual site and basic model elements, and adds decoration information after receiving decoration operations to form a target virtual scene. Another example is CN118154746A, which proposes determining whether a virtual scene is generated by a cloud rendering server or a user device based on event resource storage and user device attributes. While these solutions can improve scene generation or rendering efficiency, they still tend to focus on overall construction, resource addition, or rendering device selection, lacking mechanisms for freezing the state of confirmed scene objects, constraining semantic relationships, generating local differences, and correcting anomalies. When users continuously edit, multiple users collaborate, or the terminal load fluctuates, problems such as confirmed objects being overwritten, inconsistent spatial relationships, lost interactive scripts, and rendering stuttering in high-load areas can easily occur, making it difficult to balance scene continuity, runnability, and real-time generation efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide a method for dynamically constructing and generating metaverse scenes based on generative AI. This method solves the problems of existing methods for dynamically constructing and generating metaverse scenes, such as the overwriting of confirmed objects, inconsistencies in spatial relationships, loss of interactive scripts, and rendering stutters in high-load areas when users continuously edit, multiple people collaborate, or the terminal load fluctuates. These methods are difficult to balance scene continuity, operability, and real-time generation efficiency, and cannot meet user needs.

[0004] This invention achieves the above objective through the following technical solution: a method for dynamically constructing and generating metaverse scenes based on generative AI, the method comprising: Receive user input information and scene operation environment information, perform multimodal parsing on the user input information and scene operation environment information to obtain scene construction requirements; A scene semantic constraint graph is generated based on the scene construction requirements. The scene semantic constraint graph includes scene nodes and constraint edges used to limit the spatial relationships, functional relationships and interaction relationships between scene nodes. Read the historical running state of the current metaverse scene, establish a state memory freeze layer, and divide the scene content into frozen objects, variable regions, and dynamic regions based on the state memory freeze layer; When a scene change event is detected, a local difference generation instruction is generated based on the scene semantic constraint graph and the state memory freeze layer. The local difference generation instruction is input into the generative AI model, so that the generative AI model generates difference scene content only for the variable region or dynamic region; The operability of the differential scene content is verified. If an error occurs during the verification, the error information is written back to the scene semantic constraint graph and a local regeneration is triggered. Adjust the generation granularity of differential scene content based on user interaction popularity and terminal rendering load; The validated differential scene content is written into the metaverse scene runtime environment, and the scene semantic constraint graph and state memory freeze layer are updated synchronously.

[0005] Furthermore, the multimodal parsing includes parsing any one or any combination of text commands, voice commands, gesture commands, viewpoint movement information, object selection information, multi-person entry events, device status change events, and business process triggering events; The analysis results include scene theme, spatial scope, object category, style requirements, interaction requirements, real-time change requirements, event source, trigger time, target object identifier, and execution priority; The execution priority is used to determine the generation order of local difference generation instructions when multiple scene change events occur simultaneously.

[0006] Furthermore, the process of generating the scene semantic constraint graph includes: Generate theme nodes, spatial region nodes, and scene boundary nodes based on the scene theme and spatial range; Generate terrain nodes, building nodes, virtual object nodes, lighting nodes, and interaction nodes based on the object category; Establish constraint edges between adjacent nodes, functionally related nodes, and interactively related nodes; The constraint edges include spatial adjacency constraint edges, size ratio constraint edges, passage constraint edges, occlusion constraint edges, lighting direction constraint edges, function dependency constraint edges, and interaction trigger constraint edges. Each constraint edge is configured with constraint type, constraint strength, conflict handling priority, and version identifier.

[0007] Furthermore, the state memory freeze layer is established by reading confirmed scene anchor points, object identities, spatial coordinates, material styles, collision boundaries, interaction scripts, user editing results, business binding relationships, and multi-person collaborative states; Among them, the frozen objects are scene objects that have been confirmed by the user, have been bound to business data, have been locked, or are in the current interaction process; The variable area is a scene area that allows content replacement, style adjustment, or structural changes according to user instructions; Dynamic areas are scene areas that update according to changes in user behavior, multi-person collaboration status, business processes, or device status.

[0008] Furthermore, the generation process of the local difference generation instruction includes: Identify the target area and target object corresponding to the scene change event; The target region and target object are matched with scene nodes in the scene semantic constraint graph to obtain the changing region, changing object, and preserved object; The state memory freeze layer determines the positions, numbers, materials, collision boundaries, interaction scripts, and business binding relationships that cannot be changed; The changed area, changed object, preserved object, content that cannot be changed, style adjustment requirements, interaction retention requirements, spatial continuity requirements, and temporal continuity requirements are encapsulated into local difference generation instructions.

