Immersive 3D scene intelligent optimization generation system and method based on AIGC

Abstract

The invention discloses an immersion type 3D scene intelligent optimization generation system and method based on AIGC, and particularly relates to the crossing field of computer graphics and artificial intelligence technologies. The system comprises a scene analysis and semantic modeling module, a style control module, a semantic interactive editing module, an immersive feedback acquisition module and an intelligent optimization loop module. According to the core method, a scene semantic map is constructed by analyzing user input, and scene visual consistency is guaranteed by using a global style control vector; realizing incremental updating driven by a local instruction through a semantic graph interface; and scene automatic iterative optimization is driven by fusing user behavior data and dominant feedback. According to the method and the system, the full-process intelligentization from generation, editing to optimization is realized, and the efficiency, the consistency and the user intention fitting degree of 3D scene creation are improved.

Description

Technical Field

[0001] This invention relates to the intersection of computer graphics and artificial intelligence technologies, and more specifically, to an immersive 3D scene intelligent optimization generation system and method based on AIGC. Background Technology

[0002] In recent years, AI-generated content technology has expanded from image and text generation to 3D content creation. AIGC-based 3D scene generation, as an interdisciplinary research direction of computer graphics and artificial intelligence, is evolving from generating static, isolated models to creating dynamic, interactive, and complex scenes. This aims to lower the creative threshold and provide efficient tools for fields such as virtual reality and digital twins.

[0003] However, existing technologies have significant shortcomings: First, it is difficult to ensure the style consistency of objects, materials, lighting and other elements within the scene, which can easily lead to visual disjointedness; second, modifications require regenerating the entire scene or relying on manual adjustments, lacking efficient semantic-level editing methods; and third, it lacks the ability to utilize feedback information during user use, making it impossible to achieve continuous optimization based on experience data.

[0004] Therefore, this paper proposes an AIGC-based intelligent optimization and generation system and method for immersive 3D scenes to address the aforementioned issues. The aim is to solve key problems in the existing technologies, primarily including: how to ensure a highly consistent visual style within complex 3D scenes generated by AIGC; how to provide an efficient interactive editing method that understands scene semantics and supports local incremental updates to improve modification efficiency; and how to construct a closed-loop system that integrates multi-dimensional user feedback and enables scene self-iterative optimization, thereby narrowing the gap between the generated results and user expectations. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AIGC-based intelligent optimization generation system and method for immersive 3D scenes, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent optimization and generation method for immersive 3D scenes based on AIGC, the method comprising the following steps: S1. Scene parsing and semantic association construction steps: Receive scene description text or reference image input by the user, parse and generate initial 3D scene data through AIGC model and simultaneously construct scene semantic map; S2, Style Consistency Control Step: Generate a global style control vector based on user input, and adjust the AIGC model generation process through an attention mechanism to achieve style consistency of scene elements; S3, Semantic real-time editing and incremental generation steps, provides a graphical interface to realize the editing operation of scene semantic graph, and converts semantic-level operations into local incremental generation instructions; S4. Multi-dimensional feedback collection and optimization loop: Collect behavioral data and explicit feedback during the user's immersive experience, and integrate them to drive automated iterative optimization of the scenario.

[0007] Preferably, in the semantic real-time editing and incremental generation steps, the operation on the scene semantic graph specifically includes: responding to the user's selection of a certain entity node, providing an editable attribute list, wherein the attribute list includes one or more of the entity's geometric shape, material type, and size parameters; when the user modifies a certain attribute, the system encodes the modification instruction together with the contextual semantic information of the entity to which it belongs into a text prompt word, inputs it into the AIGC model to generate a new 3D asset that is coordinated with the original scene and replaces the original asset.

[0008] Preferably, in the multi-dimensional feedback collection and optimization loop step, the specific process of transforming implicit feedback data into an optimization target description includes: identifying scene areas where user attention is higher than a first preset threshold as high attention areas based on the gaze focus trajectory, and identifying areas where user dwell time is lower than a second preset threshold as low interest areas based on the movement path; when generating optimization instructions, prioritizing detail enhancement of the high attention areas, and simplifying or adjusting the layout of the low interest areas to improve scene smoothness.

[0009] Preferably, in the style consistency control step, the global style control vector is generated in the following ways: when the input is a style label, the corresponding vector is obtained by searching a predefined style vector library; when the input is a reference image, its high-level features are extracted using an image encoder and obtained by passing them through a mapping network.