[0009] Furthermore, the generation process of the differential scene content includes: The local difference generation instructions, scene semantic constraint graph, and state memory freeze layer are input into the generative AI model; The generative AI model identifies the boundaries of the content that can be generated this time; Generate a 3D model structure, material maps, lighting parameters, colliders, interactive scripts, and runtime configuration within the content boundaries; During the generation process, the position, number, material, collision boundary, interaction script and business binding relationship of the frozen object remain unchanged, and the spatial connectivity, proportional consistency and interaction inheritance relationship between the newly generated content and the adjacent scene content are maintained according to the constraint edges.

[0010] Furthermore, the operability check includes: Spatial connectivity verification, collision boundary verification, lighting consistency verification, object scale verification, interaction trigger verification, and rendering level verification; When an abnormal message is generated, such as when a path is blocked, collision boundaries overlap, lighting direction conflicts, object proportions exceed constraints, interaction scripts are missing, or rendering layer occlusion is detected; Write the abnormal information into the corresponding scene node or constraint edge; Based on the semantic constraint graph of the scene after writing back, a local correction instruction is generated, and the generative AI model is driven to regenerate only the abnormal region locally until the verification result meets the writing conditions.

[0011] Furthermore, the user interaction popularity is determined based on the user's perspective position, the central area of ​​vision, dwell time, click frequency, number of interactions, and the density of multiple people gathering; The terminal rendering load is determined based on the terminal frame rate, video memory usage, network latency, and rendering queue pressure. The user interaction heat and terminal rendering load are mapped to spatial region nodes in the scene semantic constraint graph, and a generation granularity level is configured for each spatial region node according to the mapping result. The generation granularity level includes fine generation level, regular generation level and simplified generation level.

[0012] Furthermore, when the user interaction popularity of the target area is higher than the first popularity threshold and the terminal rendering load is lower than the first load threshold, the target area is configured to a fine generation level, so that the generative AI model outputs a high-precision model, high-resolution materials, complete colliders, fine lighting and complete interaction scripts. When the user interaction heat of the target area is lower than the second heat threshold, the target area is in a distant location, or the terminal rendering load is higher than the second load threshold, the target area is configured to a simplified generation level, so that the generative AI model outputs low-precision contours, simplified materials, basic lighting and shadows, and placeholder interaction scripts. The first popularity threshold, the second popularity threshold, the first load threshold, and the second load threshold are determined based on historical interaction data, terminal performance data, and scenario service level.

[0013] Furthermore, when the user's perspective shifts to the target area of ​​the simplified generation level, the user's stay time in the target area exceeds the stay threshold, or the density of multiple people gathering is higher than the gathering threshold, the target area is adjusted from the simplified generation level to the fine generation level, and progressive asset completion is performed. The progressive asset completion includes replacing low-precision outlines with high-precision models, replacing simplified materials with high-resolution materials, completing colliders, enhancing lighting details, and loading complete interaction scripts. When the terminal frame rate is lower than the frame rate threshold, the video memory usage is higher than the video memory threshold, or the network latency is higher than the latency threshold, the generation granularity of non-current interactive areas is reduced, while maintaining the rendering accuracy and response speed of the current interactive object. After the asset incremental replenishment or load degradation is completed, the scene semantic constraint graph and state memory freeze layer are updated synchronously.

[0014] The beneficial effects of this invention are as follows: 1. This invention constructs a scene semantic constraint graph, which represents scene themes, spatial regions, terrain, buildings, virtual objects, lighting and shadows and interaction relationships in a graph structure. This allows generative artificial intelligence models to be constrained by clear spatial, functional and interaction relationships when generating scene content, thereby reducing spatial conflicts, scale anomalies and semantic inconsistencies between scene objects.

[0015] 2. This invention establishes a state memory freeze layer to record confirmed scene anchor points, object identities, spatial coordinates, material styles, collision boundaries, interaction scripts, user editing results, business binding relationships, and multi-person collaboration states. This allows subsequent generation tasks to inherit existing scene states, avoids confirmed content being repeatedly overwritten, and improves the stability of continuous editing of the metaverse scene.

[0016] 3. This invention uses a local differential generation mechanism to perform generation operations only on variable or dynamic regions when user input or environmental events trigger scene changes, while keeping the position, number, material, interaction state, and business binding relationship of frozen objects unchanged. Compared with the overall regeneration method, this can reduce computational overhead and improve the continuity of scene updates.

[0017] 4. This invention uses a feasibility check and anomaly write-back mechanism to check the spatial connectivity, collision boundaries, lighting consistency, object proportions, interaction triggers, and rendering levels of the generated differential scene content, and writes the abnormal results back to the scene semantic constraint graph to drive the generative artificial intelligence model to perform local regeneration, thereby improving the deployability and interactivity of the generated content in the metaverse operating environment.

[0018] 5. This invention dynamically adjusts the granularity of the output content of the generative artificial intelligence model by calculating the user interaction heat and terminal rendering load in real time. This allows the user's focus area to obtain a high-precision model, high-resolution materials and complete interactive scripts, while the distant or high-load areas adopt a simplified generation method, thus taking into account visual effects, interactive experience and terminal performance.

[0019] 6. This invention uses a progressive completion and load degradation mechanism to perform asset completion when the user's perspective shifts to a low-precision area, the dwell time increases, or the density of multiple people gathering increases. It also reduces the generation granularity of non-critical areas when the terminal frame rate decreases, the video memory usage increases, or the network latency increases, thereby achieving real-time adaptive operation of the metaverse scene. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1This is a flowchart illustrating the overall dynamic construction process of the present invention; Figure 2 This is a flowchart of the local differential instruction generation process of the present invention; Figure 3 This is a flowchart of the operability verification process of the present invention. Detailed Implementation

[0021] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0022] Example 1: Please see Figure 1-3 This invention provides a method for dynamically constructing and generating metaverse scenes based on generative AI. This method can be applied to metaverse exhibition halls, virtual parks, virtual conference spaces, immersive teaching platforms, digital twin factories, virtual commercial districts, and interactive 3D business systems. Through scene semantic constraint graphs, state memory freeze layers, local difference generation instructions, operability verification, and dynamic adjustment mechanisms for generation granularity, this method enables metaverse scenes to maintain spatial structural continuity, stable interactive relationships, and controllable operational performance under conditions of continuous editing, multi-user collaboration, and changes in terminal performance.

[0023] In this embodiment, the system first receives user input information and scene runtime environment information, and performs multimodal parsing on the user input information and scene runtime environment information to obtain scene construction requirements. User input information includes at least one of text commands, voice commands, gesture commands, viewpoint movement information, and object selection information; scene runtime environment information includes at least one of multi-person entry events, device status change events, and business process triggering events. After parsing the above information, the system extracts scene theme, spatial range, object category, style requirements, interaction requirements, real-time change requirements, event source, trigger time, target object identifier, and execution priority.

[0024] For example, in a virtual exhibition hall application, if a user inputs "Add a new energy equipment display area to the center of the exhibition hall, keeping the entrance, navigation buttons, and existing exhibits unchanged," the system will parse this text command, determine that the scene theme is new energy equipment display, the spatial scope is the central area of ​​the exhibition hall, the object categories include display stands, equipment models, explanatory panels, and lighting components, the interaction requirements include clicking to view, information pop-ups, and path navigation, and the objects to be kept include the entrance, navigation buttons, and existing exhibits. If multiple people are detected entering the central area of ​​the exhibition hall simultaneously, this event will be used as scene runtime environment information and combined with the user's input command to form the scene construction requirements.

[0025] To ensure that data from different sources can participate in unified calculations, this embodiment normalizes the dwell time, click frequency, number of interactions, multi-user gathering density, terminal frame rate, video memory usage, network latency, rendering queue pressure, spatial deviation, and other parameters involved in subsequent weighted calculations, converting them into dimensionless values ​​ranging from 0 to 1. The normalization process uses the following formula:

[0026] in, Represents the original parameters to be normalized; This indicates the lower limit of the parameter within the range of historical samples or preset samples; This indicates the upper limit of the parameter within the range of historical samples or preset samples; This represents the dimensionless parameter after normalization. If the calculated... If it is less than 0, then take 0; if the calculated value is less than 0, then take 0. If the value is greater than 1, then the value is 1.

[0027] Historical samples are preferably selected from the most recent 30 scene operation records or scene operation records within the most recent 7 days; when historical samples are insufficient, a system preset range is used. Preferably, the preset range for user dwell time is 0 to 30 seconds, the preset range for click frequency is 0 to 10 times / minute, the preset range for interaction count is 0 to 20 times / minute, the preset range for multi-person gathering density is 0 people / square meter to 3 people / square meter, and the preset range for network latency is 0 milliseconds to 200 milliseconds.