[0010] An AIGC-based intelligent optimization and generation system for immersive 3D scenes, comprising: The scene parsing and semantic modeling module is configured to use the AIGC model to process user input to generate an initial 3D scene, and to build and manage the scene semantic graph. The style control module is configured to generate the global style control vector and to ensure style consistency of scene elements by intervening in the AIGC model generation process. The semantic interactive editing module is configured to provide the graphical interface, receive and process user operations on the scene semantic graph, and convert them into incremental generation instructions; The immersive feedback collection module is configured to collect implicit and explicit feedback data from users within the immersive scene environment provided by the rendering engine. The intelligent optimization loop module is configured to integrate multi-dimensional feedback data, generate optimization target descriptions, and schedule AIGC models to complete automated iterative optimization of the scene.

[0011] Preferably, the semantic interactive editing module further includes a version management unit, which records the scene semantic graph state and the corresponding 3D scene snapshot before and after each semantic-level editing operation, allowing users to go back to any historical editing node and start a new editing branch based on that node.

[0012] Preferably, the intelligent optimization loop module integrates a preference learning unit. This unit trains a user preference profile using a machine learning model based on the user's historical feedback data, and incorporates the user preference profile as a weight factor into the generation process of the optimization target description in subsequent automated optimization.

[0013] Preferably, when constructing the scene semantic graph, the scene parsing and semantic modeling module further associates a confidence score with each entity node in the graph. This score is determined by the generation probability of the AIGC model when generating the entity. When initiating optimization, the intelligent optimization loop module prioritizes optimizing and enhancing entity nodes with confidence scores lower than the third preset threshold.

[0014] The technical effects and advantages of this invention are as follows: Compared to existing technologies, this invention introduces multi-level semantic association control based on an attention mechanism. First, it uses a global style control vector to uniformly modulate all entity nodes in the scene semantic graph. This approach ensures that all generated scene elements, from macroscopic layout to microscopic materials, follow a unified visual style instruction, effectively solving the problem of mixed and inconsistent scene element styles in existing technologies, and improving the overall artistic quality and visual consistency of the generated scene.

[0015] Compared to existing technologies, this invention constructs an interactive scene semantic graph and enables semantic real-time editing and local incremental generation. It allows users to directly modify object attributes at the semantic level, and the system automatically translates modification commands into precise generation prompts, updating only the relevant areas. This approach liberates users from tedious low-level 3D editing, achieving highly efficient scene iteration with a "what you see is what you get" approach, reducing modification costs and improving creative efficiency.

[0016] Compared to existing technologies, this invention constructs a multi-dimensional feedback-driven automated optimization loop. It systematically collects and analyzes explicit feedback and implicit behavioral data from users during immersive experiences, transforming this data into specific optimization objectives to further drive the AIGC model. This method enables the system to understand user intent and improve itself, ultimately outputting scene content that better reflects users' real needs and preferences, achieving a paradigm shift from unidirectional generation to continuous intelligent optimization. Attached Figure Description

[0017] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0019] As attached Figure 1 The method shown is an intelligent optimization and generation method for immersive 3D scenes based on AIGC, which includes the following steps: The scene parsing and semantic association construction steps are specifically implemented as follows: receiving natural language scene description text or reference image input by the user, the natural language scene description text is used to perform entity recognition, relation extraction and spatial preposition parsing through a pre-trained large language model to generate structured scene description metadata; The reference image is then segmented, depth estimated, and semantically labeled by the visual perception module, and is also converted into the structured scene description metadata. Subsequently, the scene description metadata is input into a 3D generation network based on a diffusion model. This network associates text or image features with the 3D latent space through a cross-attention mechanism. First, a low-resolution 3D voxel mesh is generated as the basic geometry of the scene. Then, a high-precision 3D scene mesh model is gradually refined through a cascaded upsampling network. Simultaneously, a material generation network synthesizes the corresponding physically based rendering material maps by querying the material database or generating adversarial networks based on the scene description metadata and geometric information; a lighting estimation module initializes global lighting parameters, including ambient light intensity, main light source direction and color temperature, based on the time, weather information and overall tone of the reference image in the scene description. During this process, the system constructs a scene semantic graph in parallel. The graph adopts a resource description framework graph structure, where nodes represent scene entity instances. Node attributes include entity type, spatial coordinates, size, orientation, and generation confidence. Edges represent spatial or logical relationships between entities. The spatial relationships are calculated by the spatial reasoning engine based on entity bounding boxes, including but not limited to proximity, containment, and support relationships. The style consistency control steps are implemented as follows: the system provides a predefined style tag library for users to choose from, or allows users to upload reference images to extract style features; when the input is a style tag, the system accesses a pre-trained style vector embedding model to map the discrete style tags into a high-dimensional, continuous global style control vector. When the input is a reference image, the system uses a deep convolutional neural network to extract its multi-scale features, and then uses a fully connected layer mapping network to compress the features into a global style control vector with the same dimension as the style label vector. in,