[0028] After receiving the scene construction requirements, the system generates a scene semantic constraint graph based on these requirements. The scene semantic constraint graph includes scene nodes and constraint edges used to define the spatial, functional, and interactive relationships between these nodes. Specifically, the system generates theme nodes, spatial region nodes, and scene boundary nodes based on the scene theme and spatial extent; it generates terrain nodes, building nodes, virtual object nodes, lighting nodes, and interactive nodes based on object categories; and it establishes constraint edges between adjacent nodes, functionally related nodes, and interactively related nodes. These constraint edges include spatial adjacency constraint edges, size ratio constraint edges, passage constraint edges, occlusion constraint edges, lighting direction constraint edges, functional dependency constraint edges, and interaction trigger constraint edges. Each constraint edge is configured with constraint type, constraint strength, conflict handling priority, and version identifier.

[0029] Among them, spatial adjacency constraint edges are used to limit the positional relationship between different scene areas; size ratio constraint edges are used to limit the proportional relationship between the new model and the original scene objects; access constraint edges are used to limit the user's walkable path and passage boundaries; occlusion constraint edges are used to limit the new object from obscuring key entrances, function buttons or business panels; lighting direction constraint edges are used to limit the consistency of the direction, intensity and shadow range of the new light with the original light source; functional dependency constraint edges are used to limit the binding relationship between scene objects and business functions; and interaction trigger constraint edges are used to limit the triggering relationship between click, drag, navigation, pop-ups and script execution.

[0030] The system reads the historical running state of the current metaverse scene, establishes a state memory freeze layer, and divides the scene content into frozen objects, variable areas, and dynamic areas based on the state memory freeze layer. The historical running state includes confirmed scene anchor points, object identities, spatial coordinates, material styles, collision boundaries, interaction scripts, user editing results, business binding relationships, and multi-person collaboration status.

[0031] When establishing the state memory freeze layer, the system calculates a freeze determination value for each scene object:

[0032] in, Represents scene object The value for determining whether the item is frozen; This indicates the user's confirmation status. It is set to 1 when the scene object has been confirmed by the user and 0 when it has not been confirmed. This indicates the business binding status. It is 1 when the scene object is bound to business data, business links, or business processes, and 0 when it is not bound. Indicates the locked state; it is 1 when the scene object is locked by the user or the system, and 0 when it is unlocked. Indicates the current interaction state. It is set to 1 when the scene object is in the process of clicking, dragging, viewing, editing, or multi-person collaborative operation, and to 0 when it is not in the process of interaction. , , , These are the corresponding weights, and the sum of the four weights is 1.

[0033] In this embodiment, the preferred value is: , , , The rules for determining the above weights are as follows: the business binding state has the greatest impact on the stability of scenario operation and the continuity of business processes, therefore... The value is higher than other weights; both user confirmation status and locked status directly affect whether an object is allowed to be modified, therefore and Take equal weight; the current interaction state is mainly used to avoid sudden changes in objects during the interaction process, therefore It is lower than the business binding weight.

[0034] when When this happens, the corresponding scene object is classified as a frozen object; when Furthermore, if the object is located within a user-specified area of ​​change, it is classified as a mutable object; When an object is updated based on changes in user behavior, multi-user collaboration status, business processes, or device status, it is classified as a dynamic object.

[0035] The freeze threshold of 0.60 is determined based on the historical erroneous modification rate. When the proportion of confirmed objects that have been erroneously modified in historical operations exceeds 5%, the freeze threshold will be lowered to 0.55 to expand the scope of freeze protection; when the proportion of variable areas that cannot respond to user edits in a timely manner in historical operations exceeds 5%, the freeze threshold will be raised to 0.65 to reduce editing obstacles caused by excessive freezing.

[0036] Through the state memory freeze layer, the system can clearly identify which objects can be adjusted by the generative artificial intelligence model and which objects must not be changed, thereby avoiding problems such as historical objects being overwritten, business binding relationships being lost, and interactive scripts becoming invalid that are prone to occur in the overall regeneration method.

[0037] When a scene change event is detected, the system generates local differential generation instructions based on the scene semantic constraint graph and the state memory freeze layer. Specifically, the system first determines the target region and target object corresponding to the scene change event, and then matches the target region and target object with scene nodes in the scene semantic constraint graph to obtain the changed region, changed object, and preserved object. Subsequently, the system determines the positions, numbers, materials, collision boundaries, interaction scripts, and business binding relationships that must not be changed based on the state memory freeze layer.

[0038] To avoid the generative AI model repeatedly generating the entire scene, the system calculates the necessity of differential generation for candidate regions:

[0039] in, Indicates the region The necessity of difference generation; Indicates the region The degree of matching with scene change events is determined based on the target object identifier, spatial range, and event type. Indicates the region The average value of the internal object freezing determination value; the higher the freezing degree, the less suitable it is to perform generation, therefore, it adopts... Participate in calculation; Indicates the region User interaction popularity; Indicates the region The proportion of internal constraints that are not met; , , , These are the weights generated by difference, and the sum of the four weights is 1.