[0020] This formula calculates the style attention weight received by the i-th entity node. .

[0021] Variable description: s: Represents the global style control vector.

[0022] : Represents the feature vector of the i-th entity node in the scene semantic graph.

[0023] : Represents the dot product operation of vectors, used to calculate similarity.

[0024] exp: represents the natural exponential function.

[0025] N: Represents the total number of entity nodes in the scene.

[0026] : Represents the normalized attention weight obtained by the i-th entity node.

[0027] When the 3D generation network generates scene elements, the global style control vector is used as an additional conditional input. It interacts with the text description features corresponding to each entity node in the scene semantic graph through a cross-modal attention layer to calculate the weight influence of the style vector on the entity generation features. Thus, when generating 3D mesh geometric details, high-frequency and low-frequency information of material maps, and surface reflection properties, the global style vector is modulated to ensure the style consistency from macro layout to micro material representation. The semantic real-time editing and incremental generation steps are implemented as follows: The system provides a graphical user interface that simultaneously presents the rendered 3D scene view and the visual graph structure of the scene semantic map; users can directly select objects in the 3D view or select corresponding nodes in the graph, and the sidebar of the interface will dynamically display the list of editable attributes of the entity node; when the user modifies an attribute, such as changing the material type of "wooden chair" to "metal", the system encapsulates the modification operation, the target entity identifier and its contextual semantic information into a structured instruction; The instruction is converted into a precise natural language prompt by an instruction parser, such as "Replace the wooden chair next to the dining table in the scene with a chair of the same style but made of metal, and keep it in harmony with the surrounding environment." This prompt, along with the global state code of the current scene, is input into the 3D generation network. However, the generation range is limited to a local space centered on the target entity by using region masking technology. The network only performs incremental generation on the masked area. The newly generated 3D asset replaces the original asset after coordinate alignment, while the geometry, material, and lighting data of the rest of the scene remain unchanged, thereby achieving efficient local updates. The multi-dimensional feedback collection and optimization loop is implemented as follows: During the immersive experience phase, users browse the scene through a virtual reality headset or a highly interactive desktop 3D application; the system continuously records the two-dimensional or three-dimensional coordinate sequence of the user's gaze focus on the screen or virtual space at a specific sampling frequency through an integrated eye tracker, forming a gaze focus trajectory; the system records the changes in the user's position coordinates in the virtual world through logs, forming a movement path; at the same time, it captures the user's interaction events with scene objects, such as clicking, grabbing, and rotating; in,

[0028] This formula calculates the attention density value D(p) of any point p in the scene, which is determined by all gaze focal points. It is generated by superimposing contributions from the Gaussian kernel function.

[0029] Variable description: p: represents any point in the 3D space of the scene.

[0030] : Represents the coordinates of the viewpoint of the k-th record.

[0031] K: Represents the total number of focal points within the sampling time period.

[0032] ||p− ||: Represents point p and focus The Euclidean distance between them.

[0033] : Represents the bandwidth parameter of the Gaussian kernel function, controlling the range of attention diffusion.

[0034] D(p): Represents the attention density value at point p.

[0035] These data are uploaded to the server in real time as implicit feedback; the system also provides a feedback panel for users to submit explicit feedback, including slider numerical ratings, text comments, or direct annotations in the 3D scene; the server-side feedback analyzer preprocesses the implicit feedback data, including using a Kalman filter to smooth the gaze trajectory and using a clustering algorithm to identify densely lingering areas and rapidly passing areas in the movement path; the analyzer marks areas as high interest areas based on preset thresholds, such as areas where a single gaze lasts for more than a certain number of milliseconds, and areas where the total dwell time within a certain radius is less than a certain number of seconds as low interest areas; Subsequently, the analyzer integrates these analysis results with explicit feedback. Through a feedback summary generation module, the structured data is transformed into a new natural language optimization description, such as "enhance the model details and texture resolution of the vase in the center of the dining table, simplify the model face count of the bookshelf in the corner of the living room, and adjust the lighting in the southwest corner to make it brighter." This optimization description serves as a new conditional input, driving the 3D generation network to perform a new round of automated iterative optimization on the entire scene or a specified area, forming a closed loop.

[0036] In the semantic real-time editing and incremental generation steps, the operation on the scene semantic graph specifically includes: responding to the user's selection of a certain entity node, providing an editable attribute list, which is dynamically generated by parsing the attribute fields predefined in the resource description framework graph structure of the entity node, and the field types include strings, enumeration values, and numerical ranges; when the user modifies a certain attribute, the system triggers an attribute change event, the event handler captures the attribute values ​​before and after the change, and retrieves all the associations of the entity node, including its parent container node, adjacent nodes, and nodes with logical constraints with it; The system packages this information with the change instructions and sends it to a prompt word generation template. This template follows specific natural language generation rules to ensure that the output prompt words are unambiguous and contain sufficient contextual constraints. The generated prompts are fed into the 3D generation network based on the diffusion model. After receiving the generation task with local region constraints, the network’s denoising process is strictly carried out within the 3D spatial region defined by the mask. Through latent spatial interpolation technology, it ensures that the newly generated local assets smoothly transition with the unmodified region at the boundary, and that the geometric topology and material UV mapping remain continuous. After a new asset is generated, the system will perform an automatic collision detection and physical property verification to ensure that the replaced object will not clip through other objects in the scene and will inherit the original physical properties. After the replacement is completed, the system will update the attribute information of the corresponding node in the scene semantic graph.

[0037] In the multi-dimensional feedback acquisition and optimization loop step, the specific process of transforming implicit feedback data into an optimization target description includes: the feedback analyzer performs spatiotemporal clustering analysis on the gaze focus trajectory data, identifies spatial clusters that are frequently gazed at and have long gaze durations, and defines the outer cubic regions of these clusters as high attention areas. Simultaneously, piecewise linear fitting is performed on the movement path data to identify path segments where the user's movement speed decreases or repeated back-and-forth behavior occurs, and the surrounding spherical areas are also included in the high interest area; for the identification of low interest areas, the cumulative dwell time of the user in each cell of the scene space grid is calculated, and a specific proportion of cell units with a time lower than the statistical average are marked as low interest areas. When generating optimization instructions, the system calls the optimization strategy knowledge base, which contains a series of rules. For example, for areas of high interest, the "detail enhancement" rule is triggered. This rule generates instructions that require improving the model-level details of objects in that area, increasing the intensity of normal maps or displacement maps, and improving texture resolution. For low-interest regions, the "simplification and streamlining" rule is triggered. This rule may generate instructions to simplify the model facets of objects in the region, use simplified material shaders, or even suggest adjusting the scene layout, such as moving obstacles to create smoother movement paths. The structured instructions output by these rules are finally integrated into a coherent natural language description by the feedback summary generation module.

[0038] In the style consistency control step, the global style control vector is generated in the following ways: the predefined style vector embedding model is a variant of the contrastive language-image pre-trained model pre-trained on a large-scale stylized image-text pairing dataset, whose text encoder is able to map style labels to a vector space in which style concepts are densely distributed. The mapping network consists of several fully connected layers with residual connections. Its function is to project the image features extracted by the convolutional neural network from the image content-dominated feature space to a shared style space aligned with the text style vector. The training objective of the mapping network is to minimize the cosine distance loss between the style vector generated from the image and the style vector generated from the text label under the same style label.