[0040] Preferably, , , , Among them, the event matching degree determines whether the area belongs to the target of this change, therefore Take the highest value; the freeze level is used to prevent accidental modification of stable objects, therefore... Secondly, interaction popularity is used to ensure that areas of user interest are processed first; the unmet constraint ratio is used to drive local corrections.

[0041] when At that time, the area This region has been identified as the local difference generation region for this study. when If the error occurs, differential generation will not be performed on that area. The differential generation necessity threshold of 0.55 is determined based on historical scene editing records. Specifically, it is calculated by statistically analyzing the distribution of scores for areas that were manually confirmed to need modification in historical generation tasks, taking the median as the initial threshold, and then adjusting it according to the false generation rate. When the false generation rate is higher than 5%, the threshold is increased by 0.05; when the missed generation rate is higher than 5%, the threshold is decreased by 0.05.

[0042] After determining the local differential generation area, the system encapsulates the changed area, changed object, preserved object, content that must not be changed, style adjustment requirements, interaction retention requirements, spatial continuity requirements, and temporal continuity requirements into local differential generation instructions. Among them, the spatial continuity requirement is used to ensure that newly added or modified content remains passable with adjacent spatial areas, without abnormal overlap or proportional conflicts; the temporal continuity requirement is used to ensure that the state inheritance relationship is maintained between multiple generations of the same scene, so that the scene does not change abruptly during continuous editing by the user.

[0043] The system inputs local differential generation instructions into the generative AI model, enabling the model to generate differential scene content only for variable or dynamic regions. Specifically, the local differential generation instructions, scene semantic constraint graph, and state memory freeze layer are input into the generative AI model. The generative AI model first identifies the allowed content boundaries and generates the 3D model structure, material maps, lighting parameters, colliders, interaction scripts, and runtime configurations within these boundaries. During the generation process, the position, number, material, collision boundaries, interaction scripts, and business binding relationships of frozen objects remain unchanged. Based on the constraint edges in the scene semantic constraint graph, the generative AI model ensures that the newly generated content maintains spatial connectivity, proportional consistency, and interactive inheritance relationships with adjacent scene content.

[0044] For example, when a new exhibition area is added to the center of a virtual exhibition hall, the generative AI model only generates display stands, equipment models, explanatory panels, and lighting components within the variable area in the center of the hall, without changing the entrance location, navigation buttons, existing exhibit numbers, or existing business links. The new exhibition area maintains traffic constraints with existing passageways, lighting direction constraints with existing light sources, and interactive trigger constraints with the explanatory panels.

[0045] After the differential scene content is generated, the system performs a runnable check on the differential scene content. If an error occurs during the check, the error information is written back to the scene semantic constraint graph, triggering a local regeneration. The runnable check includes spatial connectivity check, collision boundary check, lighting consistency check, object scale check, interaction trigger check, and rendering level check.

[0046] In this embodiment, the system calculates a feasibility score based on each verification item:

[0047] in, Indicates the region The operability score; This indicates the spatial connectivity verification result; This indicates the collision boundary verification result; This indicates the result of the illumination consistency verification. This indicates the result of the object proportion verification. This indicates the verification result triggered by the interaction; This indicates the rendering level verification result. All the above verification results are dimensionless parameters, ranging from 0 to 1. The higher the value, the more the verification result meets the running requirements. , , , , , These are the runnability verification weights, and the sum of the six weights is 1.

[0048] Preferably, , , , , , Spatial connectivity, collision boundaries, and interaction triggers directly affect whether the scene can run and whether the user can interact. , , Choose higher values; lighting consistency and object proportion affect visual quality and spatial realism, therefore , Take the median value; the rendering level mainly affects the display effect, therefore It is lower than the first three categories of mandatory performance indicators.

[0049] when If the three verification items of spatial connectivity, collision boundary and interaction trigger are all not less than 0.80, the content of the differential scene is determined to meet the writing conditions. Otherwise, anomaly information is generated and written back to the corresponding node or constraint edge in the scene semantic constraint graph, driving the generative AI model to locally regenerate the abnormal region. Here, 0.85 serves as the comprehensive runnability threshold, determined based on historical running records of online scenes. When the verification score is below 0.85, the user feedback anomaly rate increases significantly; therefore, 0.85 is used as the writing threshold. Spatial connectivity, collision boundaries, and interaction triggers are hard operating conditions. Even if the comprehensive score meets the requirements, if any of these hard conditions is below 0.80, writing to the metaverse scene's operating environment is still not allowed.