[0039] An AIGC-based intelligent optimization and generation system for immersive 3D scenes, comprising: The scene parsing and semantic modeling module is based on a server equipped with a high-performance graphics processor. At the software level, this module includes a natural language processing submodule, a computer vision submodule, a 3D generative network submodule, and a semantic graph management submodule. The Natural Language Processing submodule loads the pre-trained large language model weights and is responsible for text parsing; the Computer Vision submodule integrates the instance segmentation neural network and the monocular depth estimation network and is responsible for image parsing; the 3D Generative Network submodule maintains the model weights and inference engine of the diffusion-based 3D generative network, material generative network and illumination estimation module; and the Semantic Graph Management submodule is responsible for the construction, persistent storage and real-time querying of the resource description framework graph structure. The style control module, which acts as a conditional control middleware, stores a style vector library and runs the forward inference code of the style vector embedding model and mapping network. After receiving style input, it outputs a global style control vector and injects it into the conditional input channel of the 3D generation network through a predefined application programming interface. The semantic interactive editing module is presented as a standalone graphical user interface application or a plugin integrated into an existing 3D editing platform. Its core consists of a scene graph editor component and an instruction parser and forwarder. The scene graph editor component is responsible for rendering the visualization interface of the scene semantic graph and processing user interaction events. The instruction parser and forwarder is responsible for converting interaction events into machine-readable instructions and sending them to the backend generation service. The immersive feedback acquisition module consists of a client component and a server component. The client component is embedded in the immersive application and is responsible for calling the device interface to collect raw sensor data and interaction logs. The server component provides a data receiving interface and a real-time data processing pipeline to perform preliminary cleaning and formatting on the uploaded raw data. The intelligent optimization loop module, deployed on the server side, includes the feedback analyzer, optimization strategy knowledge base, and feedback summary generator. It periodically pulls processed feedback data from the immersive feedback acquisition module, performs analysis, strategy matching, and instruction generation, and finally calls the application programming interface of the 3D generative network submodule to initiate the optimization task.

[0040] The semantic interactive editing module also includes a version management unit, which is implemented as a data version control system based on a directed acyclic graph structure. Each semantic-level editing operation is recorded as a new version node. The node data includes the difference data of the scene semantic graph before and after the operation, the incremental snapshot of the 3D scene resource file, and the prompt words and random seeds used to generate the version. The system provides a version comparison view, allowing users to visualize the differences between different versions, and supports rolling back the scene state to any historical version node. The rollback operation is to apply the difference data in reverse to the current state. Users can also branch out new editing branches from any historical node to achieve non-linear editing history.

[0041] The intelligent optimization loop module integrates a preference learning unit, which is essentially an online learning machine learning model, such as a logistic regression model or a shallow neural network. The model takes feature vectors extracted from the user's historical feedback data as input. The feature vectors include, but are not limited to, the user's preferred style keywords, the types of regions they frequently visit, the categories of objects they frequently interact with, and the number of times they have given positive feedback to different optimization strategies in the past. The model outputs a user preference profile vector, which, when the feedback summary generator is working, acts as an additive bias term or a key-value pair in the attention mechanism, influencing the generation direction of the optimization target description, making it more inclined to generate scene content that conforms to the user's historical preferences, thereby achieving personalized optimization.

[0042] When constructing the scene semantic graph, the scene parsing and semantic modeling module further associates a confidence score with each entity node in the graph. This score is calculated by the spatial average of the probability values ​​of each voxel or vertex in the generated region belonging to the target category in the final step of the denoising process when the 3D generation network generates the entity. When the intelligent optimization loop module initiates optimization, its optimization strategy knowledge base contains one optimization rule for entities with low confidence. When the confidence score of an entity node is lower than the threshold dynamically calculated based on the confidence distribution of all entities, the system will generate an instruction to regenerate or enhance the details of the entity, and emphasize the semantic description of the entity in the prompt words to improve its generation accuracy. Example

[0043] The detailed workflow of the AIGC-based immersive 3D scene intelligent optimization generation system and method described in this invention is as follows.

[0044] The workflow of this invention begins with the user inputting scene description text or reference image. The system first performs deep analysis through a pre-trained large language model or computer vision module to identify entities, attributes and spatial logical relationships in the scene. Then, it drives a 3D generative network based on a diffusion model to produce an initial mesh model, material map and lighting parameters, and simultaneously constructs a structured scene semantic map. Based on this, the system generates a global style control vector according to the style tags or reference image features specified by the user. This vector interacts with each node in the semantic graph through a cross-modal attention mechanism, thereby uniformly regulating the visual style of all elements during the generation process and ensuring overall coordination. Subsequently, users can directly manipulate nodes or attributes in the semantic graph through the graphical interface. The system will convert such advanced semantic commands into accurate natural language prompts in real time, and start local incremental generation only for the changed areas, efficiently completing scene modifications while keeping the rest stable. When users enter the generated scene for an immersive experience, the system continuously collects implicit data such as their gaze trajectory, movement path, and interaction behavior, and combines it with possible explicit scores. Through analysis algorithms, it identifies high attention areas and low interest areas, and then automatically generates optimization description instructions. This drives the AIGC model to perform targeted iterative optimization of scene details, layout, or lighting, thus forming a complete closed loop from intelligent generation and intuitive editing to continuous optimization.