[0050] During scene execution, the system further adjusts the granularity of differential scene content generation based on user interaction intensity and terminal rendering load. First, the system collects user viewpoint position, center area of ​​vision, dwell time, click frequency, number of interactions, and density of multiple users, and calculates user interaction intensity.

[0051] in, Indicates the region User interaction popularity; This indicates the user's level of attention, measured by the center and area of ​​the user's field of vision. The distance to the center is normalized; the closer the distance, the better. The larger; Indicates the user is in the region Normalized dwell time within; Indicates the region The combined value of normalized click frequency and interaction count within the range; Indicates the region Normalized multi-person cluster density within the area; , , , These are the popularity weights, and the sum of the four weights is 1.

[0052] Preferably, , , , Among these, visual focus most directly reflects the user's current attention location, therefore Take the highest value; dwell time and number of interactions reflect the user's sustained attention and operational intent, therefore and Take the same value; the density of multiple users is used to reflect the demand for collaborative access, and its direct impact on the real-time operation of a single user is slightly lower, therefore Below The above weights can be adjusted based on historical interaction logs. The adjustment rules are as follows: statistically analyze the regions where users actually triggered high-precision loading during historical operations, and perform correlation analysis between each parameter and the actual triggering results; the higher the correlation, the greater the corresponding weight, and the sum of the adjusted weights remains 1.

[0053] The system also collects terminal frame rate, video memory usage, network latency, and rendering queue pressure, and calculates the terminal rendering load:

[0054] in, Indicates the region The corresponding terminal rendering load; This represents the normalized terminal frame rate metric. A higher frame rate indicates a lower load risk; therefore, it is used... Participate in calculation; This indicates the normalized video memory usage; This represents the normalized network latency; This indicates the normalized rendering queue pressure; , , , These are the load weights, and the sum of the four weights is 1.

[0055] Preferably, , , , Among them, frame rate directly reflects the smoothness of terminal rendering, therefore Take the highest value; VRAM usage and network latency affect graphics resource loading and scene synchronization, respectively, therefore... and Take the same value; the rendering queue pressure is used to help determine short-term congestion, therefore The values ​​are lower than the first three. For frame rate parameters, 60 frames per second is preferred as the standard frame rate. When the actual frame rate of the terminal reaches or exceeds 60 frames per second, Set to 1; when the actual frame rate of the terminal is lower than 30 frames per second, Set to 0; when the actual frame rate of the terminal is between 30 frames / second and 60 frames / second, normalize according to the linear rule.

[0056] After obtaining user interaction popularity and terminal rendering load, the system calculates the granular score of the region:

[0057] in, Indicates the region The generation granularity score; Indicates the region User interaction popularity; Indicates the region Terminal rendering load; Indicates the region Business level parameters; , , These are the granularity adjustment weights, and the sum of the three weights is 1.

[0058] Preferably, , , Among these factors, user interaction activity is the primary factor determining whether refined generation is necessary. Take the highest value; the terminal rendering load determines whether the system has the ability to generate detailed content, therefore Secondly, business priority is used to ensure that important business areas receive higher production priority. Used to compensate for business importance. Business level parameter. The value is determined by the business configuration of the metaverse scene. Preferably, the value is 0.30 for the general display area, 0.60 for the key interaction area, and 0.90 for the core business area. The business level is pre-configured by the scene management terminal or determined based on historical visit volume and business conversion frequency.

[0059] when At that time, the area Configured to the fine generation level, it generates high-precision models, high-resolution materials, complete colliders, fine lighting, and complete interaction scripts; when At that time, the area Configured to the normal generation level, it generates medium-precision models, standard materials, basic colliders, standard lighting, and main interaction scripts; when At that time, the area Configured to simplify the generation level, generating low-precision outlines, simplified materials, basic lighting and shadows, and placeholder interaction scripts.

[0060] Among them, 0.70 and 0.40 are the generation granularity thresholds, which are determined by the following rules: based on the actual dwell time, number of interactions and frame rate changes of users after entering a certain area in historical operation data, when the actual interaction probability of users in a certain scoring interval is higher than 70% and the average frame rate of the terminal is not lower than 45 frames / second, the interval is classified as fine generation level; when the actual interaction probability of users is lower than 30% or the average frame rate of the terminal is lower than 30 frames / second, the interval is classified as simplified generation level.