[0045] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent optimization and generation of immersive 3D scenes based on AIGC, characterized in that, The method includes the following steps: S1. Scene parsing and semantic association construction steps: Receive scene description text or reference image input by the user, parse and generate initial 3D scene data through AIGC model and simultaneously construct scene semantic map; S2, Style Consistency Control Step: Generate a global style control vector based on user input, and adjust the AIGC model generation process through an attention mechanism to achieve style consistency of scene elements; S3, Semantic real-time editing and incremental generation steps, provides a graphical interface to realize the editing operation of scene semantic graph, and converts semantic-level operations into local incremental generation instructions; S4. Multi-dimensional feedback collection and optimization loop: Collect behavioral data and explicit feedback during the user's immersive experience, and integrate them to drive automated iterative optimization of the scenario.

2. The method for intelligent optimization and generation of immersive 3D scenes based on AIGC according to claim 1, characterized in that, In the semantic real-time editing and incremental generation steps, the operation on the scene semantic graph specifically includes: responding to the user's selection of a certain entity node, providing an editable attribute list, the attribute list including one or more of the entity's geometric shape, material type, and size parameters; when the user modifies a certain attribute, the system encodes the modification instruction together with the contextual semantic information of the entity to which it belongs into a text prompt word, inputs it into the AIGC model to generate a new 3D asset that is coordinated with the original scene and replaces the original asset.

3. The method for intelligent optimization and generation of immersive 3D scenes based on AIGC according to claim 1, characterized in that, In the multi-dimensional feedback collection and optimization loop, the specific process of transforming implicit feedback data into an optimization target description includes: identifying scene areas where user attention is higher than a first preset threshold based on the gaze focus trajectory as high attention areas, and identifying areas where user dwell time is lower than a second preset threshold based on the movement path as low interest areas; when generating optimization instructions, priority is given to enhancing the details of the high attention areas, and simplifying or adjusting the layout of the low interest areas to improve scene smoothness.

4. The method for intelligent optimization and generation of immersive 3D scenes based on AIGC according to any one of claims 1 to 3, characterized in that, In the style consistency control step, the global style control vector is generated in the following ways: when the input is a style label, the corresponding vector is obtained by searching a predefined style vector library; when the input is a reference image, its high-level features are extracted using an image encoder and obtained by passing them through a mapping network.

5. An AIGC-based intelligent optimization and generation system for immersive 3D scenes, implementing the method as described in any one of claims 1 to 4, characterized in that, The system includes: The scene parsing and semantic modeling module is configured to use the AIGC model to process user input to generate an initial 3D scene, and to build and manage the scene semantic graph. The style control module is configured to generate the global style control vector and to ensure style consistency of scene elements by intervening in the AIGC model generation process. The semantic interactive editing module is configured to provide the graphical interface, receive and process user operations on the scene semantic graph, and convert them into incremental generation instructions; The immersive feedback collection module is configured to collect implicit and explicit feedback data from users within the immersive scene environment provided by the rendering engine. The intelligent optimization loop module is configured to integrate multi-dimensional feedback data, generate optimization target descriptions, and schedule AIGC models to complete automated iterative optimization of the scene.

6. The immersive 3D scene intelligent optimization and generation system based on AIGC according to claim 5, characterized in that, The semantic interactive editing module also includes a version management unit, which records the scene semantic graph state and the corresponding 3D scene snapshot before and after each semantic-level editing operation, allowing users to go back to any historical editing node and start a new editing branch based on that node.

7. The immersive 3D scene intelligent optimization and generation system based on AIGC according to claim 5, characterized in that, The intelligent optimization loop module integrates a preference learning unit. This unit trains a user preference profile using a machine learning model based on the user's historical feedback data, and incorporates the user preference profile as a weight factor into the generation process of the optimization target description in subsequent automated optimization.

8. The immersive 3D scene intelligent optimization and generation system based on AIGC according to claim 5, characterized in that, When constructing a scene semantic graph, the scene parsing and semantic modeling module further associates a confidence score with each entity node in the graph. This score is determined by the generation probability of the AIGC model when generating the entity. When initiating optimization, the intelligent optimization loop module prioritizes optimizing and enhancing entity nodes with confidence scores lower than the third preset threshold.

Images