[0061] When a user's view shifts to a target area at a simplified generation level, the user's dwell time in that target area exceeds a dwell time threshold, or the density of multiple users exceeds a gathering threshold, the system adjusts the target area from a simplified generation level to a refined generation level and performs progressive asset completion. Progressive asset completion includes replacing low-precision outlines with high-precision models, replacing simplified materials with high-resolution materials, completing colliders, enhancing lighting details, and loading complete interaction scripts. The dwell time threshold is preferably set between 5 and 10 seconds, with the specific value determined by the historical average dwell time of users; the gathering threshold is preferably set between 1 and 2 people per square meter, with the specific value determined by the scene design capacity and the terminal's carrying capacity.

[0062] When the terminal frame rate is lower than the frame rate threshold, the video memory usage is higher than the video memory threshold, or the network latency is higher than the latency threshold, the system reduces the generation granularity of non-current interactive areas while maintaining the rendering accuracy and response speed of the current interactive object. The preferred frame rate threshold is 30 to 45 frames per second, the preferred video memory threshold is 80% of the terminal's available video memory, and the preferred latency threshold is 100 to 150 milliseconds. These thresholds are determined based on the terminal performance level, network environment, and scenario service level. For high-performance terminals, the finer generation level trigger condition can be increased; for low-performance terminals, the load degradation trigger condition can be reduced to ensure overall operational stability.

[0063] The system writes the validated differential scene content into the metaverse scene runtime environment and simultaneously updates the scene semantic constraint graph and state memory freeze layer. The written content includes the 3D model structure, material textures, lighting configurations, collision boundaries, interaction scripts, and scene runtime parameters. After writing, the system updates the scene semantic constraint graph with newly added or modified scene objects, updating the corresponding scene nodes, constraint edges, spatial relationships, functional relationships, and interaction relationships. Simultaneously, it writes the user-confirmed object state, editing results, business binding relationships, and multi-user collaboration state into the state memory freeze layer, ensuring that subsequent generation tasks inherit the current scene's spatial structure, object state, interaction relationships, and user editing results.

[0064] Through the above processing, this embodiment forms a closed loop of scene semantic constraint graph - state memory freeze layer - local differential generation - operability verification - abnormal write-back and regeneration - heat load granularity adjustment. This enables the generative artificial intelligence model to no longer generate the metaverse scene without constraints, but to perform controllable local generation based on clear semantic relationships, historical states and operating conditions. This improves the stability of continuous scene editing, the operability of generated content and the real-time response capability under different terminal environments.

[0065] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for dynamically constructing and generating metaverse scenes based on generative AI, characterized in that, The method includes: Receive user input information and scene operation environment information, perform multimodal parsing on the user input information and scene operation environment information to obtain scene construction requirements; A scene semantic constraint graph is generated based on the scene construction requirements. The scene semantic constraint graph includes scene nodes and constraint edges used to limit the spatial relationships, functional relationships and interaction relationships between scene nodes. Read the historical running state of the current metaverse scene, establish a state memory freeze layer, and divide the scene content into frozen objects, variable regions, and dynamic regions based on the state memory freeze layer; When a scene change event is detected, a local difference generation instruction is generated based on the scene semantic constraint graph and the state memory freeze layer. The local difference generation instruction is input into the generative AI model, so that the generative AI model generates difference scene content only for the variable region or dynamic region; The operability of the differential scene content is verified. If an error occurs during the verification, the error information is written back to the scene semantic constraint graph and a local regeneration is triggered. Adjust the generation granularity of differential scene content based on user interaction popularity and terminal rendering load; The validated differential scene content is written into the metaverse scene runtime environment, and the scene semantic constraint graph and state memory freeze layer are updated synchronously.

2. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 1, characterized in that: The multimodal parsing includes parsing any one or any combination of text commands, voice commands, gesture commands, viewpoint movement information, object selection information, multi-person entry events, device status change events, and business process triggering events; The analysis results include scene theme, spatial scope, object category, style requirements, interaction requirements, real-time change requirements, event source, trigger time, target object identifier, and execution priority; The execution priority is used to determine the generation order of local difference generation instructions when multiple scene change events occur simultaneously.

3. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 2, characterized in that, The process of generating the scene semantic constraint graph includes: Generate theme nodes, spatial region nodes, and scene boundary nodes based on the scene theme and spatial range; Generate terrain nodes, building nodes, virtual object nodes, lighting nodes, and interaction nodes based on the object category; Establish constraint edges between adjacent nodes, functionally related nodes, and interactively related nodes; The constraint edges include spatial adjacency constraint edges, size ratio constraint edges, passage constraint edges, occlusion constraint edges, lighting direction constraint edges, function dependency constraint edges, and interaction trigger constraint edges. Each constraint edge is configured with constraint type, constraint strength, conflict handling priority, and version identifier.

4. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 1, characterized in that: The state memory freeze layer is established by reading confirmed scene anchor points, object identities, spatial coordinates, material styles, collision boundaries, interaction scripts, user editing results, business binding relationships, and multi-person collaborative states; Among them, the frozen objects are scene objects that have been confirmed by the user, have been bound to business data, have been locked, or are in the current interaction process; The variable area is a scene area that allows content replacement, style adjustment, or structural changes according to user instructions; Dynamic areas are scene areas that update according to changes in user behavior, multi-person collaboration status, business processes, or device status.

5. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 4, characterized in that, The generation process of the local difference generation instruction includes: Identify the target area and target object corresponding to the scene change event; The target region and target object are matched with scene nodes in the scene semantic constraint graph to obtain the changing region, changing object, and preserved object; The state memory freeze layer determines the positions, numbers, materials, collision boundaries, interaction scripts, and business binding relationships that cannot be changed; The changed area, changed object, preserved object, content that cannot be changed, style adjustment requirements, interaction retention requirements, spatial continuity requirements, and temporal continuity requirements are encapsulated into local difference generation instructions.

6. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 5, characterized in that, The generation process of the differential scene content includes: The local difference generation instructions, scene semantic constraint graph, and state memory freeze layer are input into the generative AI model; The generative AI model identifies the boundaries of the content that can be generated this time; Generate a 3D model structure, material maps, lighting parameters, colliders, interactive scripts, and runtime configuration within the content boundaries; During the generation process, the position, number, material, collision boundary, interaction script and business binding relationship of the frozen object remain unchanged, and the spatial connectivity, proportional consistency and interaction inheritance relationship between the newly generated content and the adjacent scene content are maintained according to the constraint edges.

7. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 6, characterized in that, The operability check includes: Spatial connectivity verification, collision boundary verification, lighting consistency verification, object scale verification, interaction trigger verification, and rendering level verification; When an abnormal message is generated, such as when a path is blocked, collision boundaries overlap, lighting direction conflicts, object proportions exceed constraints, interaction scripts are missing, or rendering layer occlusion is detected; Write the abnormal information into the corresponding scene node or constraint edge; Based on the semantic constraint graph of the scene after writing back, a local correction instruction is generated, and the generative AI model is driven to regenerate only the abnormal region locally until the verification result meets the writing conditions.

8. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 1, characterized in that: The user interaction popularity is determined based on the user's perspective position, the center area of ​​the field of vision, the duration of stay, the frequency of clicks, the number of interactions, and the density of multiple people gathering; The terminal rendering load is determined based on the terminal frame rate, video memory usage, network latency, and rendering queue pressure. The user interaction heat and terminal rendering load are mapped to spatial region nodes in the scene semantic constraint graph, and a generation granularity level is configured for each spatial region node according to the mapping result. The generation granularity level includes fine generation level, regular generation level and simplified generation level.

9. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 8, characterized in that: When the user interaction popularity of the target area is higher than the first popularity threshold and the terminal rendering load is lower than the first load threshold, the target area is configured to a fine generation level, so that the generative AI model outputs a high-precision model, high-resolution materials, complete colliders, fine lighting and complete interaction scripts. When the user interaction heat of the target area is lower than the second heat threshold, the target area is in a distant location, or the terminal rendering load is higher than the second load threshold, the target area is configured to a simplified generation level, so that the generative AI model outputs low-precision contours, simplified materials, basic lighting and shadows, and placeholder interaction scripts. The first popularity threshold, the second popularity threshold, the first load threshold, and the second load threshold are determined based on historical interaction data, terminal performance data, and scenario service level.

10. The method for dynamic construction and generation of metaverse scenes based on generative AI according to claim 9, characterized in that: When the user's perspective shifts to the target area of ​​the simplified generation level, the user's stay time in the target area exceeds the stay threshold, or the density of multiple people gathering is higher than the gathering threshold, the target area is adjusted from the simplified generation level to the fine generation level, and progressive asset completion is performed. The progressive asset completion includes replacing low-precision outlines with high-precision models, replacing simplified materials with high-resolution materials, completing colliders, enhancing lighting details, and loading complete interaction scripts. When the terminal frame rate is lower than the frame rate threshold, the video memory usage is higher than the video memory threshold, or the network latency is higher than the latency threshold, the generation granularity of non-current interactive areas is reduced, while maintaining the rendering accuracy and response speed of the current interactive object. After the asset incremental replenishment or load degradation is completed, the scene semantic constraint graph and state memory freeze layer are updated synchronously.

Images